Upload 624 files

ootd/inference_ootd.py

@@ -0,0 +1,133 @@
+import pdb
+from pathlib import Path
+import sys
+PROJECT_ROOT = Path(__file__).absolute().parents[0].absolute()
+sys.path.insert(0, str(PROJECT_ROOT))
+import os
+
+import torch
+import numpy as np
+from PIL import Image
+import cv2
+
+import random
+import time
+import pdb
+
+from pipelines_ootd.pipeline_ootd import OotdPipeline
+from pipelines_ootd.unet_garm_2d_condition import UNetGarm2DConditionModel
+from pipelines_ootd.unet_vton_2d_condition import UNetVton2DConditionModel
+from diffusers import UniPCMultistepScheduler
+from diffusers import AutoencoderKL
+
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoProcessor, CLIPVisionModelWithProjection
+from transformers import CLIPTextModel, CLIPTokenizer
+
+VIT_PATH = "../checkpoints/clip-vit-large-patch14"
+VAE_PATH = "../checkpoints/ootd"
+UNET_PATH = "../checkpoints/ootd/ootd_hd/checkpoint-36000"
+MODEL_PATH = "../checkpoints/ootd"
+
+class OOTDiffusion:
+
+    def __init__(self, gpu_id):
+        self.gpu_id = 'cuda:' + str(gpu_id)
+
+        vae = AutoencoderKL.from_pretrained(
+            VAE_PATH,
+            subfolder="vae",
+            torch_dtype=torch.float16,
+        )
+
+        unet_garm = UNetGarm2DConditionModel.from_pretrained(
+            UNET_PATH,
+            subfolder="unet_garm",
+            torch_dtype=torch.float16,
+            use_safetensors=True,
+        )
+        unet_vton = UNetVton2DConditionModel.from_pretrained(
+            UNET_PATH,
+            subfolder="unet_vton",
+            torch_dtype=torch.float16,
+            use_safetensors=True,
+        )
+
+        self.pipe = OotdPipeline.from_pretrained(
+            MODEL_PATH,
+            unet_garm=unet_garm,
+            unet_vton=unet_vton,
+            vae=vae,
+            torch_dtype=torch.float16,
+            variant="fp16",
+            use_safetensors=True,
+            safety_checker=None,
+            requires_safety_checker=False,
+        ).to(self.gpu_id)
+
+        self.pipe.scheduler = UniPCMultistepScheduler.from_config(self.pipe.scheduler.config)
+        
+        self.auto_processor = AutoProcessor.from_pretrained(VIT_PATH)
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(VIT_PATH).to(self.gpu_id)
+
+        self.tokenizer = CLIPTokenizer.from_pretrained(
+            MODEL_PATH,
+            subfolder="tokenizer",
+        )
+        self.text_encoder = CLIPTextModel.from_pretrained(
+            MODEL_PATH,
+            subfolder="text_encoder",
+        ).to(self.gpu_id)
+
+
+    def tokenize_captions(self, captions, max_length):
+        inputs = self.tokenizer(
+            captions, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt"
+        )
+        return inputs.input_ids
+
+
+    def __call__(self,
+                model_type='hd',
+                category='upperbody',
+                image_garm=None,
+                image_vton=None,
+                mask=None,
+                image_ori=None,
+                num_samples=1,
+                num_steps=20,
+                image_scale=1.0,
+                seed=-1,
+    ):
+        if seed == -1:
+            random.seed(time.time())
+            seed = random.randint(0, 2147483647)
+        print('Initial seed: ' + str(seed))
+        generator = torch.manual_seed(seed)
+
+        with torch.no_grad():
+            prompt_image = self.auto_processor(images=image_garm, return_tensors="pt").to(self.gpu_id)
+            prompt_image = self.image_encoder(prompt_image.data['pixel_values']).image_embeds
+            prompt_image = prompt_image.unsqueeze(1)
+            if model_type == 'hd':
+                prompt_embeds = self.text_encoder(self.tokenize_captions([""], 2).to(self.gpu_id))[0]
+                prompt_embeds[:, 1:] = prompt_image[:]
+            elif model_type == 'dc':
+                prompt_embeds = self.text_encoder(self.tokenize_captions([category], 3).to(self.gpu_id))[0]
+                prompt_embeds = torch.cat([prompt_embeds, prompt_image], dim=1)
+            else:
+                raise ValueError("model_type must be \'hd\' or \'dc\'!")
+
+            images = self.pipe(prompt_embeds=prompt_embeds,
+                        image_garm=image_garm,
+                        image_vton=image_vton, 
+                        mask=mask,
+                        image_ori=image_ori,
+                        num_inference_steps=num_steps,
+                        image_guidance_scale=image_scale,
+                        num_images_per_prompt=num_samples,
+                        generator=generator,
+            ).images
+
+        return images

ootd/inference_ootd_dc.py

@@ -0,0 +1,132 @@
+import pdb
+from pathlib import Path
+import sys
+PROJECT_ROOT = Path(__file__).absolute().parents[0].absolute()
+sys.path.insert(0, str(PROJECT_ROOT))
+import os
+import torch
+import numpy as np
+from PIL import Image
+import cv2
+
+import random
+import time
+import pdb
+
+from pipelines_ootd.pipeline_ootd import OotdPipeline
+from pipelines_ootd.unet_garm_2d_condition import UNetGarm2DConditionModel
+from pipelines_ootd.unet_vton_2d_condition import UNetVton2DConditionModel
+from diffusers import UniPCMultistepScheduler
+from diffusers import AutoencoderKL
+
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoProcessor, CLIPVisionModelWithProjection
+from transformers import CLIPTextModel, CLIPTokenizer
+
+VIT_PATH = "../checkpoints/clip-vit-large-patch14"
+VAE_PATH = "../checkpoints/ootd"
+UNET_PATH = "../checkpoints/ootd/ootd_dc/checkpoint-36000"
+MODEL_PATH = "../checkpoints/ootd"
+
+class OOTDiffusionDC:
+
+    def __init__(self, gpu_id):
+        self.gpu_id = 'cuda:' + str(gpu_id)
+
+        vae = AutoencoderKL.from_pretrained(
+            VAE_PATH,
+            subfolder="vae",
+            torch_dtype=torch.float16,
+        )
+
+        unet_garm = UNetGarm2DConditionModel.from_pretrained(
+            UNET_PATH,
+            subfolder="unet_garm",
+            torch_dtype=torch.float16,
+            use_safetensors=True,
+        )
+        unet_vton = UNetVton2DConditionModel.from_pretrained(
+            UNET_PATH,
+            subfolder="unet_vton",
+            torch_dtype=torch.float16,
+            use_safetensors=True,
+        )
+
+        self.pipe = OotdPipeline.from_pretrained(
+            MODEL_PATH,
+            unet_garm=unet_garm,
+            unet_vton=unet_vton,
+            vae=vae,
+            torch_dtype=torch.float16,
+            variant="fp16",
+            use_safetensors=True,
+            safety_checker=None,
+            requires_safety_checker=False,
+        ).to(self.gpu_id)
+
+        self.pipe.scheduler = UniPCMultistepScheduler.from_config(self.pipe.scheduler.config)
+        
+        self.auto_processor = AutoProcessor.from_pretrained(VIT_PATH)
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(VIT_PATH).to(self.gpu_id)
+
+        self.tokenizer = CLIPTokenizer.from_pretrained(
+            MODEL_PATH,
+            subfolder="tokenizer",
+        )
+        self.text_encoder = CLIPTextModel.from_pretrained(
+            MODEL_PATH,
+            subfolder="text_encoder",
+        ).to(self.gpu_id)
+
+
+    def tokenize_captions(self, captions, max_length):
+        inputs = self.tokenizer(
+            captions, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt"
+        )
+        return inputs.input_ids
+
+
+    def __call__(self,
+                model_type='hd',
+                category='upperbody',
+                image_garm=None,
+                image_vton=None,
+                mask=None,
+                image_ori=None,
+                num_samples=1,
+                num_steps=20,
+                image_scale=1.0,
+                seed=-1,
+    ):
+        if seed == -1:
+            random.seed(time.time())
+            seed = random.randint(0, 2147483647)
+        print('Initial seed: ' + str(seed))
+        generator = torch.manual_seed(seed)
+
+        with torch.no_grad():
+            prompt_image = self.auto_processor(images=image_garm, return_tensors="pt").to(self.gpu_id)
+            prompt_image = self.image_encoder(prompt_image.data['pixel_values']).image_embeds
+            prompt_image = prompt_image.unsqueeze(1)
+            if model_type == 'hd':
+                prompt_embeds = self.text_encoder(self.tokenize_captions([""], 2).to(self.gpu_id))[0]
+                prompt_embeds[:, 1:] = prompt_image[:]
+            elif model_type == 'dc':
+                prompt_embeds = self.text_encoder(self.tokenize_captions([category], 3).to(self.gpu_id))[0]
+                prompt_embeds = torch.cat([prompt_embeds, prompt_image], dim=1)
+            else:
+                raise ValueError("model_type must be \'hd\' or \'dc\'!")
+
+            images = self.pipe(prompt_embeds=prompt_embeds,
+                        image_garm=image_garm,
+                        image_vton=image_vton, 
+                        mask=mask,
+                        image_ori=image_ori,
+                        num_inference_steps=num_steps,
+                        image_guidance_scale=image_scale,
+                        num_images_per_prompt=num_samples,
+                        generator=generator,
+            ).images
+
+        return images

ootd/inference_ootd_hd.py

@@ -0,0 +1,132 @@
+import pdb
+from pathlib import Path
+import sys
+PROJECT_ROOT = Path(__file__).absolute().parents[0].absolute()
+sys.path.insert(0, str(PROJECT_ROOT))
+import os
+import torch
+import numpy as np
+from PIL import Image
+import cv2
+
+import random
+import time
+import pdb
+
+from pipelines_ootd.pipeline_ootd import OotdPipeline
+from pipelines_ootd.unet_garm_2d_condition import UNetGarm2DConditionModel
+from pipelines_ootd.unet_vton_2d_condition import UNetVton2DConditionModel
+from diffusers import UniPCMultistepScheduler
+from diffusers import AutoencoderKL
+
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoProcessor, CLIPVisionModelWithProjection
+from transformers import CLIPTextModel, CLIPTokenizer
+
+VIT_PATH = "../checkpoints/clip-vit-large-patch14"
+VAE_PATH = "../checkpoints/ootd"
+UNET_PATH = "../checkpoints/ootd/ootd_hd/checkpoint-36000"
+MODEL_PATH = "../checkpoints/ootd"
+
+class OOTDiffusionHD:
+
+    def __init__(self, gpu_id):
+        self.gpu_id = 'cuda:' + str(gpu_id)
+
+        vae = AutoencoderKL.from_pretrained(
+            VAE_PATH,
+            subfolder="vae",
+            torch_dtype=torch.float16,
+        )
+
+        unet_garm = UNetGarm2DConditionModel.from_pretrained(
+            UNET_PATH,
+            subfolder="unet_garm",
+            torch_dtype=torch.float16,
+            use_safetensors=True,
+        )
+        unet_vton = UNetVton2DConditionModel.from_pretrained(
+            UNET_PATH,
+            subfolder="unet_vton",
+            torch_dtype=torch.float16,
+            use_safetensors=True,
+        )
+
+        self.pipe = OotdPipeline.from_pretrained(
+            MODEL_PATH,
+            unet_garm=unet_garm,
+            unet_vton=unet_vton,
+            vae=vae,
+            torch_dtype=torch.float16,
+            variant="fp16",
+            use_safetensors=True,
+            safety_checker=None,
+            requires_safety_checker=False,
+        ).to(self.gpu_id)
+
+        self.pipe.scheduler = UniPCMultistepScheduler.from_config(self.pipe.scheduler.config)
+        
+        self.auto_processor = AutoProcessor.from_pretrained(VIT_PATH)
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(VIT_PATH).to(self.gpu_id)
+
+        self.tokenizer = CLIPTokenizer.from_pretrained(
+            MODEL_PATH,
+            subfolder="tokenizer",
+        )
+        self.text_encoder = CLIPTextModel.from_pretrained(
+            MODEL_PATH,
+            subfolder="text_encoder",
+        ).to(self.gpu_id)
+
+
+    def tokenize_captions(self, captions, max_length):
+        inputs = self.tokenizer(
+            captions, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt"
+        )
+        return inputs.input_ids
+
+
+    def __call__(self,
+                model_type='hd',
+                category='upperbody',
+                image_garm=None,
+                image_vton=None,
+                mask=None,
+                image_ori=None,
+                num_samples=1,
+                num_steps=20,
+                image_scale=1.0,
+                seed=-1,
+    ):
+        if seed == -1:
+            random.seed(time.time())
+            seed = random.randint(0, 2147483647)
+        print('Initial seed: ' + str(seed))
+        generator = torch.manual_seed(seed)
+
+        with torch.no_grad():
+            prompt_image = self.auto_processor(images=image_garm, return_tensors="pt").to(self.gpu_id)
+            prompt_image = self.image_encoder(prompt_image.data['pixel_values']).image_embeds
+            prompt_image = prompt_image.unsqueeze(1)
+            if model_type == 'hd':
+                prompt_embeds = self.text_encoder(self.tokenize_captions([""], 2).to(self.gpu_id))[0]
+                prompt_embeds[:, 1:] = prompt_image[:]
+            elif model_type == 'dc':
+                prompt_embeds = self.text_encoder(self.tokenize_captions([category], 3).to(self.gpu_id))[0]
+                prompt_embeds = torch.cat([prompt_embeds, prompt_image], dim=1)
+            else:
+                raise ValueError("model_type must be \'hd\' or \'dc\'!")
+
+            images = self.pipe(prompt_embeds=prompt_embeds,
+                        image_garm=image_garm,
+                        image_vton=image_vton, 
+                        mask=mask,
+                        image_ori=image_ori,
+                        num_inference_steps=num_steps,
+                        image_guidance_scale=image_scale,
+                        num_images_per_prompt=num_samples,
+                        generator=generator,
+            ).images
+
+        return images

ootd/pipelines_ootd/attention_garm.py

@@ -0,0 +1,402 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
+from typing import Any, Dict, Optional
+
+import torch
+from torch import nn
+
+from diffusers.utils import USE_PEFT_BACKEND
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
+from diffusers.models.attention_processor import Attention
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from diffusers.models.lora import LoRACompatibleLinear
+from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        spatial_attn_inputs = [],
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        spatial_attn_input = hidden_states
+        spatial_attn_inputs.append(spatial_attn_input)
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [
+                    self.ff(hid_slice, scale=lora_scale)
+                    for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
+                ],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states, spatial_attn_inputs
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh")
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+        compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
+        for module in self.net:
+            if isinstance(module, compatible_cls):
+                hidden_states = module(hidden_states, scale)
+            else:
+                hidden_states = module(hidden_states)
+        return hidden_states

ootd/pipelines_ootd/attention_vton.py

@@ -0,0 +1,407 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
+from typing import Any, Dict, Optional
+
+import torch
+from torch import nn
+
+from diffusers.utils import USE_PEFT_BACKEND
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
+from diffusers.models.attention_processor import Attention
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from diffusers.models.lora import LoRACompatibleLinear
+from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        spatial_attn_inputs = [],
+        spatial_attn_idx = 0,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        spatial_attn_input = spatial_attn_inputs[spatial_attn_idx]
+        spatial_attn_idx += 1
+        hidden_states = torch.cat((hidden_states, spatial_attn_input), dim=1)
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        
+        hidden_states = attn_output + hidden_states
+        hidden_states, _ = hidden_states.chunk(2, dim=1)
+
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [
+                    self.ff(hid_slice, scale=lora_scale)
+                    for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
+                ],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states, spatial_attn_inputs, spatial_attn_idx
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh")
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+        compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
+        for module in self.net:
+            if isinstance(module, compatible_cls):
+                hidden_states = module(hidden_states, scale)
+            else:
+                hidden_states = module(hidden_states)
+        return hidden_states

ootd/pipelines_ootd/pipeline_ootd.py

@@ -0,0 +1,846 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from transformers import AutoProcessor, CLIPVisionModelWithProjection
+
+from .unet_vton_2d_condition import UNetVton2DConditionModel
+from .unet_garm_2d_condition import UNetGarm2DConditionModel
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+    PIL_INTERPOLATION,
+    USE_PEFT_BACKEND,
+    deprecate,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def preprocess(image):
+    deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+    deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+    if isinstance(image, torch.Tensor):
+        return image
+    elif isinstance(image, PIL.Image.Image):
+        image = [image]
+
+    if isinstance(image[0], PIL.Image.Image):
+        w, h = image[0].size
+        w, h = (x - x % 8 for x in (w, h))  # resize to integer multiple of 8
+
+        image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+        image = np.concatenate(image, axis=0)
+        image = np.array(image).astype(np.float32) / 255.0
+        image = image.transpose(0, 3, 1, 2)
+        image = 2.0 * image - 1.0
+        image = torch.from_numpy(image)
+    elif isinstance(image[0], torch.Tensor):
+        image = torch.cat(image, dim=0)
+    return image
+
+
+class OotdPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+    r"""
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        text_encoder ([`~transformers.CLIPTextModel`]):
+            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+        tokenizer ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        unet ([`UNet2DConditionModel`]):
+            A `UNet2DConditionModel` to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+            about a model's potential harms.
+        feature_extractor ([`~transformers.CLIPImageProcessor`]):
+            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+    """
+    model_cpu_offload_seq = "text_encoder->unet->vae"
+    _optional_components = ["safety_checker", "feature_extractor"]
+    _exclude_from_cpu_offload = ["safety_checker"]
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "vton_latents"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet_garm: UNetGarm2DConditionModel,
+        unet_vton: UNetVton2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPImageProcessor,
+        requires_safety_checker: bool = True,
+    ):
+        super().__init__()
+
+        if safety_checker is None and requires_safety_checker:
+            logger.warning(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet_garm=unet_garm,
+            unet_vton=unet_vton,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        image_garm: PipelineImageInput = None,
+        image_vton: PipelineImageInput = None,
+        mask: PipelineImageInput = None,
+        image_ori: PipelineImageInput = None,
+        num_inference_steps: int = 100,
+        guidance_scale: float = 7.5,
+        image_guidance_scale: float = 1.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        **kwargs,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            image (`torch.FloatTensor` `np.ndarray`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+                `Image` or tensor representing an image batch to be repainted according to `prompt`. Can also accept
+                image latents as `image`, but if passing latents directly it is not encoded again.
+            num_inference_steps (`int`, *optional*, defaults to 100):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            image_guidance_scale (`float`, *optional*, defaults to 1.5):
+                Push the generated image towards the initial `image`. Image guidance scale is enabled by setting
+                `image_guidance_scale > 1`. Higher image guidance scale encourages generated images that are closely
+                linked to the source `image`, usually at the expense of lower image quality. This pipeline requires a
+                value of at least `1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images and the
+                second element is a list of `bool`s indicating whether the corresponding generated image contains
+                "not-safe-for-work" (nsfw) content.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        # 0. Check inputs
+        self.check_inputs(
+            prompt,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+        self._guidance_scale = guidance_scale
+        self._image_guidance_scale = image_guidance_scale
+
+        if (image_vton is None) or (image_garm is None):
+            raise ValueError("`image` input cannot be undefined.")
+
+        # 1. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # check if scheduler is in sigmas space
+        scheduler_is_in_sigma_space = hasattr(self.scheduler, "sigmas")
+
+        # 2. Encode input prompt
+        prompt_embeds = self._encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+        )
+
+        # 3. Preprocess image
+        image_garm = self.image_processor.preprocess(image_garm)
+        image_vton = self.image_processor.preprocess(image_vton)
+        image_ori = self.image_processor.preprocess(image_ori)
+        mask = np.array(mask)
+        mask[mask < 127] = 0
+        mask[mask >= 127] = 255
+        mask = torch.tensor(mask)
+        mask = mask / 255
+        mask = mask.reshape(-1, 1, mask.size(-2), mask.size(-1))
+
+        # 4. set timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare Image latents
+        garm_latents = self.prepare_garm_latents(
+            image_garm,
+            batch_size,
+            num_images_per_prompt,
+            prompt_embeds.dtype,
+            device,
+            self.do_classifier_free_guidance,
+            generator,
+        )
+
+        vton_latents, mask_latents, image_ori_latents = self.prepare_vton_latents(
+            image_vton,
+            mask,
+            image_ori,
+            batch_size,
+            num_images_per_prompt,
+            prompt_embeds.dtype,
+            device,
+            self.do_classifier_free_guidance,
+            generator,
+        )
+
+        height, width = vton_latents.shape[-2:]
+        height = height * self.vae_scale_factor
+        width = width * self.vae_scale_factor
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.vae.config.latent_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        noise = latents.clone()
+
+        # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 9. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        _, spatial_attn_outputs = self.unet_garm(
+            garm_latents,
+            0,
+            encoder_hidden_states=prompt_embeds,
+            return_dict=False,
+        )
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+                # concat latents, image_latents in the channel dimension
+                scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+                latent_vton_model_input = torch.cat([scaled_latent_model_input, vton_latents], dim=1)
+                # latent_vton_model_input = scaled_latent_model_input + vton_latents
+
+                spatial_attn_inputs = spatial_attn_outputs.copy()
+
+                # predict the noise residual
+                noise_pred = self.unet_vton(
+                    latent_vton_model_input,
+                    spatial_attn_inputs,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    return_dict=False,
+                )[0]
+
+                # Hack:
+                # For karras style schedulers the model does classifer free guidance using the
+                # predicted_original_sample instead of the noise_pred. So we need to compute the
+                # predicted_original_sample here if we are using a karras style scheduler.
+                if scheduler_is_in_sigma_space:
+                    step_index = (self.scheduler.timesteps == t).nonzero()[0].item()
+                    sigma = self.scheduler.sigmas[step_index]
+                    noise_pred = latent_model_input - sigma * noise_pred
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_text_image, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = (
+                        noise_pred_text
+                        + self.image_guidance_scale * (noise_pred_text_image - noise_pred_text)
+                    )
+
+                # Hack:
+                # For karras style schedulers the model does classifer free guidance using the
+                # predicted_original_sample instead of the noise_pred. But the scheduler.step function
+                # expects the noise_pred and computes the predicted_original_sample internally. So we
+                # need to overwrite the noise_pred here such that the value of the computed
+                # predicted_original_sample is correct.
+                if scheduler_is_in_sigma_space:
+                    noise_pred = (noise_pred - latents) / (-sigma)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                init_latents_proper = image_ori_latents * self.vae.config.scaling_factor
+
+                # repainting
+                if i < len(timesteps) - 1:
+                    noise_timestep = timesteps[i + 1]
+                    init_latents_proper = self.scheduler.add_noise(
+                        init_latents_proper, noise, torch.tensor([noise_timestep])
+                    )
+
+                latents = (1 - mask_latents) * init_latents_proper + mask_latents * latents
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    vton_latents = callback_outputs.pop("vton_latents", vton_latents)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+             prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_ prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+        """
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            # textual inversion: procecss multi-vector tokens if necessary
+            if isinstance(self, TextualInversionLoaderMixin):
+                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(
+                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = text_inputs.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            prompt_embeds = self.text_encoder(
+                text_input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            prompt_embeds = prompt_embeds[0]
+
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            # textual inversion: procecss multi-vector tokens if necessary
+            if isinstance(self, TextualInversionLoaderMixin):
+                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([prompt_embeds, prompt_embeds])
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+    def run_safety_checker(self, image, device, dtype):
+        if self.safety_checker is None:
+            has_nsfw_concept = None
+        else:
+            if torch.is_tensor(image):
+                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+            else:
+                feature_extractor_input = self.image_processor.numpy_to_pil(image)
+            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        return image, has_nsfw_concept
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    def check_inputs(
+        self,
+        prompt,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def prepare_garm_latents(
+        self, image, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None
+    ):
+        if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+            raise ValueError(
+                f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+            )
+
+        image = image.to(device=device, dtype=dtype)
+
+        batch_size = batch_size * num_images_per_prompt
+
+        if image.shape[1] == 4:
+            image_latents = image
+        else:
+            if isinstance(generator, list) and len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            if isinstance(generator, list):
+                image_latents = [self.vae.encode(image[i : i + 1]).latent_dist.mode() for i in range(batch_size)]
+                image_latents = torch.cat(image_latents, dim=0)
+            else:
+                image_latents = self.vae.encode(image).latent_dist.mode()
+
+        if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+            additional_image_per_prompt = batch_size // image_latents.shape[0]
+            image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+        elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+            raise ValueError(
+                f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+            )
+        else:
+            image_latents = torch.cat([image_latents], dim=0)
+
+        if do_classifier_free_guidance:
+            uncond_image_latents = torch.zeros_like(image_latents)
+            image_latents = torch.cat([image_latents, uncond_image_latents], dim=0)
+
+        return image_latents
+    
+    def prepare_vton_latents(
+        self, image, mask, image_ori, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None
+    ):
+        if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+            raise ValueError(
+                f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+            )
+
+        image = image.to(device=device, dtype=dtype)
+        image_ori = image_ori.to(device=device, dtype=dtype)
+
+        batch_size = batch_size * num_images_per_prompt
+
+        if image.shape[1] == 4:
+            image_latents = image
+            image_ori_latents = image_ori
+        else:
+            if isinstance(generator, list) and len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            if isinstance(generator, list):
+                image_latents = [self.vae.encode(image[i : i + 1]).latent_dist.mode() for i in range(batch_size)]
+                image_latents = torch.cat(image_latents, dim=0)
+                image_ori_latents = [self.vae.encode(image_ori[i : i + 1]).latent_dist.mode() for i in range(batch_size)]
+                image_ori_latents = torch.cat(image_ori_latents, dim=0)
+            else:
+                image_latents = self.vae.encode(image).latent_dist.mode()
+                image_ori_latents = self.vae.encode(image_ori).latent_dist.mode()
+
+        mask = torch.nn.functional.interpolate(
+            mask, size=(image_latents.size(-2), image_latents.size(-1))
+        )
+        mask = mask.to(device=device, dtype=dtype)
+
+        if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+            additional_image_per_prompt = batch_size // image_latents.shape[0]
+            image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+            mask = torch.cat([mask] * additional_image_per_prompt, dim=0)
+            image_ori_latents = torch.cat([image_ori_latents] * additional_image_per_prompt, dim=0)
+        elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+            raise ValueError(
+                f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+            )
+        else:
+            image_latents = torch.cat([image_latents], dim=0)
+            mask = torch.cat([mask], dim=0)
+            image_ori_latents = torch.cat([image_ori_latents], dim=0)
+
+        if do_classifier_free_guidance:
+            # uncond_image_latents = torch.zeros_like(image_latents)
+            image_latents = torch.cat([image_latents] * 2, dim=0)
+
+        return image_latents, mask, image_ori_latents
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet_vton.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet_vton.disable_freeu()
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def image_guidance_scale(self):
+        return self._image_guidance_scale
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self.image_guidance_scale >= 1.0

ootd/pipelines_ootd/transformer_garm_2d.py

@@ -0,0 +1,449 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from .attention_garm import BasicTransformerBlock
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.embeddings import ImagePositionalEmbeddings
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
+# from diffusers.models.attention import BasicTransformerBlock
+from diffusers.models.embeddings import CaptionProjection, PatchEmbed
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        if self.is_input_continuous:
+            self.in_channels = in_channels
+
+            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+            if use_linear_projection:
+                self.proj_in = linear_cls(in_channels, inner_dim)
+            else:
+                self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
+
+            self.height = sample_size
+            self.width = sample_size
+            self.num_vector_embeds = num_vector_embeds
+            self.num_latent_pixels = self.height * self.width
+
+            self.latent_image_embedding = ImagePositionalEmbeddings(
+                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
+            )
+        elif self.is_input_patches:
+            assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+            self.height = sample_size
+            self.width = sample_size
+
+            self.patch_size = patch_size
+            interpolation_scale = self.config.sample_size // 64  # => 64 (= 512 pixart) has interpolation scale 1
+            interpolation_scale = max(interpolation_scale, 1)
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+                interpolation_scale=interpolation_scale,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        if self.is_input_continuous:
+            # TODO: should use out_channels for continuous projections
+            if use_linear_projection:
+                self.proj_out = linear_cls(inner_dim, in_channels)
+            else:
+                self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            self.norm_out = nn.LayerNorm(inner_dim)
+            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
+        elif self.is_input_patches and norm_type != "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
+            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+        elif self.is_input_patches and norm_type == "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+            self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        spatial_attn_inputs = [],
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 1. Input
+        if self.is_input_continuous:
+            batch, _, height, width = hidden_states.shape
+            residual = hidden_states
+
+            hidden_states = self.norm(hidden_states)
+            if not self.use_linear_projection:
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+            else:
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+            hidden_states = self.pos_embed(hidden_states)
+
+            if self.adaln_single is not None:
+                if self.use_additional_conditions and added_cond_kwargs is None:
+                    raise ValueError(
+                        "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+                    )
+                batch_size = hidden_states.shape[0]
+                timestep, embedded_timestep = self.adaln_single(
+                    timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+                )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+                hidden_states, spatial_attn_inputs = torch.utils.checkpoint.checkpoint(
+                    block,
+                    hidden_states,
+                    spatial_attn_inputs,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states, spatial_attn_inputs = block(
+                    hidden_states,
+                    spatial_attn_inputs,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+            output = hidden_states + residual
+        elif self.is_input_vectorized:
+            hidden_states = self.norm_out(hidden_states)
+            logits = self.out(hidden_states)
+            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+            logits = logits.permute(0, 2, 1)
+
+            # log(p(x_0))
+            output = F.log_softmax(logits.double(), dim=1).float()
+
+        if self.is_input_patches:
+            if self.config.norm_type != "ada_norm_single":
+                conditioning = self.transformer_blocks[0].norm1.emb(
+                    timestep, class_labels, hidden_dtype=hidden_states.dtype
+                )
+                shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+                hidden_states = self.proj_out_2(hidden_states)
+            elif self.config.norm_type == "ada_norm_single":
+                shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states)
+                # Modulation
+                hidden_states = hidden_states * (1 + scale) + shift
+                hidden_states = self.proj_out(hidden_states)
+                hidden_states = hidden_states.squeeze(1)
+
+            # unpatchify
+            if self.adaln_single is None:
+                height = width = int(hidden_states.shape[1] ** 0.5)
+            hidden_states = hidden_states.reshape(
+                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+            )
+            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+            output = hidden_states.reshape(
+                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+            )
+
+        if not return_dict:
+            return (output,), spatial_attn_inputs
+
+        return Transformer2DModelOutput(sample=output), spatial_attn_inputs

ootd/pipelines_ootd/transformer_vton_2d.py

@@ -0,0 +1,452 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from .attention_vton import BasicTransformerBlock
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.embeddings import ImagePositionalEmbeddings
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
+# from diffusers.models.attention import BasicTransformerBlock
+from diffusers.models.embeddings import CaptionProjection, PatchEmbed
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        if self.is_input_continuous:
+            self.in_channels = in_channels
+
+            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+            if use_linear_projection:
+                self.proj_in = linear_cls(in_channels, inner_dim)
+            else:
+                self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
+
+            self.height = sample_size
+            self.width = sample_size
+            self.num_vector_embeds = num_vector_embeds
+            self.num_latent_pixels = self.height * self.width
+
+            self.latent_image_embedding = ImagePositionalEmbeddings(
+                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
+            )
+        elif self.is_input_patches:
+            assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+            self.height = sample_size
+            self.width = sample_size
+
+            self.patch_size = patch_size
+            interpolation_scale = self.config.sample_size // 64  # => 64 (= 512 pixart) has interpolation scale 1
+            interpolation_scale = max(interpolation_scale, 1)
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+                interpolation_scale=interpolation_scale,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        if self.is_input_continuous:
+            # TODO: should use out_channels for continuous projections
+            if use_linear_projection:
+                self.proj_out = linear_cls(inner_dim, in_channels)
+            else:
+                self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            self.norm_out = nn.LayerNorm(inner_dim)
+            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
+        elif self.is_input_patches and norm_type != "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
+            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+        elif self.is_input_patches and norm_type == "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+            self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        spatial_attn_inputs = [],
+        spatial_attn_idx = 0,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 1. Input
+        if self.is_input_continuous:
+            batch, _, height, width = hidden_states.shape
+            residual = hidden_states
+
+            hidden_states = self.norm(hidden_states)
+            if not self.use_linear_projection:
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+            else:
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+            hidden_states = self.pos_embed(hidden_states)
+
+            if self.adaln_single is not None:
+                if self.use_additional_conditions and added_cond_kwargs is None:
+                    raise ValueError(
+                        "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+                    )
+                batch_size = hidden_states.shape[0]
+                timestep, embedded_timestep = self.adaln_single(
+                    timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+                )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+                hidden_states, spatial_attn_inputs, spatial_attn_idx = torch.utils.checkpoint.checkpoint(
+                    block,
+                    hidden_states,
+                    spatial_attn_inputs,
+                    spatial_attn_idx,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states, spatial_attn_inputs, spatial_attn_idx = block(
+                    hidden_states,
+                    spatial_attn_inputs,
+                    spatial_attn_idx,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+            output = hidden_states + residual
+        elif self.is_input_vectorized:
+            hidden_states = self.norm_out(hidden_states)
+            logits = self.out(hidden_states)
+            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+            logits = logits.permute(0, 2, 1)
+
+            # log(p(x_0))
+            output = F.log_softmax(logits.double(), dim=1).float()
+
+        if self.is_input_patches:
+            if self.config.norm_type != "ada_norm_single":
+                conditioning = self.transformer_blocks[0].norm1.emb(
+                    timestep, class_labels, hidden_dtype=hidden_states.dtype
+                )
+                shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+                hidden_states = self.proj_out_2(hidden_states)
+            elif self.config.norm_type == "ada_norm_single":
+                shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states)
+                # Modulation
+                hidden_states = hidden_states * (1 + scale) + shift
+                hidden_states = self.proj_out(hidden_states)
+                hidden_states = hidden_states.squeeze(1)
+
+            # unpatchify
+            if self.adaln_single is None:
+                height = width = int(hidden_states.shape[1] ** 0.5)
+            hidden_states = hidden_states.reshape(
+                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+            )
+            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+            output = hidden_states.reshape(
+                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+            )
+
+        if not return_dict:
+            return (output,), spatial_attn_inputs, spatial_attn_idx
+
+        return Transformer2DModelOutput(sample=output), spatial_attn_inputs, spatial_attn_idx

ootd/pipelines_ootd/unet_garm_2d_condition.py

@@ -0,0 +1,1183 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from .unet_garm_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    UNetMidBlock2DSimpleCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+# from diffusers.models.unet_2d_blocks import (
+#     UNetMidBlock2D,
+#     UNetMidBlock2DCrossAttn,
+#     UNetMidBlock2DSimpleCrossAttn,
+#     get_down_block,
+#     get_up_block,
+# )
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+class UNetGarm2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                cross_attention_dim=cross_attention_dim[-1],
+                attention_head_dim=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                skip_time_act=resnet_skip_time_act,
+                only_cross_attention=mid_block_only_cross_attention,
+                cross_attention_norm=cross_attention_norm,
+            )
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+        )
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # For Vton
+        spatial_attn_inputs = []
+
+        # 3. down
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples, spatial_attn_inputs = downsample_block(
+                    hidden_states=sample,
+                    spatial_attn_inputs=spatial_attn_inputs,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample, 
+                    temb=emb, 
+                    scale=lora_scale,
+                )
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        # if is_controlnet:
+        #     new_down_block_res_samples = ()
+
+        #     for down_block_res_sample, down_block_additional_residual in zip(
+        #         down_block_res_samples, down_block_additional_residuals
+        #     ):
+        #         down_block_res_sample = down_block_res_sample + down_block_additional_residual
+        #         new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+        #     down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample, spatial_attn_inputs = self.mid_block(
+                    sample,
+                    spatial_attn_inputs=spatial_attn_inputs,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample, spatial_attn_inputs = upsample_block(
+                    hidden_states=sample,
+                    spatial_attn_inputs=spatial_attn_inputs,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,), spatial_attn_inputs
+
+        return UNet2DConditionOutput(sample=sample), spatial_attn_inputs

ootd/pipelines_ootd/unet_vton_2d_condition.py

@@ -0,0 +1,1183 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from .unet_vton_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    UNetMidBlock2DSimpleCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+# from ..diffusers.src.diffusers.models.unet_2d_blocks import (
+#     UNetMidBlock2D,
+#     UNetMidBlock2DCrossAttn,
+#     UNetMidBlock2DSimpleCrossAttn,
+#     get_down_block,
+#     get_up_block,
+# )
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+class UNetVton2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                cross_attention_dim=cross_attention_dim[-1],
+                attention_head_dim=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                skip_time_act=resnet_skip_time_act,
+                only_cross_attention=mid_block_only_cross_attention,
+                cross_attention_norm=cross_attention_norm,
+            )
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+        )
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        spatial_attn_inputs,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # for spatial attention
+        spatial_attn_idx = 0
+
+        # 3. down
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples, spatial_attn_inputs, spatial_attn_idx = downsample_block(
+                    hidden_states=sample,
+                    spatial_attn_inputs=spatial_attn_inputs,
+                    spatial_attn_idx=spatial_attn_idx,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale)
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample, spatial_attn_inputs, spatial_attn_idx = self.mid_block(
+                    sample,
+                    spatial_attn_inputs=spatial_attn_inputs,
+                    spatial_attn_idx=spatial_attn_idx,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample, spatial_attn_inputs, spatial_attn_idx = upsample_block(
+                    hidden_states=sample,
+                    spatial_attn_inputs=spatial_attn_inputs,
+                    spatial_attn_idx=spatial_attn_idx,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

preprocess/humanparsing/datasets/datasets.py

@@ -0,0 +1,201 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   [email protected]
+@File    :   datasets.py
+@Time    :   8/4/19 3:35 PM
+@Desc    :
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import os
+import numpy as np
+import random
+import torch
+import cv2
+from torch.utils import data
+from utils.transforms import get_affine_transform
+
+
+class LIPDataSet(data.Dataset):
+    def __init__(self, root, dataset, crop_size=[473, 473], scale_factor=0.25,
+                 rotation_factor=30, ignore_label=255, transform=None):
+        self.root = root
+        self.aspect_ratio = crop_size[1] * 1.0 / crop_size[0]
+        self.crop_size = np.asarray(crop_size)
+        self.ignore_label = ignore_label
+        self.scale_factor = scale_factor
+        self.rotation_factor = rotation_factor
+        self.flip_prob = 0.5
+        self.transform = transform
+        self.dataset = dataset
+
+        list_path = os.path.join(self.root, self.dataset + '_id.txt')
+        train_list = [i_id.strip() for i_id in open(list_path)]
+
+        self.train_list = train_list
+        self.number_samples = len(self.train_list)
+
+    def __len__(self):
+        return self.number_samples
+
+    def _box2cs(self, box):
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x, y, w, h):
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w * 1.0, h * 1.0], dtype=np.float32)
+        return center, scale
+
+    def __getitem__(self, index):
+        train_item = self.train_list[index]
+
+        im_path = os.path.join(self.root, self.dataset + '_images', train_item + '.jpg')
+        parsing_anno_path = os.path.join(self.root, self.dataset + '_segmentations', train_item + '.png')
+
+        im = cv2.imread(im_path, cv2.IMREAD_COLOR)
+        h, w, _ = im.shape
+        parsing_anno = np.zeros((h, w), dtype=np.long)
+
+        # Get person center and scale
+        person_center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+
+        if self.dataset != 'test':
+            # Get pose annotation
+            parsing_anno = cv2.imread(parsing_anno_path, cv2.IMREAD_GRAYSCALE)
+            if self.dataset == 'train' or self.dataset == 'trainval':
+                sf = self.scale_factor
+                rf = self.rotation_factor
+                s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
+                r = np.clip(np.random.randn() * rf, -rf * 2, rf * 2) if random.random() <= 0.6 else 0
+
+                if random.random() <= self.flip_prob:
+                    im = im[:, ::-1, :]
+                    parsing_anno = parsing_anno[:, ::-1]
+                    person_center[0] = im.shape[1] - person_center[0] - 1
+                    right_idx = [15, 17, 19]
+                    left_idx = [14, 16, 18]
+                    for i in range(0, 3):
+                        right_pos = np.where(parsing_anno == right_idx[i])
+                        left_pos = np.where(parsing_anno == left_idx[i])
+                        parsing_anno[right_pos[0], right_pos[1]] = left_idx[i]
+                        parsing_anno[left_pos[0], left_pos[1]] = right_idx[i]
+
+        trans = get_affine_transform(person_center, s, r, self.crop_size)
+        input = cv2.warpAffine(
+            im,
+            trans,
+            (int(self.crop_size[1]), int(self.crop_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0))
+
+        if self.transform:
+            input = self.transform(input)
+
+        meta = {
+            'name': train_item,
+            'center': person_center,
+            'height': h,
+            'width': w,
+            'scale': s,
+            'rotation': r
+        }
+
+        if self.dataset == 'val' or self.dataset == 'test':
+            return input, meta
+        else:
+            label_parsing = cv2.warpAffine(
+                parsing_anno,
+                trans,
+                (int(self.crop_size[1]), int(self.crop_size[0])),
+                flags=cv2.INTER_NEAREST,
+                borderMode=cv2.BORDER_CONSTANT,
+                borderValue=(255))
+
+            label_parsing = torch.from_numpy(label_parsing)
+
+            return input, label_parsing, meta
+
+
+class LIPDataValSet(data.Dataset):
+    def __init__(self, root, dataset='val', crop_size=[473, 473], transform=None, flip=False):
+        self.root = root
+        self.crop_size = crop_size
+        self.transform = transform
+        self.flip = flip
+        self.dataset = dataset
+        self.root = root
+        self.aspect_ratio = crop_size[1] * 1.0 / crop_size[0]
+        self.crop_size = np.asarray(crop_size)
+
+        list_path = os.path.join(self.root, self.dataset + '_id.txt')
+        val_list = [i_id.strip() for i_id in open(list_path)]
+
+        self.val_list = val_list
+        self.number_samples = len(self.val_list)
+
+    def __len__(self):
+        return len(self.val_list)
+
+    def _box2cs(self, box):
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x, y, w, h):
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w * 1.0, h * 1.0], dtype=np.float32)
+
+        return center, scale
+
+    def __getitem__(self, index):
+        val_item = self.val_list[index]
+        # Load training image
+        im_path = os.path.join(self.root, self.dataset + '_images', val_item + '.jpg')
+        im = cv2.imread(im_path, cv2.IMREAD_COLOR)
+        h, w, _ = im.shape
+        # Get person center and scale
+        person_center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+        trans = get_affine_transform(person_center, s, r, self.crop_size)
+        input = cv2.warpAffine(
+            im,
+            trans,
+            (int(self.crop_size[1]), int(self.crop_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0))
+        input = self.transform(input)
+        flip_input = input.flip(dims=[-1])
+        if self.flip:
+            batch_input_im = torch.stack([input, flip_input])
+        else:
+            batch_input_im = input
+
+        meta = {
+            'name': val_item,
+            'center': person_center,
+            'height': h,
+            'width': w,
+            'scale': s,
+            'rotation': r
+        }
+
+        return batch_input_im, meta

preprocess/humanparsing/datasets/simple_extractor_dataset.py

@@ -0,0 +1,89 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   [email protected]
+@File    :   dataset.py
+@Time    :   8/30/19 9:12 PM
+@Desc    :   Dataset Definition
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import os
+import pdb
+
+import cv2
+import numpy as np
+from PIL import Image
+from torch.utils import data
+from utils.transforms import get_affine_transform
+
+
+class SimpleFolderDataset(data.Dataset):
+    def __init__(self, root, input_size=[512, 512], transform=None):
+        self.root = root
+        self.input_size = input_size
+        self.transform = transform
+        self.aspect_ratio = input_size[1] * 1.0 / input_size[0]
+        self.input_size = np.asarray(input_size)
+        self.is_pil_image = False
+        if isinstance(root, Image.Image):
+            self.file_list = [root]
+            self.is_pil_image = True
+        elif os.path.isfile(root):
+            self.file_list = [os.path.basename(root)]
+            self.root = os.path.dirname(root)
+        else:
+            self.file_list = os.listdir(self.root)
+
+    def __len__(self):
+        return len(self.file_list)
+
+    def _box2cs(self, box):
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x, y, w, h):
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w, h], dtype=np.float32)
+        return center, scale
+
+    def __getitem__(self, index):
+        if self.is_pil_image:
+            img = np.asarray(self.file_list[index])[:, :, [2, 1, 0]]
+        else:
+            img_name = self.file_list[index]
+            img_path = os.path.join(self.root, img_name)
+            img = cv2.imread(img_path, cv2.IMREAD_COLOR)
+        h, w, _ = img.shape
+
+        # Get person center and scale
+        person_center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+        trans = get_affine_transform(person_center, s, r, self.input_size)
+        input = cv2.warpAffine(
+            img,
+            trans,
+            (int(self.input_size[1]), int(self.input_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0))
+
+        input = self.transform(input)
+        meta = {
+            'center': person_center,
+            'height': h,
+            'width': w,
+            'scale': s,
+            'rotation': r
+        }
+
+        return input, meta

preprocess/humanparsing/datasets/target_generation.py

@@ -0,0 +1,40 @@
+import torch
+from torch.nn import functional as F
+
+
+def generate_edge_tensor(label, edge_width=3):
+    label = label.type(torch.cuda.FloatTensor)
+    if len(label.shape) == 2:
+        label = label.unsqueeze(0)
+    n, h, w = label.shape
+    edge = torch.zeros(label.shape, dtype=torch.float).cuda()
+    # right
+    edge_right = edge[:, 1:h, :]
+    edge_right[(label[:, 1:h, :] != label[:, :h - 1, :]) & (label[:, 1:h, :] != 255)
+               & (label[:, :h - 1, :] != 255)] = 1
+
+    # up
+    edge_up = edge[:, :, :w - 1]
+    edge_up[(label[:, :, :w - 1] != label[:, :, 1:w])
+            & (label[:, :, :w - 1] != 255)
+            & (label[:, :, 1:w] != 255)] = 1
+
+    # upright
+    edge_upright = edge[:, :h - 1, :w - 1]
+    edge_upright[(label[:, :h - 1, :w - 1] != label[:, 1:h, 1:w])
+                 & (label[:, :h - 1, :w - 1] != 255)
+                 & (label[:, 1:h, 1:w] != 255)] = 1
+
+    # bottomright
+    edge_bottomright = edge[:, :h - 1, 1:w]
+    edge_bottomright[(label[:, :h - 1, 1:w] != label[:, 1:h, :w - 1])
+                     & (label[:, :h - 1, 1:w] != 255)
+                     & (label[:, 1:h, :w - 1] != 255)] = 1
+
+    kernel = torch.ones((1, 1, edge_width, edge_width), dtype=torch.float).cuda()
+    with torch.no_grad():
+        edge = edge.unsqueeze(1)
+        edge = F.conv2d(edge, kernel, stride=1, padding=1)
+    edge[edge!=0] = 1
+    edge = edge.squeeze()
+    return edge

preprocess/humanparsing/mhp_extension/coco_style_annotation_creator/human_to_coco.py

@@ -0,0 +1,166 @@
+import argparse
+import datetime
+import json
+import os
+from PIL import Image
+import numpy as np
+
+import pycococreatortools
+
+
+def get_arguments():
+    parser = argparse.ArgumentParser(description="transform mask annotation to coco annotation")
+    parser.add_argument("--dataset", type=str, default='CIHP', help="name of dataset (CIHP, MHPv2 or VIP)")
+    parser.add_argument("--json_save_dir", type=str, default='../data/msrcnn_finetune_annotations',
+                        help="path to save coco-style annotation json file")
+    parser.add_argument("--use_val", type=bool, default=False,
+                        help="use train+val set for finetuning or not")
+    parser.add_argument("--train_img_dir", type=str, default='../data/instance-level_human_parsing/Training/Images',
+                        help="train image path")
+    parser.add_argument("--train_anno_dir", type=str,
+                        default='../data/instance-level_human_parsing/Training/Human_ids',
+                        help="train human mask path")
+    parser.add_argument("--val_img_dir", type=str, default='../data/instance-level_human_parsing/Validation/Images',
+                        help="val image path")
+    parser.add_argument("--val_anno_dir", type=str,
+                        default='../data/instance-level_human_parsing/Validation/Human_ids',
+                        help="val human mask path")
+    return parser.parse_args()
+
+
+def main(args):
+    INFO = {
+        "description": args.split_name + " Dataset",
+        "url": "",
+        "version": "",
+        "year": 2019,
+        "contributor": "xyq",
+        "date_created": datetime.datetime.utcnow().isoformat(' ')
+    }
+
+    LICENSES = [
+        {
+            "id": 1,
+            "name": "",
+            "url": ""
+        }
+    ]
+
+    CATEGORIES = [
+        {
+            'id': 1,
+            'name': 'person',
+            'supercategory': 'person',
+        },
+    ]
+
+    coco_output = {
+        "info": INFO,
+        "licenses": LICENSES,
+        "categories": CATEGORIES,
+        "images": [],
+        "annotations": []
+    }
+
+    image_id = 1
+    segmentation_id = 1
+
+    for image_name in os.listdir(args.train_img_dir):
+        image = Image.open(os.path.join(args.train_img_dir, image_name))
+        image_info = pycococreatortools.create_image_info(
+            image_id, image_name, image.size
+        )
+        coco_output["images"].append(image_info)
+
+        human_mask_name = os.path.splitext(image_name)[0] + '.png'
+        human_mask = np.asarray(Image.open(os.path.join(args.train_anno_dir, human_mask_name)))
+        human_gt_labels = np.unique(human_mask)
+
+        for i in range(1, len(human_gt_labels)):
+            category_info = {'id': 1, 'is_crowd': 0}
+            binary_mask = np.uint8(human_mask == i)
+            annotation_info = pycococreatortools.create_annotation_info(
+                segmentation_id, image_id, category_info, binary_mask,
+                image.size, tolerance=10
+            )
+            if annotation_info is not None:
+                coco_output["annotations"].append(annotation_info)
+
+            segmentation_id += 1
+        image_id += 1
+
+    if not os.path.exists(args.json_save_dir):
+        os.makedirs(args.json_save_dir)
+    if not args.use_val:
+        with open('{}/{}_train.json'.format(args.json_save_dir, args.split_name), 'w') as output_json_file:
+            json.dump(coco_output, output_json_file)
+    else:
+        for image_name in os.listdir(args.val_img_dir):
+            image = Image.open(os.path.join(args.val_img_dir, image_name))
+            image_info = pycococreatortools.create_image_info(
+                image_id, image_name, image.size
+            )
+            coco_output["images"].append(image_info)
+
+            human_mask_name = os.path.splitext(image_name)[0] + '.png'
+            human_mask = np.asarray(Image.open(os.path.join(args.val_anno_dir, human_mask_name)))
+            human_gt_labels = np.unique(human_mask)
+
+            for i in range(1, len(human_gt_labels)):
+                category_info = {'id': 1, 'is_crowd': 0}
+                binary_mask = np.uint8(human_mask == i)
+                annotation_info = pycococreatortools.create_annotation_info(
+                    segmentation_id, image_id, category_info, binary_mask,
+                    image.size, tolerance=10
+                )
+                if annotation_info is not None:
+                    coco_output["annotations"].append(annotation_info)
+
+                segmentation_id += 1
+            image_id += 1
+
+        with open('{}/{}_trainval.json'.format(args.json_save_dir, args.split_name), 'w') as output_json_file:
+            json.dump(coco_output, output_json_file)
+
+    coco_output_val = {
+        "info": INFO,
+        "licenses": LICENSES,
+        "categories": CATEGORIES,
+        "images": [],
+        "annotations": []
+    }
+
+    image_id_val = 1
+    segmentation_id_val = 1
+
+    for image_name in os.listdir(args.val_img_dir):
+        image = Image.open(os.path.join(args.val_img_dir, image_name))
+        image_info = pycococreatortools.create_image_info(
+            image_id_val, image_name, image.size
+        )
+        coco_output_val["images"].append(image_info)
+
+        human_mask_name = os.path.splitext(image_name)[0] + '.png'
+        human_mask = np.asarray(Image.open(os.path.join(args.val_anno_dir, human_mask_name)))
+        human_gt_labels = np.unique(human_mask)
+
+        for i in range(1, len(human_gt_labels)):
+            category_info = {'id': 1, 'is_crowd': 0}
+            binary_mask = np.uint8(human_mask == i)
+            annotation_info = pycococreatortools.create_annotation_info(
+                segmentation_id_val, image_id_val, category_info, binary_mask,
+                image.size, tolerance=10
+            )
+            if annotation_info is not None:
+                coco_output_val["annotations"].append(annotation_info)
+
+            segmentation_id_val += 1
+        image_id_val += 1
+
+    with open('{}/{}_val.json'.format(args.json_save_dir, args.split_name), 'w') as output_json_file_val:
+        json.dump(coco_output_val, output_json_file_val)
+
+
+if __name__ == "__main__":
+    args = get_arguments()
+    main(args)

preprocess/humanparsing/mhp_extension/coco_style_annotation_creator/pycococreatortools.py

@@ -0,0 +1,114 @@
+import re
+import datetime
+import numpy as np
+from itertools import groupby
+from skimage import measure
+from PIL import Image
+from pycocotools import mask
+
+convert = lambda text: int(text) if text.isdigit() else text.lower()
+natrual_key = lambda key: [convert(c) for c in re.split('([0-9]+)', key)]
+
+
+def resize_binary_mask(array, new_size):
+    image = Image.fromarray(array.astype(np.uint8) * 255)
+    image = image.resize(new_size)
+    return np.asarray(image).astype(np.bool_)
+
+
+def close_contour(contour):
+    if not np.array_equal(contour[0], contour[-1]):
+        contour = np.vstack((contour, contour[0]))
+    return contour
+
+
+def binary_mask_to_rle(binary_mask):
+    rle = {'counts': [], 'size': list(binary_mask.shape)}
+    counts = rle.get('counts')
+    for i, (value, elements) in enumerate(groupby(binary_mask.ravel(order='F'))):
+        if i == 0 and value == 1:
+            counts.append(0)
+        counts.append(len(list(elements)))
+
+    return rle
+
+
+def binary_mask_to_polygon(binary_mask, tolerance=0):
+    """Converts a binary mask to COCO polygon representation
+    Args:
+        binary_mask: a 2D binary numpy array where '1's represent the object
+        tolerance: Maximum distance from original points of polygon to approximated
+            polygonal chain. If tolerance is 0, the original coordinate array is returned.
+    """
+    polygons = []
+    # pad mask to close contours of shapes which start and end at an edge
+    padded_binary_mask = np.pad(binary_mask, pad_width=1, mode='constant', constant_values=0)
+    contours = measure.find_contours(padded_binary_mask, 0.5)
+    contours = np.subtract(contours, 1)
+    for contour in contours:
+        contour = close_contour(contour)
+        contour = measure.approximate_polygon(contour, tolerance)
+        if len(contour) < 3:
+            continue
+        contour = np.flip(contour, axis=1)
+        segmentation = contour.ravel().tolist()
+        # after padding and subtracting 1 we may get -0.5 points in our segmentation 
+        segmentation = [0 if i < 0 else i for i in segmentation]
+        polygons.append(segmentation)
+
+    return polygons
+
+
+def create_image_info(image_id, file_name, image_size,
+                      date_captured=datetime.datetime.utcnow().isoformat(' '),
+                      license_id=1, coco_url="", flickr_url=""):
+    image_info = {
+        "id": image_id,
+        "file_name": file_name,
+        "width": image_size[0],
+        "height": image_size[1],
+        "date_captured": date_captured,
+        "license": license_id,
+        "coco_url": coco_url,
+        "flickr_url": flickr_url
+    }
+
+    return image_info
+
+
+def create_annotation_info(annotation_id, image_id, category_info, binary_mask,
+                           image_size=None, tolerance=2, bounding_box=None):
+    if image_size is not None:
+        binary_mask = resize_binary_mask(binary_mask, image_size)
+
+    binary_mask_encoded = mask.encode(np.asfortranarray(binary_mask.astype(np.uint8)))
+
+    area = mask.area(binary_mask_encoded)
+    if area < 1:
+        return None
+
+    if bounding_box is None:
+        bounding_box = mask.toBbox(binary_mask_encoded)
+
+    if category_info["is_crowd"]:
+        is_crowd = 1
+        segmentation = binary_mask_to_rle(binary_mask)
+    else:
+        is_crowd = 0
+        segmentation = binary_mask_to_polygon(binary_mask, tolerance)
+        if not segmentation:
+            return None
+
+    annotation_info = {
+        "id": annotation_id,
+        "image_id": image_id,
+        "category_id": category_info["id"],
+        "iscrowd": is_crowd,
+        "area": area.tolist(),
+        "bbox": bounding_box.tolist(),
+        "segmentation": segmentation,
+        "width": binary_mask.shape[1],
+        "height": binary_mask.shape[0],
+    }
+
+    return annotation_info

preprocess/humanparsing/mhp_extension/coco_style_annotation_creator/test_human2coco_format.py

@@ -0,0 +1,74 @@
+import argparse
+import datetime
+import json
+import os
+from PIL import Image
+
+import pycococreatortools
+
+
+def get_arguments():
+    parser = argparse.ArgumentParser(description="transform mask annotation to coco annotation")
+    parser.add_argument("--dataset", type=str, default='CIHP', help="name of dataset (CIHP, MHPv2 or VIP)")
+    parser.add_argument("--json_save_dir", type=str, default='../data/CIHP/annotations',
+                        help="path to save coco-style annotation json file")
+    parser.add_argument("--test_img_dir", type=str, default='../data/CIHP/Testing/Images',
+                        help="test image path")
+    return parser.parse_args()
+
+args = get_arguments()
+
+INFO = {
+    "description": args.dataset + "Dataset",
+    "url": "",
+    "version": "",
+    "year": 2020,
+    "contributor": "yunqiuxu",
+    "date_created": datetime.datetime.utcnow().isoformat(' ')
+}
+
+LICENSES = [
+    {
+        "id": 1,
+        "name": "",
+        "url": ""
+    }
+]
+
+CATEGORIES = [
+    {
+        'id': 1,
+        'name': 'person',
+        'supercategory': 'person',
+    },
+]
+
+
+def main(args):
+    coco_output = {
+        "info": INFO,
+        "licenses": LICENSES,
+        "categories": CATEGORIES,
+        "images": [],
+        "annotations": []
+    }
+
+    image_id = 1
+
+    for image_name in os.listdir(args.test_img_dir):
+        image = Image.open(os.path.join(args.test_img_dir, image_name))
+        image_info = pycococreatortools.create_image_info(
+            image_id, image_name, image.size
+        )
+        coco_output["images"].append(image_info)
+        image_id += 1
+
+    if not os.path.exists(os.path.join(args.json_save_dir)):
+        os.mkdir(os.path.join(args.json_save_dir))
+
+    with open('{}/{}.json'.format(args.json_save_dir, args.dataset), 'w') as output_json_file:
+        json.dump(coco_output, output_json_file)
+
+
+if __name__ == "__main__":
+    main(args)

preprocess/humanparsing/mhp_extension/detectron2/.circleci/config.yml

@@ -0,0 +1,179 @@
+# Python CircleCI 2.0 configuration file
+#
+# Check https://circleci.com/docs/2.0/language-python/ for more details
+#
+version: 2
+
+# -------------------------------------------------------------------------------------
+# Environments to run the jobs in
+# -------------------------------------------------------------------------------------
+cpu: &cpu
+  docker:
+    - image: circleci/python:3.6.8-stretch
+  resource_class: medium
+
+gpu: &gpu
+  machine:
+    image: ubuntu-1604:201903-01
+    docker_layer_caching: true
+  resource_class: gpu.small
+
+# -------------------------------------------------------------------------------------
+# Re-usable commands
+# -------------------------------------------------------------------------------------
+install_python: &install_python
+  - run:
+      name: Install Python
+      working_directory: ~/
+      command: |
+        pyenv install 3.6.1
+        pyenv global 3.6.1
+
+setup_venv: &setup_venv
+  - run:
+      name: Setup Virtual Env
+      working_directory: ~/
+      command: |
+        python -m venv ~/venv
+        echo ". ~/venv/bin/activate" >> $BASH_ENV
+        . ~/venv/bin/activate
+        python --version
+        which python
+        which pip
+        pip install --upgrade pip
+
+install_dep: &install_dep
+  - run:
+      name: Install Dependencies
+      command: |
+        pip install --progress-bar off -U 'git+https://github.com/facebookresearch/fvcore'
+        pip install --progress-bar off cython opencv-python
+        pip install --progress-bar off 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
+        pip install --progress-bar off torch torchvision
+
+install_detectron2: &install_detectron2
+  - run:
+      name: Install Detectron2
+      command: |
+        gcc --version
+        pip install -U --progress-bar off -e .[dev]
+        python -m detectron2.utils.collect_env
+
+install_nvidia_driver: &install_nvidia_driver
+  - run:
+      name: Install nvidia driver
+      working_directory: ~/
+      command: |
+        wget -q 'https://s3.amazonaws.com/ossci-linux/nvidia_driver/NVIDIA-Linux-x86_64-430.40.run'
+        sudo /bin/bash ./NVIDIA-Linux-x86_64-430.40.run -s --no-drm
+        nvidia-smi
+
+run_unittests: &run_unittests
+  - run:
+      name: Run Unit Tests
+      command: |
+        python -m unittest discover -v -s tests
+
+# -------------------------------------------------------------------------------------
+# Jobs to run
+# -------------------------------------------------------------------------------------
+jobs:
+  cpu_tests:
+    <<: *cpu
+
+    working_directory: ~/detectron2
+
+    steps:
+      - checkout
+      - <<: *setup_venv
+
+      # Cache the venv directory that contains dependencies
+      - restore_cache:
+          keys:
+            - cache-key-{{ .Branch }}-ID-20200425
+
+      - <<: *install_dep
+
+      - save_cache:
+          paths:
+            - ~/venv
+          key: cache-key-{{ .Branch }}-ID-20200425
+
+      - <<: *install_detectron2
+
+      - run:
+          name: isort
+          command: |
+            isort -c -sp .
+      - run:
+          name: black
+          command: |
+            black --check -l 100 .
+      - run:
+          name: flake8
+          command: |
+            flake8 .
+
+      - <<: *run_unittests
+
+  gpu_tests:
+    <<: *gpu
+
+    working_directory: ~/detectron2
+
+    steps:
+      - checkout
+      - <<: *install_nvidia_driver
+
+      - run:
+          name: Install nvidia-docker
+          working_directory: ~/
+          command: |
+            curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
+            distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
+            curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | \
+            sudo tee /etc/apt/sources.list.d/nvidia-docker.list
+            sudo apt-get update && sudo apt-get install -y nvidia-docker2
+            # reload the docker daemon configuration
+            sudo pkill -SIGHUP dockerd
+
+      - run:
+          name: Launch docker
+          working_directory: ~/detectron2/docker
+          command: |
+            nvidia-docker build -t detectron2:v0 -f Dockerfile-circleci .
+            nvidia-docker run -itd --name d2 detectron2:v0
+            docker exec -it d2 nvidia-smi
+
+      - run:
+          name: Build Detectron2
+          command: |
+            docker exec -it d2 pip install 'git+https://github.com/facebookresearch/fvcore'
+            docker cp ~/detectron2 d2:/detectron2
+            # This will build d2 for the target GPU arch only
+            docker exec -it d2 pip install -e /detectron2
+            docker exec -it d2 python3 -m detectron2.utils.collect_env
+            docker exec -it d2 python3 -c 'import torch; assert(torch.cuda.is_available())'
+
+      - run:
+          name: Run Unit Tests
+          command: |
+            docker exec -e CIRCLECI=true -it d2 python3 -m unittest discover -v -s /detectron2/tests
+
+workflows:
+  version: 2
+  regular_test:
+    jobs:
+      - cpu_tests
+      - gpu_tests
+
+  #nightly_test:
+    #jobs:
+      #- gpu_tests
+    #triggers:
+      #- schedule:
+          #cron: "0 0 * * *"
+          #filters:
+            #branches:
+              #only:
+                #- master

preprocess/humanparsing/mhp_extension/detectron2/.clang-format

@@ -0,0 +1,85 @@
+AccessModifierOffset: -1
+AlignAfterOpenBracket: AlwaysBreak
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+AlignEscapedNewlinesLeft: true
+AlignOperands:   false
+AlignTrailingComments: false
+AllowAllParametersOfDeclarationOnNextLine: false
+AllowShortBlocksOnASingleLine: false
+AllowShortCaseLabelsOnASingleLine: false
+AllowShortFunctionsOnASingleLine: Empty
+AllowShortIfStatementsOnASingleLine: false
+AllowShortLoopsOnASingleLine: false
+AlwaysBreakAfterReturnType: None
+AlwaysBreakBeforeMultilineStrings: true
+AlwaysBreakTemplateDeclarations: true
+BinPackArguments: false
+BinPackParameters: false
+BraceWrapping:
+  AfterClass:      false
+  AfterControlStatement: false
+  AfterEnum:       false
+  AfterFunction:   false
+  AfterNamespace:  false
+  AfterObjCDeclaration: false
+  AfterStruct:     false
+  AfterUnion:      false
+  BeforeCatch:     false
+  BeforeElse:      false
+  IndentBraces:    false
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Attach
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: false
+BreakAfterJavaFieldAnnotations: false
+BreakStringLiterals: false
+ColumnLimit:     80
+CommentPragmas:  '^ IWYU pragma:'
+ConstructorInitializerAllOnOneLineOrOnePerLine: true
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: true
+DerivePointerAlignment: false
+DisableFormat:   false
+ForEachMacros:   [ FOR_EACH, FOR_EACH_ENUMERATE, FOR_EACH_KV, FOR_EACH_R, FOR_EACH_RANGE, ]
+IncludeCategories:
+  - Regex:           '^<.*\.h(pp)?>'
+    Priority:        1
+  - Regex:           '^<.*'
+    Priority:        2
+  - Regex:           '.*'
+    Priority:        3
+IndentCaseLabels: true
+IndentWidth:     2
+IndentWrappedFunctionNames: false
+KeepEmptyLinesAtTheStartOfBlocks: false
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: false
+PenaltyBreakBeforeFirstCallParameter: 1
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Left
+ReflowComments:  true
+SortIncludes:    true
+SpaceAfterCStyleCast: false
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeParens: ControlStatements
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 1
+SpacesInAngles:  false
+SpacesInContainerLiterals: true
+SpacesInCStyleCastParentheses: false
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+Standard:        Cpp11
+TabWidth:        8
+UseTab:          Never

preprocess/humanparsing/mhp_extension/detectron2/.flake8

@@ -0,0 +1,9 @@
+# This is an example .flake8 config, used when developing *Black* itself.
+# Keep in sync with setup.cfg which is used for source packages.
+
+[flake8]
+ignore = W503, E203, E221, C901, C408, E741
+max-line-length = 100
+max-complexity = 18
+select = B,C,E,F,W,T4,B9
+exclude = build,__init__.py

preprocess/humanparsing/mhp_extension/detectron2/.github/CODE_OF_CONDUCT.md

@@ -0,0 +1,5 @@
+# Code of Conduct
+
+Facebook has adopted a Code of Conduct that we expect project participants to adhere to.
+Please read the [full text](https://code.fb.com/codeofconduct/)
+so that you can understand what actions will and will not be tolerated.

preprocess/humanparsing/mhp_extension/detectron2/.github/CONTRIBUTING.md

@@ -0,0 +1,49 @@
+# Contributing to detectron2
+
+## Issues
+We use GitHub issues to track public bugs and questions.
+Please make sure to follow one of the
+[issue templates](https://github.com/facebookresearch/detectron2/issues/new/choose)
+when reporting any issues.
+
+Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+However, if you're adding any significant features (e.g. > 50 lines), please
+make sure to have a corresponding issue to discuss your motivation and proposals,
+before sending a PR. We do not always accept new features, and we take the following
+factors into consideration:
+
+1. Whether the same feature can be achieved without modifying detectron2.
+Detectron2 is designed so that you can implement many extensions from the outside, e.g.
+those in [projects](https://github.com/facebookresearch/detectron2/tree/master/projects).
+If some part is not as extensible, you can also bring up the issue to make it more extensible.
+2. Whether the feature is potentially useful to a large audience, or only to a small portion of users.
+3. Whether the proposed solution has a good design / interface.
+4. Whether the proposed solution adds extra mental/practical overhead to users who don't
+   need such feature.
+5. Whether the proposed solution breaks existing APIs.
+
+When sending a PR, please do:
+
+1. If a PR contains multiple orthogonal changes, split it to several PRs.
+2. If you've added code that should be tested, add tests.
+3. For PRs that need experiments (e.g. adding a new model or new methods),
+	 you don't need to update model zoo, but do provide experiment results in the description of the PR.
+4. If APIs are changed, update the documentation.
+5. Make sure your code lints with `./dev/linter.sh`.
+
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## License
+By contributing to detectron2, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE.md

@@ -0,0 +1,5 @@
+
+Please select an issue template from
+https://github.com/facebookresearch/detectron2/issues/new/choose .
+
+Otherwise your issue will be closed.

preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/bugs.md

@@ -0,0 +1,36 @@
+---
+name: "🐛 Bugs"
+about: Report bugs in detectron2
+title: Please read & provide the following
+
+---
+
+## Instructions To Reproduce the 🐛 Bug:
+
+1. what changes you made (`git diff`) or what code you wrote
+```
+<put diff or code here>
+```
+2. what exact command you run:
+3. what you observed (including __full logs__):
+```
+<put logs here>
+```
+4. please simplify the steps as much as possible so they do not require additional resources to
+	 run, such as a private dataset.
+
+## Expected behavior:
+
+If there are no obvious error in "what you observed" provided above,
+please tell us the expected behavior.
+
+## Environment:
+
+Provide your environment information using the following command:
+```
+wget -nc -q https://github.com/facebookresearch/detectron2/raw/master/detectron2/utils/collect_env.py && python collect_env.py
+```
+
+If your issue looks like an installation issue / environment issue,
+please first try to solve it yourself with the instructions in
+https://detectron2.readthedocs.io/tutorials/install.html#common-installation-issues

preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,9 @@
+# require an issue template to be chosen
+blank_issues_enabled: false
+
+# Unexpected behaviors & bugs are split to two templates.
+# When they are one template, users think "it's not a bug" and don't choose the template.
+#
+# But the file name is still "unexpected-problems-bugs.md" so that old references
+# to this issue template still works.
+# It's ok since this template should be a superset of "bugs.md" (unexpected behaviors is a superset of bugs)

preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/feature-request.md

@@ -0,0 +1,31 @@
+---
+name: "\U0001F680Feature Request"
+about: Submit a proposal/request for a new detectron2 feature
+
+---
+
+## 🚀 Feature
+A clear and concise description of the feature proposal.
+
+
+## Motivation & Examples
+
+Tell us why the feature is useful.
+
+Describe what the feature would look like, if it is implemented.
+Best demonstrated using **code examples** in addition to words.
+
+## Note
+
+We only consider adding new features if they are relevant to many users.
+
+If you request implementation of research papers --
+we only consider papers that have enough significance and prevalance in the object detection field.
+
+We do not take requests for most projects in the `projects/` directory,
+because they are research code release that is mainly for other researchers to reproduce results.
+
+Instead of adding features inside detectron2,
+you can implement many features by [extending detectron2](https://detectron2.readthedocs.io/tutorials/extend.html).
+The [projects/](https://github.com/facebookresearch/detectron2/tree/master/projects/) directory contains many of such examples.
+

preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/questions-help-support.md

@@ -0,0 +1,26 @@
+---
+name: "❓How to do something?"
+about: How to do something using detectron2? What does an API do?
+
+---
+
+## ❓ How to do something using detectron2
+
+Describe what you want to do, including:
+1. what inputs you will provide, if any:
+2. what outputs you are expecting:
+
+## ❓ What does an API do and how to use it?
+Please link to which API or documentation you're asking about from
+https://detectron2.readthedocs.io/
+
+
+NOTE:
+
+1. Only general answers are provided.
+   If you want to ask about "why X did not work", please use the
+   [Unexpected behaviors](https://github.com/facebookresearch/detectron2/issues/new/choose) issue template.
+
+2. About how to implement new models / new dataloader / new training logic, etc., check documentation first.
+
+3. We do not answer general machine learning / computer vision questions that are not specific to detectron2, such as how a model works, how to improve your training/make it converge, or what algorithm/methods can be used to achieve X.

preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/unexpected-problems-bugs.md

@@ -0,0 +1,45 @@
+---
+name: "Unexpected behaviors"
+about: Run into unexpected behaviors when using detectron2
+title: Please read & provide the following
+
+---
+
+If you do not know the root cause of the problem, and wish someone to help you, please
+post according to this template:
+
+## Instructions To Reproduce the Issue:
+
+1. what changes you made (`git diff`) or what code you wrote
+```
+<put diff or code here>
+```
+2. what exact command you run:
+3. what you observed (including __full logs__):
+```
+<put logs here>
+```
+4. please simplify the steps as much as possible so they do not require additional resources to
+	 run, such as a private dataset.
+
+## Expected behavior:
+
+If there are no obvious error in "what you observed" provided above,
+please tell us the expected behavior.
+
+If you expect the model to converge / work better, note that we do not give suggestions
+on how to train a new model.
+Only in one of the two conditions we will help with it:
+(1) You're unable to reproduce the results in detectron2 model zoo.
+(2) It indicates a detectron2 bug.
+
+## Environment:
+
+Provide your environment information using the following command:
+```
+wget -nc -q https://github.com/facebookresearch/detectron2/raw/master/detectron2/utils/collect_env.py && python collect_env.py
+```
+
+If your issue looks like an installation issue / environment issue,
+please first try to solve it yourself with the instructions in
+https://detectron2.readthedocs.io/tutorials/install.html#common-installation-issues

preprocess/humanparsing/mhp_extension/detectron2/.github/pull_request_template.md

@@ -0,0 +1,9 @@
+Thanks for your contribution!
+
+If you're sending a large PR (e.g., >50 lines),
+please open an issue first about the feature / bug, and indicate how you want to contribute.
+
+Before submitting a PR, please run `dev/linter.sh` to lint the code.
+
+See https://detectron2.readthedocs.io/notes/contributing.html#pull-requests
+about how we handle PRs.

preprocess/humanparsing/mhp_extension/detectron2/.gitignore

@@ -0,0 +1,46 @@
+# output dir
+output
+instant_test_output
+inference_test_output
+
+
+*.jpg
+*.png
+*.txt
+*.json
+*.diff
+
+# compilation and distribution
+__pycache__
+_ext
+*.pyc
+*.so
+detectron2.egg-info/
+build/
+dist/
+wheels/
+
+# pytorch/python/numpy formats
+*.pth
+*.pkl
+*.npy
+
+# ipython/jupyter notebooks
+*.ipynb
+**/.ipynb_checkpoints/
+
+# Editor temporaries
+*.swn
+*.swo
+*.swp
+*~
+
+# editor settings
+.idea
+.vscode
+
+# project dirs
+/detectron2/model_zoo/configs
+/datasets
+/projects/*/datasets
+/models

preprocess/humanparsing/mhp_extension/detectron2/GETTING_STARTED.md

@@ -0,0 +1,79 @@
+## Getting Started with Detectron2
+
+This document provides a brief intro of the usage of builtin command-line tools in detectron2.
+
+For a tutorial that involves actual coding with the API,
+see our [Colab Notebook](https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5)
+which covers how to run inference with an
+existing model, and how to train a builtin model on a custom dataset.
+
+For more advanced tutorials, refer to our [documentation](https://detectron2.readthedocs.io/tutorials/extend.html).
+
+
+### Inference Demo with Pre-trained Models
+
+1. Pick a model and its config file from
+	[model zoo](MODEL_ZOO.md),
+	for example, `mask_rcnn_R_50_FPN_3x.yaml`.
+2. We provide `demo.py` that is able to run builtin standard models. Run it with:
+```
+cd demo/
+python demo.py --config-file ../configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml \
+  --input input1.jpg input2.jpg \
+  [--other-options]
+  --opts MODEL.WEIGHTS detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl
+```
+The configs are made for training, therefore we need to specify `MODEL.WEIGHTS` to a model from model zoo for evaluation.
+This command will run the inference and show visualizations in an OpenCV window.
+
+For details of the command line arguments, see `demo.py -h` or look at its source code
+to understand its behavior. Some common arguments are:
+* To run __on your webcam__, replace `--input files` with `--webcam`.
+* To run __on a video__, replace `--input files` with `--video-input video.mp4`.
+* To run __on cpu__, add `MODEL.DEVICE cpu` after `--opts`.
+* To save outputs to a directory (for images) or a file (for webcam or video), use `--output`.
+
+
+### Training & Evaluation in Command Line
+
+We provide a script in "tools/{,plain_}train_net.py", that is made to train
+all the configs provided in detectron2.
+You may want to use it as a reference to write your own training script.
+
+To train a model with "train_net.py", first
+setup the corresponding datasets following
+[datasets/README.md](./datasets/README.md),
+then run:
+```
+cd tools/
+./train_net.py --num-gpus 8 \
+	--config-file ../configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml
+```
+
+The configs are made for 8-GPU training.
+To train on 1 GPU, you may need to [change some parameters](https://arxiv.org/abs/1706.02677), e.g.:
+```
+./train_net.py \
+	--config-file ../configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml \
+	--num-gpus 1 SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.0025
+```
+
+For most models, CPU training is not supported.
+
+To evaluate a model's performance, use
+```
+./train_net.py \
+	--config-file ../configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml \
+	--eval-only MODEL.WEIGHTS /path/to/checkpoint_file
+```
+For more options, see `./train_net.py -h`.
+
+### Use Detectron2 APIs in Your Code
+
+See our [Colab Notebook](https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5)
+to learn how to use detectron2 APIs to:
+1. run inference with an existing model
+2. train a builtin model on a custom dataset
+
+See [detectron2/projects](https://github.com/facebookresearch/detectron2/tree/master/projects)
+for more ways to build your project on detectron2.

preprocess/humanparsing/mhp_extension/detectron2/INSTALL.md

@@ -0,0 +1,184 @@
+## Installation
+
+Our [Colab Notebook](https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5)
+has step-by-step instructions that install detectron2.
+The [Dockerfile](docker)
+also installs detectron2 with a few simple commands.
+
+### Requirements
+- Linux or macOS with Python ≥ 3.6
+- PyTorch ≥ 1.4
+- [torchvision](https://github.com/pytorch/vision/) that matches the PyTorch installation.
+	You can install them together at [pytorch.org](https://pytorch.org) to make sure of this.
+- OpenCV, optional, needed by demo and visualization
+- pycocotools: `pip install cython; pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'`
+
+
+### Build Detectron2 from Source
+
+gcc & g++ ≥ 5 are required. [ninja](https://ninja-build.org/) is recommended for faster build.
+After having them, run:
+```
+python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
+# (add --user if you don't have permission)
+
+# Or, to install it from a local clone:
+git clone https://github.com/facebookresearch/detectron2.git
+python -m pip install -e detectron2
+
+# Or if you are on macOS
+# CC=clang CXX=clang++ python -m pip install -e .
+```
+
+To __rebuild__ detectron2 that's built from a local clone, use `rm -rf build/ **/*.so` to clean the
+old build first. You often need to rebuild detectron2 after reinstalling PyTorch.
+
+### Install Pre-Built Detectron2 (Linux only)
+```
+# for CUDA 10.1:
+python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/index.html
+```
+You can replace cu101 with "cu{100,92}" or "cpu".
+
+Note that:
+1. Such installation has to be used with certain version of official PyTorch release.
+   See [releases](https://github.com/facebookresearch/detectron2/releases) for requirements.
+   It will not work with a different version of PyTorch or a non-official build of PyTorch.
+2. Such installation is out-of-date w.r.t. master branch of detectron2. It may not be
+	 compatible with the master branch of a research project that uses detectron2 (e.g. those in
+	 [projects](projects) or [meshrcnn](https://github.com/facebookresearch/meshrcnn/)).
+
+### Common Installation Issues
+
+If you met issues using the pre-built detectron2, please uninstall it and try building it from source.
+
+Click each issue for its solutions:
+
+<details>
+<summary>
+Undefined torch/aten/caffe2 symbols, or segmentation fault immediately when running the library.
+</summary>
+<br/>
+
+This usually happens when detectron2 or torchvision is not
+compiled with the version of PyTorch you're running.
+
+Pre-built torchvision or detectron2 has to work with the corresponding official release of pytorch.
+If the error comes from a pre-built torchvision, uninstall torchvision and pytorch and reinstall them
+following [pytorch.org](http://pytorch.org). So the versions will match.
+
+If the error comes from a pre-built detectron2, check [release notes](https://github.com/facebookresearch/detectron2/releases)
+to see the corresponding pytorch version required for each pre-built detectron2.
+
+If the error comes from detectron2 or torchvision that you built manually from source,
+remove files you built (`build/`, `**/*.so`) and rebuild it so it can pick up the version of pytorch currently in your environment.
+
+If you cannot resolve this problem, please include the output of `gdb -ex "r" -ex "bt" -ex "quit" --args python -m detectron2.utils.collect_env`
+in your issue.
+</details>
+
+<details>
+<summary>
+Undefined C++ symbols (e.g. `GLIBCXX`) or C++ symbols not found.
+</summary>
+<br/>
+Usually it's because the library is compiled with a newer C++ compiler but run with an old C++ runtime.
+
+This often happens with old anaconda.
+Try `conda update libgcc`. Then rebuild detectron2.
+
+The fundamental solution is to run the code with proper C++ runtime.
+One way is to use `LD_PRELOAD=/path/to/libstdc++.so`.
+
+</details>
+
+<details>
+<summary>
+"Not compiled with GPU support" or "Detectron2 CUDA Compiler: not available".
+</summary>
+<br/>
+CUDA is not found when building detectron2.
+You should make sure
+
+```
+python -c 'import torch; from torch.utils.cpp_extension import CUDA_HOME; print(torch.cuda.is_available(), CUDA_HOME)'
+```
+
+print valid outputs at the time you build detectron2.
+
+Most models can run inference (but not training) without GPU support. To use CPUs, set `MODEL.DEVICE='cpu'` in the config.
+</details>
+
+<details>
+<summary>
+"invalid device function" or "no kernel image is available for execution".
+</summary>
+<br/>
+Two possibilities:
+
+* You build detectron2 with one version of CUDA but run it with a different version.
+
+  To check whether it is the case,
+  use `python -m detectron2.utils.collect_env` to find out inconsistent CUDA versions.
+	In the output of this command, you should expect "Detectron2 CUDA Compiler", "CUDA_HOME", "PyTorch built with - CUDA"
+	to contain cuda libraries of the same version.
+
+	When they are inconsistent,
+	you need to either install a different build of PyTorch (or build by yourself)
+	to match your local CUDA installation, or install a different version of CUDA to match PyTorch.
+
+* Detectron2 or PyTorch/torchvision is not built for the correct GPU architecture (compute compatibility).
+
+	The GPU architecture for PyTorch/detectron2/torchvision is available in the "architecture flags" in
+	`python -m detectron2.utils.collect_env`.
+
+	The GPU architecture flags of detectron2/torchvision by default matches the GPU model detected
+	during compilation. This means the compiled code may not work on a different GPU model.
+	To overwrite the GPU architecture for detectron2/torchvision, use `TORCH_CUDA_ARCH_LIST` environment variable during compilation.
+
+	For example, `export TORCH_CUDA_ARCH_LIST=6.0,7.0` makes it compile for both P100s and V100s.
+	Visit [developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus) to find out
+	the correct compute compatibility number for your device.
+
+</details>
+
+<details>
+<summary>
+Undefined CUDA symbols; cannot open libcudart.so; other nvcc failures.
+</summary>
+<br/>
+The version of NVCC you use to build detectron2 or torchvision does
+not match the version of CUDA you are running with.
+This often happens when using anaconda's CUDA runtime.
+
+Use `python -m detectron2.utils.collect_env` to find out inconsistent CUDA versions.
+In the output of this command, you should expect "Detectron2 CUDA Compiler", "CUDA_HOME", "PyTorch built with - CUDA"
+to contain cuda libraries of the same version.
+
+When they are inconsistent,
+you need to either install a different build of PyTorch (or build by yourself)
+to match your local CUDA installation, or install a different version of CUDA to match PyTorch.
+</details>
+
+
+<details>
+<summary>
+"ImportError: cannot import name '_C'".
+</summary>
+<br/>
+Please build and install detectron2 following the instructions above.
+
+If you are running code from detectron2's root directory, `cd` to a different one.
+Otherwise you may not import the code that you installed.
+</details>
+
+<details>
+<summary>
+ONNX conversion segfault after some "TraceWarning".
+</summary>
+<br/>
+The ONNX package is compiled with too old compiler.
+
+Please build and install ONNX from its source code using a compiler
+whose version is closer to what's used by PyTorch (available in `torch.__config__.show()`).
+</details>

preprocess/humanparsing/mhp_extension/detectron2/LICENSE

@@ -0,0 +1,201 @@
+Apache License
+Version 2.0, January 2004
+http://www.apache.org/licenses/
+
+TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+1. Definitions.
+
+"License" shall mean the terms and conditions for use, reproduction,
+and distribution as defined by Sections 1 through 9 of this document.
+
+"Licensor" shall mean the copyright owner or entity authorized by
+the copyright owner that is granting the License.
+
+"Legal Entity" shall mean the union of the acting entity and all
+other entities that control, are controlled by, or are under common
+control with that entity. For the purposes of this definition,
+"control" means (i) the power, direct or indirect, to cause the
+direction or management of such entity, whether by contract or
+otherwise, or (ii) ownership of fifty percent (50%) or more of the
+outstanding shares, or (iii) beneficial ownership of such entity.
+
+"You" (or "Your") shall mean an individual or Legal Entity
+exercising permissions granted by this License.
+
+"Source" form shall mean the preferred form for making modifications,
+including but not limited to software source code, documentation
+source, and configuration files.
+
+"Object" form shall mean any form resulting from mechanical
+transformation or translation of a Source form, including but
+not limited to compiled object code, generated documentation,
+and conversions to other media types.
+
+"Work" shall mean the work of authorship, whether in Source or
+Object form, made available under the License, as indicated by a
+copyright notice that is included in or attached to the work
+(an example is provided in the Appendix below).
+
+"Derivative Works" shall mean any work, whether in Source or Object
+form, that is based on (or derived from) the Work and for which the
+editorial revisions, annotations, elaborations, or other modifications
+represent, as a whole, an original work of authorship. For the purposes
+of this License, Derivative Works shall not include works that remain
+separable from, or merely link (or bind by name) to the interfaces of,
+the Work and Derivative Works thereof.
+
+"Contribution" shall mean any work of authorship, including
+the original version of the Work and any modifications or additions
+to that Work or Derivative Works thereof, that is intentionally
+submitted to Licensor for inclusion in the Work by the copyright owner
+or by an individual or Legal Entity authorized to submit on behalf of
+the copyright owner. For the purposes of this definition, "submitted"
+means any form of electronic, verbal, or written communication sent
+to the Licensor or its representatives, including but not limited to
+communication on electronic mailing lists, source code control systems,
+and issue tracking systems that are managed by, or on behalf of, the
+Licensor for the purpose of discussing and improving the Work, but
+excluding communication that is conspicuously marked or otherwise
+designated in writing by the copyright owner as "Not a Contribution."
+
+"Contributor" shall mean Licensor and any individual or Legal Entity
+on behalf of whom a Contribution has been received by Licensor and
+subsequently incorporated within the Work.
+
+2. Grant of Copyright License. Subject to the terms and conditions of
+this License, each Contributor hereby grants to You a perpetual,
+worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+copyright license to reproduce, prepare Derivative Works of,
+publicly display, publicly perform, sublicense, and distribute the
+Work and such Derivative Works in Source or Object form.
+
+3. Grant of Patent License. Subject to the terms and conditions of
+this License, each Contributor hereby grants to You a perpetual,
+worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+(except as stated in this section) patent license to make, have made,
+use, offer to sell, sell, import, and otherwise transfer the Work,
+where such license applies only to those patent claims licensable
+by such Contributor that are necessarily infringed by their
+Contribution(s) alone or by combination of their Contribution(s)
+with the Work to which such Contribution(s) was submitted. If You
+institute patent litigation against any entity (including a
+cross-claim or counterclaim in a lawsuit) alleging that the Work
+or a Contribution incorporated within the Work constitutes direct
+or contributory patent infringement, then any patent licenses
+granted to You under this License for that Work shall terminate
+as of the date such litigation is filed.
+
+4. Redistribution. You may reproduce and distribute copies of the
+Work or Derivative Works thereof in any medium, with or without
+modifications, and in Source or Object form, provided that You
+meet the following conditions:
+
+(a) You must give any other recipients of the Work or
+Derivative Works a copy of this License; and
+
+(b) You must cause any modified files to carry prominent notices
+stating that You changed the files; and
+
+(c) You must retain, in the Source form of any Derivative Works
+that You distribute, all copyright, patent, trademark, and
+attribution notices from the Source form of the Work,
+excluding those notices that do not pertain to any part of
+the Derivative Works; and
+
+(d) If the Work includes a "NOTICE" text file as part of its
+distribution, then any Derivative Works that You distribute must
+include a readable copy of the attribution notices contained
+within such NOTICE file, excluding those notices that do not
+pertain to any part of the Derivative Works, in at least one
+of the following places: within a NOTICE text file distributed
+as part of the Derivative Works; within the Source form or
+documentation, if provided along with the Derivative Works; or,
+within a display generated by the Derivative Works, if and
+wherever such third-party notices normally appear. The contents
+of the NOTICE file are for informational purposes only and
+do not modify the License. You may add Your own attribution
+notices within Derivative Works that You distribute, alongside
+or as an addendum to the NOTICE text from the Work, provided
+that such additional attribution notices cannot be construed
+as modifying the License.
+
+You may add Your own copyright statement to Your modifications and
+may provide additional or different license terms and conditions
+for use, reproduction, or distribution of Your modifications, or
+for any such Derivative Works as a whole, provided Your use,
+reproduction, and distribution of the Work otherwise complies with
+the conditions stated in this License.
+
+5. Submission of Contributions. Unless You explicitly state otherwise,
+any Contribution intentionally submitted for inclusion in the Work
+by You to the Licensor shall be under the terms and conditions of
+this License, without any additional terms or conditions.
+Notwithstanding the above, nothing herein shall supersede or modify
+the terms of any separate license agreement you may have executed
+with Licensor regarding such Contributions.
+
+6. Trademarks. This License does not grant permission to use the trade
+names, trademarks, service marks, or product names of the Licensor,
+except as required for reasonable and customary use in describing the
+origin of the Work and reproducing the content of the NOTICE file.
+
+7. Disclaimer of Warranty. Unless required by applicable law or
+agreed to in writing, Licensor provides the Work (and each
+Contributor provides its Contributions) on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+implied, including, without limitation, any warranties or conditions
+of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+PARTICULAR PURPOSE. You are solely responsible for determining the
+appropriateness of using or redistributing the Work and assume any
+risks associated with Your exercise of permissions under this License.
+
+8. Limitation of Liability. In no event and under no legal theory,
+whether in tort (including negligence), contract, or otherwise,
+unless required by applicable law (such as deliberate and grossly
+negligent acts) or agreed to in writing, shall any Contributor be
+liable to You for damages, including any direct, indirect, special,
+incidental, or consequential damages of any character arising as a
+result of this License or out of the use or inability to use the
+Work (including but not limited to damages for loss of goodwill,
+work stoppage, computer failure or malfunction, or any and all
+other commercial damages or losses), even if such Contributor
+has been advised of the possibility of such damages.
+
+9. Accepting Warranty or Additional Liability. While redistributing
+the Work or Derivative Works thereof, You may choose to offer,
+and charge a fee for, acceptance of support, warranty, indemnity,
+or other liability obligations and/or rights consistent with this
+License. However, in accepting such obligations, You may act only
+on Your own behalf and on Your sole responsibility, not on behalf
+of any other Contributor, and only if You agree to indemnify,
+defend, and hold each Contributor harmless for any liability
+incurred by, or claims asserted against, such Contributor by reason
+of your accepting any such warranty or additional liability.
+
+END OF TERMS AND CONDITIONS
+
+APPENDIX: How to apply the Apache License to your work.
+
+To apply the Apache License to your work, attach the following
+boilerplate notice, with the fields enclosed by brackets "[]"
+replaced with your own identifying information. (Don't include
+the brackets!)  The text should be enclosed in the appropriate
+comment syntax for the file format. We also recommend that a
+file or class name and description of purpose be included on the
+same "printed page" as the copyright notice for easier
+identification within third-party archives.
+
+Copyright 2019 - present, Facebook, Inc
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.

preprocess/humanparsing/mhp_extension/detectron2/MODEL_ZOO.md

@@ -0,0 +1,903 @@
+# Detectron2 Model Zoo and Baselines
+
+## Introduction
+
+This file documents a large collection of baselines trained
+with detectron2 in Sep-Oct, 2019.
+All numbers were obtained on [Big Basin](https://engineering.fb.com/data-center-engineering/introducing-big-basin-our-next-generation-ai-hardware/)
+servers with 8 NVIDIA V100 GPUs & NVLink. The software in use were PyTorch 1.3, CUDA 9.2, cuDNN 7.4.2 or 7.6.3.
+You can access these models from code using [detectron2.model_zoo](https://detectron2.readthedocs.io/modules/model_zoo.html) APIs.
+
+In addition to these official baseline models, you can find more models in [projects/](projects/).
+
+#### How to Read the Tables
+* The "Name" column contains a link to the config file. Running `tools/train_net.py` with this config file
+	and 8 GPUs will reproduce the model.
+* Training speed is averaged across the entire training.
+	We keep updating the speed with latest version of detectron2/pytorch/etc.,
+	so they might be different from the `metrics` file.
+	Training speed for multi-machine jobs is not provided.
+* Inference speed is measured by `tools/train_net.py --eval-only`, or [inference_on_dataset()](https://detectron2.readthedocs.io/modules/evaluation.html#detectron2.evaluation.inference_on_dataset),
+  with batch size 1 in detectron2 directly.
+	Measuring it with your own code will likely introduce other overhead.
+  Actual deployment in production should in general be faster than the given inference
+  speed due to more optimizations.
+* The *model id* column is provided for ease of reference.
+  To check downloaded file integrity, any model on this page contains its md5 prefix in its file name.
+* Training curves and other statistics can be found in `metrics` for each model.
+
+#### Common Settings for COCO Models
+* All COCO models were trained on `train2017` and evaluated on `val2017`.
+* The default settings are __not directly comparable__ with Detectron's standard settings.
+  For example, our default training data augmentation uses scale jittering in addition to horizontal flipping.
+
+  To make fair comparisons with Detectron's settings, see
+  [Detectron1-Comparisons](configs/Detectron1-Comparisons/) for accuracy comparison,
+  and [benchmarks](https://detectron2.readthedocs.io/notes/benchmarks.html)
+  for speed comparison.
+* For Faster/Mask R-CNN, we provide baselines based on __3 different backbone combinations__:
+  * __FPN__: Use a ResNet+FPN backbone with standard conv and FC heads for mask and box prediction,
+    respectively. It obtains the best
+    speed/accuracy tradeoff, but the other two are still useful for research.
+  * __C4__: Use a ResNet conv4 backbone with conv5 head. The original baseline in the Faster R-CNN paper.
+  * __DC5__ (Dilated-C5): Use a ResNet conv5 backbone with dilations in conv5, and standard conv and FC heads
+    for mask and box prediction, respectively.
+    This is used by the Deformable ConvNet paper.
+* Most models are trained with the 3x schedule (~37 COCO epochs).
+  Although 1x models are heavily under-trained, we provide some ResNet-50 models with the 1x (~12 COCO epochs)
+  training schedule for comparison when doing quick research iteration.
+
+#### ImageNet Pretrained Models
+
+We provide backbone models pretrained on ImageNet-1k dataset.
+These models have __different__ format from those provided in Detectron: we do not fuse BatchNorm into an affine layer.
+* [R-50.pkl](https://dl.fbaipublicfiles.com/detectron2/ImageNetPretrained/MSRA/R-50.pkl): converted copy of [MSRA's original ResNet-50](https://github.com/KaimingHe/deep-residual-networks) model.
+* [R-101.pkl](https://dl.fbaipublicfiles.com/detectron2/ImageNetPretrained/MSRA/R-101.pkl): converted copy of [MSRA's original ResNet-101](https://github.com/KaimingHe/deep-residual-networks) model.
+* [X-101-32x8d.pkl](https://dl.fbaipublicfiles.com/detectron2/ImageNetPretrained/FAIR/X-101-32x8d.pkl): ResNeXt-101-32x8d model trained with Caffe2 at FB.
+
+Pretrained models in Detectron's format can still be used. For example:
+* [X-152-32x8d-IN5k.pkl](https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/25093814/X-152-32x8d-IN5k.pkl):
+  ResNeXt-152-32x8d model trained on ImageNet-5k with Caffe2 at FB (see ResNeXt paper for details on ImageNet-5k).
+* [R-50-GN.pkl](https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl):
+  ResNet-50 with Group Normalization.
+* [R-101-GN.pkl](https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/47592356/R-101-GN.pkl):
+  ResNet-101 with Group Normalization.
+
+Torchvision's ResNet models can be used after converted by [this script](tools/convert-torchvision-to-d2.py).
+
+#### License
+
+All models available for download through this document are licensed under the
+[Creative Commons Attribution-ShareAlike 3.0 license](https://creativecommons.org/licenses/by-sa/3.0/).
+
+### COCO Object Detection Baselines
+
+#### Faster R-CNN:
+<!--
+(fb only) To update the table in vim:
+1. Remove the old table: d}
+2. Copy the below command to the place of the table
+3. :.!bash
+
+./gen_html_table.py --config 'COCO-Detection/faster*50*'{1x,3x}'*' 'COCO-Detection/faster*101*' --name R50-C4 R50-DC5 R50-FPN R50-C4 R50-DC5 R50-FPN R101-C4 R101-DC5 R101-FPN X101-FPN --fields lr_sched train_speed inference_speed mem box_AP
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: faster_rcnn_R_50_C4_1x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_50_C4_1x.yaml">R50-C4</a></td>
+<td align="center">1x</td>
+<td align="center">0.551</td>
+<td align="center">0.102</td>
+<td align="center">4.8</td>
+<td align="center">35.7</td>
+<td align="center">137257644</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_C4_1x/137257644/model_final_721ade.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_C4_1x/137257644/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_50_DC5_1x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_50_DC5_1x.yaml">R50-DC5</a></td>
+<td align="center">1x</td>
+<td align="center">0.380</td>
+<td align="center">0.068</td>
+<td align="center">5.0</td>
+<td align="center">37.3</td>
+<td align="center">137847829</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_DC5_1x/137847829/model_final_51d356.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_DC5_1x/137847829/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_50_FPN_1x.yaml">R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.210</td>
+<td align="center">0.038</td>
+<td align="center">3.0</td>
+<td align="center">37.9</td>
+<td align="center">137257794</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_FPN_1x/137257794/model_final_b275ba.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_FPN_1x/137257794/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_50_C4_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_50_C4_3x.yaml">R50-C4</a></td>
+<td align="center">3x</td>
+<td align="center">0.543</td>
+<td align="center">0.104</td>
+<td align="center">4.8</td>
+<td align="center">38.4</td>
+<td align="center">137849393</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_C4_3x/137849393/model_final_f97cb7.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_C4_3x/137849393/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_50_DC5_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_50_DC5_3x.yaml">R50-DC5</a></td>
+<td align="center">3x</td>
+<td align="center">0.378</td>
+<td align="center">0.070</td>
+<td align="center">5.0</td>
+<td align="center">39.0</td>
+<td align="center">137849425</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_DC5_3x/137849425/model_final_68d202.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_DC5_3x/137849425/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_50_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml">R50-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.209</td>
+<td align="center">0.038</td>
+<td align="center">3.0</td>
+<td align="center">40.2</td>
+<td align="center">137849458</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_FPN_3x/137849458/model_final_280758.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_50_FPN_3x/137849458/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_101_C4_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_101_C4_3x.yaml">R101-C4</a></td>
+<td align="center">3x</td>
+<td align="center">0.619</td>
+<td align="center">0.139</td>
+<td align="center">5.9</td>
+<td align="center">41.1</td>
+<td align="center">138204752</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_101_C4_3x/138204752/model_final_298dad.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_101_C4_3x/138204752/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_101_DC5_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_101_DC5_3x.yaml">R101-DC5</a></td>
+<td align="center">3x</td>
+<td align="center">0.452</td>
+<td align="center">0.086</td>
+<td align="center">6.1</td>
+<td align="center">40.6</td>
+<td align="center">138204841</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_101_DC5_3x/138204841/model_final_3e0943.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_101_DC5_3x/138204841/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_101_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml">R101-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.286</td>
+<td align="center">0.051</td>
+<td align="center">4.1</td>
+<td align="center">42.0</td>
+<td align="center">137851257</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_101_FPN_3x/137851257/model_final_f6e8b1.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_R_101_FPN_3x/137851257/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_X_101_32x8d_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml">X101-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.638</td>
+<td align="center">0.098</td>
+<td align="center">6.7</td>
+<td align="center">43.0</td>
+<td align="center">139173657</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x/139173657/model_final_68b088.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x/139173657/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+#### RetinaNet:
+<!--
+./gen_html_table.py --config 'COCO-Detection/retina*50*' 'COCO-Detection/retina*101*' --name R50 R50 R101 --fields lr_sched train_speed inference_speed mem box_AP
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: retinanet_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/retinanet_R_50_FPN_1x.yaml">R50</a></td>
+<td align="center">1x</td>
+<td align="center">0.200</td>
+<td align="center">0.055</td>
+<td align="center">3.9</td>
+<td align="center">36.5</td>
+<td align="center">137593951</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet_R_50_FPN_1x/137593951/model_final_b796dc.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet_R_50_FPN_1x/137593951/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: retinanet_R_50_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/retinanet_R_50_FPN_3x.yaml">R50</a></td>
+<td align="center">3x</td>
+<td align="center">0.201</td>
+<td align="center">0.055</td>
+<td align="center">3.9</td>
+<td align="center">37.9</td>
+<td align="center">137849486</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet_R_50_FPN_3x/137849486/model_final_4cafe0.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet_R_50_FPN_3x/137849486/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: retinanet_R_101_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/retinanet_R_101_FPN_3x.yaml">R101</a></td>
+<td align="center">3x</td>
+<td align="center">0.280</td>
+<td align="center">0.068</td>
+<td align="center">5.1</td>
+<td align="center">39.9</td>
+<td align="center">138363263</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet_R_101_FPN_3x/138363263/model_final_59f53c.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet_R_101_FPN_3x/138363263/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+#### RPN & Fast R-CNN:
+<!--
+./gen_html_table.py --config 'COCO-Detection/rpn*' 'COCO-Detection/fast_rcnn*' --name "RPN R50-C4" "RPN R50-FPN" "Fast R-CNN R50-FPN" --fields lr_sched train_speed inference_speed mem box_AP prop_AR
+-->
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">prop.<br/>AR</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: rpn_R_50_C4_1x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/rpn_R_50_C4_1x.yaml">RPN R50-C4</a></td>
+<td align="center">1x</td>
+<td align="center">0.130</td>
+<td align="center">0.034</td>
+<td align="center">1.5</td>
+<td align="center"></td>
+<td align="center">51.6</td>
+<td align="center">137258005</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn_R_50_C4_1x/137258005/model_final_450694.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn_R_50_C4_1x/137258005/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: rpn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/rpn_R_50_FPN_1x.yaml">RPN R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.186</td>
+<td align="center">0.032</td>
+<td align="center">2.7</td>
+<td align="center"></td>
+<td align="center">58.0</td>
+<td align="center">137258492</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn_R_50_FPN_1x/137258492/model_final_02ce48.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn_R_50_FPN_1x/137258492/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: fast_rcnn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/COCO-Detection/fast_rcnn_R_50_FPN_1x.yaml">Fast R-CNN R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.140</td>
+<td align="center">0.029</td>
+<td align="center">2.6</td>
+<td align="center">37.8</td>
+<td align="center"></td>
+<td align="center">137635226</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/fast_rcnn_R_50_FPN_1x/137635226/model_final_e5f7ce.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/fast_rcnn_R_50_FPN_1x/137635226/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+### COCO Instance Segmentation Baselines with Mask R-CNN
+<!--
+./gen_html_table.py --config 'COCO-InstanceSegmentation/mask*50*'{1x,3x}'*' 'COCO-InstanceSegmentation/mask*101*' --name R50-C4 R50-DC5 R50-FPN R50-C4 R50-DC5 R50-FPN R101-C4 R101-DC5 R101-FPN X101-FPN --fields lr_sched train_speed inference_speed mem box_AP mask_AP
+-->
+
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">mask<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: mask_rcnn_R_50_C4_1x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_C4_1x.yaml">R50-C4</a></td>
+<td align="center">1x</td>
+<td align="center">0.584</td>
+<td align="center">0.110</td>
+<td align="center">5.2</td>
+<td align="center">36.8</td>
+<td align="center">32.2</td>
+<td align="center">137259246</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_C4_1x/137259246/model_final_9243eb.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_C4_1x/137259246/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_DC5_1x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_1x.yaml">R50-DC5</a></td>
+<td align="center">1x</td>
+<td align="center">0.471</td>
+<td align="center">0.076</td>
+<td align="center">6.5</td>
+<td align="center">38.3</td>
+<td align="center">34.2</td>
+<td align="center">137260150</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_1x/137260150/model_final_4f86c3.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_1x/137260150/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml">R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.261</td>
+<td align="center">0.043</td>
+<td align="center">3.4</td>
+<td align="center">38.6</td>
+<td align="center">35.2</td>
+<td align="center">137260431</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x/137260431/model_final_a54504.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x/137260431/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_C4_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_C4_3x.yaml">R50-C4</a></td>
+<td align="center">3x</td>
+<td align="center">0.575</td>
+<td align="center">0.111</td>
+<td align="center">5.2</td>
+<td align="center">39.8</td>
+<td align="center">34.4</td>
+<td align="center">137849525</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_C4_3x/137849525/model_final_4ce675.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_C4_3x/137849525/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_DC5_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_3x.yaml">R50-DC5</a></td>
+<td align="center">3x</td>
+<td align="center">0.470</td>
+<td align="center">0.076</td>
+<td align="center">6.5</td>
+<td align="center">40.0</td>
+<td align="center">35.9</td>
+<td align="center">137849551</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_3x/137849551/model_final_84107b.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_3x/137849551/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml">R50-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.261</td>
+<td align="center">0.043</td>
+<td align="center">3.4</td>
+<td align="center">41.0</td>
+<td align="center">37.2</td>
+<td align="center">137849600</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_101_C4_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_101_C4_3x.yaml">R101-C4</a></td>
+<td align="center">3x</td>
+<td align="center">0.652</td>
+<td align="center">0.145</td>
+<td align="center">6.3</td>
+<td align="center">42.6</td>
+<td align="center">36.7</td>
+<td align="center">138363239</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_101_C4_3x/138363239/model_final_a2914c.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_101_C4_3x/138363239/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_101_DC5_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_101_DC5_3x.yaml">R101-DC5</a></td>
+<td align="center">3x</td>
+<td align="center">0.545</td>
+<td align="center">0.092</td>
+<td align="center">7.6</td>
+<td align="center">41.9</td>
+<td align="center">37.3</td>
+<td align="center">138363294</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_101_DC5_3x/138363294/model_final_0464b7.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_101_DC5_3x/138363294/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_101_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_101_FPN_3x.yaml">R101-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.340</td>
+<td align="center">0.056</td>
+<td align="center">4.6</td>
+<td align="center">42.9</td>
+<td align="center">38.6</td>
+<td align="center">138205316</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_101_FPN_3x/138205316/model_final_a3ec72.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_101_FPN_3x/138205316/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_X_101_32x8d_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml">X101-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.690</td>
+<td align="center">0.103</td>
+<td align="center">7.2</td>
+<td align="center">44.3</td>
+<td align="center">39.5</td>
+<td align="center">139653917</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x/139653917/model_final_2d9806.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x/139653917/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+### COCO Person Keypoint Detection Baselines with Keypoint R-CNN
+<!--
+./gen_html_table.py --config 'COCO-Keypoints/*50*' 'COCO-Keypoints/*101*'  --name R50-FPN R50-FPN R101-FPN X101-FPN --fields lr_sched train_speed inference_speed mem box_AP keypoint_AP
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">kp.<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: keypoint_rcnn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/COCO-Keypoints/keypoint_rcnn_R_50_FPN_1x.yaml">R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.315</td>
+<td align="center">0.072</td>
+<td align="center">5.0</td>
+<td align="center">53.6</td>
+<td align="center">64.0</td>
+<td align="center">137261548</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_R_50_FPN_1x/137261548/model_final_04e291.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_R_50_FPN_1x/137261548/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: keypoint_rcnn_R_50_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x.yaml">R50-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.316</td>
+<td align="center">0.066</td>
+<td align="center">5.0</td>
+<td align="center">55.4</td>
+<td align="center">65.5</td>
+<td align="center">137849621</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x/137849621/model_final_a6e10b.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x/137849621/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: keypoint_rcnn_R_101_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Keypoints/keypoint_rcnn_R_101_FPN_3x.yaml">R101-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.390</td>
+<td align="center">0.076</td>
+<td align="center">6.1</td>
+<td align="center">56.4</td>
+<td align="center">66.1</td>
+<td align="center">138363331</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_R_101_FPN_3x/138363331/model_final_997cc7.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_R_101_FPN_3x/138363331/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: keypoint_rcnn_X_101_32x8d_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-Keypoints/keypoint_rcnn_X_101_32x8d_FPN_3x.yaml">X101-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.738</td>
+<td align="center">0.121</td>
+<td align="center">8.7</td>
+<td align="center">57.3</td>
+<td align="center">66.0</td>
+<td align="center">139686956</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_X_101_32x8d_FPN_3x/139686956/model_final_5ad38f.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-Keypoints/keypoint_rcnn_X_101_32x8d_FPN_3x/139686956/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+### COCO Panoptic Segmentation Baselines with Panoptic FPN
+<!--
+./gen_html_table.py --config 'COCO-PanopticSegmentation/*50*' 'COCO-PanopticSegmentation/*101*'  --name R50-FPN R50-FPN R101-FPN --fields lr_sched train_speed inference_speed mem box_AP mask_AP PQ
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">mask<br/>AP</th>
+<th valign="bottom">PQ</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: panoptic_fpn_R_50_1x -->
+ <tr><td align="left"><a href="configs/COCO-PanopticSegmentation/panoptic_fpn_R_50_1x.yaml">R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.304</td>
+<td align="center">0.053</td>
+<td align="center">4.8</td>
+<td align="center">37.6</td>
+<td align="center">34.7</td>
+<td align="center">39.4</td>
+<td align="center">139514544</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-PanopticSegmentation/panoptic_fpn_R_50_1x/139514544/model_final_dbfeb4.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-PanopticSegmentation/panoptic_fpn_R_50_1x/139514544/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: panoptic_fpn_R_50_3x -->
+ <tr><td align="left"><a href="configs/COCO-PanopticSegmentation/panoptic_fpn_R_50_3x.yaml">R50-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.302</td>
+<td align="center">0.053</td>
+<td align="center">4.8</td>
+<td align="center">40.0</td>
+<td align="center">36.5</td>
+<td align="center">41.5</td>
+<td align="center">139514569</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-PanopticSegmentation/panoptic_fpn_R_50_3x/139514569/model_final_c10459.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-PanopticSegmentation/panoptic_fpn_R_50_3x/139514569/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: panoptic_fpn_R_101_3x -->
+ <tr><td align="left"><a href="configs/COCO-PanopticSegmentation/panoptic_fpn_R_101_3x.yaml">R101-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.392</td>
+<td align="center">0.066</td>
+<td align="center">6.0</td>
+<td align="center">42.4</td>
+<td align="center">38.5</td>
+<td align="center">43.0</td>
+<td align="center">139514519</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-PanopticSegmentation/panoptic_fpn_R_101_3x/139514519/model_final_cafdb1.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-PanopticSegmentation/panoptic_fpn_R_101_3x/139514519/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+
+### LVIS Instance Segmentation Baselines with Mask R-CNN
+
+Mask R-CNN baselines on the [LVIS dataset](https://lvisdataset.org), v0.5.
+These baselines are described in Table 3(c) of the [LVIS paper](https://arxiv.org/abs/1908.03195).
+
+NOTE: the 1x schedule here has the same amount of __iterations__ as the COCO 1x baselines.
+They are roughly 24 epochs of LVISv0.5 data.
+The final results of these configs have large variance across different runs.
+
+<!--
+./gen_html_table.py --config 'LVIS-InstanceSegmentation/mask*50*' 'LVIS-InstanceSegmentation/mask*101*' --name R50-FPN R101-FPN X101-FPN --fields lr_sched train_speed inference_speed mem box_AP mask_AP
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">mask<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: mask_rcnn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/LVIS-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml">R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.292</td>
+<td align="center">0.107</td>
+<td align="center">7.1</td>
+<td align="center">23.6</td>
+<td align="center">24.4</td>
+<td align="center">144219072</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/LVIS-InstanceSegmentation/mask_rcnn_R_50_FPN_1x/144219072/model_final_571f7c.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/LVIS-InstanceSegmentation/mask_rcnn_R_50_FPN_1x/144219072/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_101_FPN_1x -->
+ <tr><td align="left"><a href="configs/LVIS-InstanceSegmentation/mask_rcnn_R_101_FPN_1x.yaml">R101-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.371</td>
+<td align="center">0.114</td>
+<td align="center">7.8</td>
+<td align="center">25.6</td>
+<td align="center">25.9</td>
+<td align="center">144219035</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/LVIS-InstanceSegmentation/mask_rcnn_R_101_FPN_1x/144219035/model_final_824ab5.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/LVIS-InstanceSegmentation/mask_rcnn_R_101_FPN_1x/144219035/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_X_101_32x8d_FPN_1x -->
+ <tr><td align="left"><a href="configs/LVIS-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_1x.yaml">X101-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.712</td>
+<td align="center">0.151</td>
+<td align="center">10.2</td>
+<td align="center">26.7</td>
+<td align="center">27.1</td>
+<td align="center">144219108</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/LVIS-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_1x/144219108/model_final_5e3439.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/LVIS-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_1x/144219108/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+
+
+### Cityscapes & Pascal VOC Baselines
+
+Simple baselines for
+* Mask R-CNN on Cityscapes instance segmentation (initialized from COCO pre-training, then trained on Cityscapes fine annotations only)
+* Faster R-CNN on PASCAL VOC object detection (trained on VOC 2007 train+val + VOC 2012 train+val, tested on VOC 2007 using 11-point interpolated AP)
+
+<!--
+./gen_html_table.py --config 'Cityscapes/*' 'PascalVOC-Detection/*' --name "R50-FPN, Cityscapes" "R50-C4, VOC" --fields train_speed inference_speed mem box_AP box_AP50 mask_AP
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">box<br/>AP50</th>
+<th valign="bottom">mask<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: mask_rcnn_R_50_FPN -->
+ <tr><td align="left"><a href="configs/Cityscapes/mask_rcnn_R_50_FPN.yaml">R50-FPN, Cityscapes</a></td>
+<td align="center">0.240</td>
+<td align="center">0.078</td>
+<td align="center">4.4</td>
+<td align="center"></td>
+<td align="center"></td>
+<td align="center">36.5</td>
+<td align="center">142423278</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Cityscapes/mask_rcnn_R_50_FPN/142423278/model_final_af9cf5.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Cityscapes/mask_rcnn_R_50_FPN/142423278/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: faster_rcnn_R_50_C4 -->
+ <tr><td align="left"><a href="configs/PascalVOC-Detection/faster_rcnn_R_50_C4.yaml">R50-C4, VOC</a></td>
+<td align="center">0.537</td>
+<td align="center">0.081</td>
+<td align="center">4.8</td>
+<td align="center">51.9</td>
+<td align="center">80.3</td>
+<td align="center"></td>
+<td align="center">142202221</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/PascalVOC-Detection/faster_rcnn_R_50_C4/142202221/model_final_b1acc2.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/PascalVOC-Detection/faster_rcnn_R_50_C4/142202221/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+
+
+### Other Settings
+
+Ablations for Deformable Conv and Cascade R-CNN:
+
+<!--
+./gen_html_table.py --config 'COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml' 'Misc/*R_50_FPN_1x_dconv*' 'Misc/cascade*1x.yaml' 'COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml' 'Misc/*R_50_FPN_3x_dconv*' 'Misc/cascade*3x.yaml' --name "Baseline R50-FPN" "Deformable Conv" "Cascade R-CNN" "Baseline R50-FPN" "Deformable Conv" "Cascade R-CNN"  --fields lr_sched train_speed inference_speed mem box_AP mask_AP
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">mask<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: mask_rcnn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml">Baseline R50-FPN</a></td>
+<td align="center">1x</td>
+<td align="center">0.261</td>
+<td align="center">0.043</td>
+<td align="center">3.4</td>
+<td align="center">38.6</td>
+<td align="center">35.2</td>
+<td align="center">137260431</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x/137260431/model_final_a54504.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x/137260431/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_FPN_1x_dconv_c3-c5 -->
+ <tr><td align="left"><a href="configs/Misc/mask_rcnn_R_50_FPN_1x_dconv_c3-c5.yaml">Deformable Conv</a></td>
+<td align="center">1x</td>
+<td align="center">0.342</td>
+<td align="center">0.048</td>
+<td align="center">3.5</td>
+<td align="center">41.5</td>
+<td align="center">37.5</td>
+<td align="center">138602867</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_1x_dconv_c3-c5/138602867/model_final_65c703.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_1x_dconv_c3-c5/138602867/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: cascade_mask_rcnn_R_50_FPN_1x -->
+ <tr><td align="left"><a href="configs/Misc/cascade_mask_rcnn_R_50_FPN_1x.yaml">Cascade R-CNN</a></td>
+<td align="center">1x</td>
+<td align="center">0.317</td>
+<td align="center">0.052</td>
+<td align="center">4.0</td>
+<td align="center">42.1</td>
+<td align="center">36.4</td>
+<td align="center">138602847</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_R_50_FPN_1x/138602847/model_final_e9d89b.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_R_50_FPN_1x/138602847/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml">Baseline R50-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.261</td>
+<td align="center">0.043</td>
+<td align="center">3.4</td>
+<td align="center">41.0</td>
+<td align="center">37.2</td>
+<td align="center">137849600</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_FPN_3x_dconv_c3-c5 -->
+ <tr><td align="left"><a href="configs/Misc/mask_rcnn_R_50_FPN_3x_dconv_c3-c5.yaml">Deformable Conv</a></td>
+<td align="center">3x</td>
+<td align="center">0.349</td>
+<td align="center">0.047</td>
+<td align="center">3.5</td>
+<td align="center">42.7</td>
+<td align="center">38.5</td>
+<td align="center">144998336</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_3x_dconv_c3-c5/144998336/model_final_821d0b.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_3x_dconv_c3-c5/144998336/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: cascade_mask_rcnn_R_50_FPN_3x -->
+ <tr><td align="left"><a href="configs/Misc/cascade_mask_rcnn_R_50_FPN_3x.yaml">Cascade R-CNN</a></td>
+<td align="center">3x</td>
+<td align="center">0.328</td>
+<td align="center">0.053</td>
+<td align="center">4.0</td>
+<td align="center">44.3</td>
+<td align="center">38.5</td>
+<td align="center">144998488</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_R_50_FPN_3x/144998488/model_final_480dd8.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_R_50_FPN_3x/144998488/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+
+Ablations for normalization methods, and a few models trained from scratch following [Rethinking ImageNet Pre-training](https://arxiv.org/abs/1811.08883).
+(Note: The baseline uses `2fc` head while the others use [`4conv1fc` head](https://arxiv.org/abs/1803.08494))
+<!--
+./gen_html_table.py --config 'COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml' 'Misc/mask*50_FPN_3x_gn.yaml' 'Misc/mask*50_FPN_3x_syncbn.yaml' 'Misc/scratch*' --name "Baseline R50-FPN" "GN" "SyncBN" "GN (from scratch)" "GN (from scratch)" "SyncBN (from scratch)" --fields lr_sched train_speed inference_speed mem box_AP mask_AP
+   -->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">lr<br/>sched</th>
+<th valign="bottom">train<br/>time<br/>(s/iter)</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">mask<br/>AP</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: mask_rcnn_R_50_FPN_3x -->
+ <tr><td align="left"><a href="configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml">Baseline R50-FPN</a></td>
+<td align="center">3x</td>
+<td align="center">0.261</td>
+<td align="center">0.043</td>
+<td align="center">3.4</td>
+<td align="center">41.0</td>
+<td align="center">37.2</td>
+<td align="center">137849600</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_FPN_3x_gn -->
+ <tr><td align="left"><a href="configs/Misc/mask_rcnn_R_50_FPN_3x_gn.yaml">GN</a></td>
+<td align="center">3x</td>
+<td align="center">0.356</td>
+<td align="center">0.069</td>
+<td align="center">7.3</td>
+<td align="center">42.6</td>
+<td align="center">38.6</td>
+<td align="center">138602888</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_3x_gn/138602888/model_final_dc5d9e.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_3x_gn/138602888/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: mask_rcnn_R_50_FPN_3x_syncbn -->
+ <tr><td align="left"><a href="configs/Misc/mask_rcnn_R_50_FPN_3x_syncbn.yaml">SyncBN</a></td>
+<td align="center">3x</td>
+<td align="center">0.371</td>
+<td align="center">0.053</td>
+<td align="center">5.5</td>
+<td align="center">41.9</td>
+<td align="center">37.8</td>
+<td align="center">169527823</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_3x_syncbn/169527823/model_final_3b3c51.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/mask_rcnn_R_50_FPN_3x_syncbn/169527823/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: scratch_mask_rcnn_R_50_FPN_3x_gn -->
+ <tr><td align="left"><a href="configs/Misc/scratch_mask_rcnn_R_50_FPN_3x_gn.yaml">GN (from scratch)</a></td>
+<td align="center">3x</td>
+<td align="center">0.400</td>
+<td align="center">0.069</td>
+<td align="center">9.8</td>
+<td align="center">39.9</td>
+<td align="center">36.6</td>
+<td align="center">138602908</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/scratch_mask_rcnn_R_50_FPN_3x_gn/138602908/model_final_01ca85.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/scratch_mask_rcnn_R_50_FPN_3x_gn/138602908/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: scratch_mask_rcnn_R_50_FPN_9x_gn -->
+ <tr><td align="left"><a href="configs/Misc/scratch_mask_rcnn_R_50_FPN_9x_gn.yaml">GN (from scratch)</a></td>
+<td align="center">9x</td>
+<td align="center">N/A</td>
+<td align="center">0.070</td>
+<td align="center">9.8</td>
+<td align="center">43.7</td>
+<td align="center">39.6</td>
+<td align="center">183808979</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/scratch_mask_rcnn_R_50_FPN_9x_gn/183808979/model_final_da7b4c.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/scratch_mask_rcnn_R_50_FPN_9x_gn/183808979/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: scratch_mask_rcnn_R_50_FPN_9x_syncbn -->
+ <tr><td align="left"><a href="configs/Misc/scratch_mask_rcnn_R_50_FPN_9x_syncbn.yaml">SyncBN (from scratch)</a></td>
+<td align="center">9x</td>
+<td align="center">N/A</td>
+<td align="center">0.055</td>
+<td align="center">7.2</td>
+<td align="center">43.6</td>
+<td align="center">39.3</td>
+<td align="center">184226666</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/scratch_mask_rcnn_R_50_FPN_9x_syncbn/184226666/model_final_5ce33e.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/scratch_mask_rcnn_R_50_FPN_9x_syncbn/184226666/metrics.json">metrics</a></td>
+</tr>
+</tbody></table>
+
+
+A few very large models trained for a long time, for demo purposes. They are trained using multiple machines:
+
+<!--
+./gen_html_table.py --config 'Misc/panoptic_*dconv*' 'Misc/cascade_*152*' --name "Panoptic FPN R101" "Mask R-CNN X152" --fields inference_speed mem box_AP mask_AP PQ
+# manually add TTA results
+-->
+
+
+<table><tbody>
+<!-- START TABLE -->
+<!-- TABLE HEADER -->
+<th valign="bottom">Name</th>
+<th valign="bottom">inference<br/>time<br/>(s/im)</th>
+<th valign="bottom">train<br/>mem<br/>(GB)</th>
+<th valign="bottom">box<br/>AP</th>
+<th valign="bottom">mask<br/>AP</th>
+<th valign="bottom">PQ</th>
+<th valign="bottom">model id</th>
+<th valign="bottom">download</th>
+<!-- TABLE BODY -->
+<!-- ROW: panoptic_fpn_R_101_dconv_cascade_gn_3x -->
+ <tr><td align="left"><a href="configs/Misc/panoptic_fpn_R_101_dconv_cascade_gn_3x.yaml">Panoptic FPN R101</a></td>
+<td align="center">0.107</td>
+<td align="center">11.4</td>
+<td align="center">47.4</td>
+<td align="center">41.3</td>
+<td align="center">46.1</td>
+<td align="center">139797668</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/panoptic_fpn_R_101_dconv_cascade_gn_3x/139797668/model_final_be35db.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/panoptic_fpn_R_101_dconv_cascade_gn_3x/139797668/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv -->
+ <tr><td align="left"><a href="configs/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv.yaml">Mask R-CNN X152</a></td>
+<td align="center">0.242</td>
+<td align="center">15.1</td>
+<td align="center">50.2</td>
+<td align="center">44.0</td>
+<td align="center"></td>
+<td align="center">18131413</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv/18131413/model_0039999_e76410.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv/18131413/metrics.json">metrics</a></td>
+</tr>
+<!-- ROW: TTA cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv -->
+ <tr><td align="left">above + test-time aug.</td>
+<td align="center"></td>
+<td align="center"></td>
+<td align="center">51.9</td>
+<td align="center">45.9</td>
+<td align="center"></td>
+<td align="center"></td>
+<td align="center"></td>
+</tr>
+</tbody></table>

preprocess/humanparsing/mhp_extension/detectron2/README.md

@@ -0,0 +1,56 @@
+<img src=".github/Detectron2-Logo-Horz.svg" width="300" >
+
+Detectron2 is Facebook AI Research's next generation software system
+that implements state-of-the-art object detection algorithms.
+It is a ground-up rewrite of the previous version,
+[Detectron](https://github.com/facebookresearch/Detectron/),
+and it originates from [maskrcnn-benchmark](https://github.com/facebookresearch/maskrcnn-benchmark/).
+
+<div align="center">
+  <img src="https://user-images.githubusercontent.com/1381301/66535560-d3422200-eace-11e9-9123-5535d469db19.png"/>
+</div>
+
+### What's New
+* It is powered by the [PyTorch](https://pytorch.org) deep learning framework.
+* Includes more features such as panoptic segmentation, densepose, Cascade R-CNN, rotated bounding boxes, etc.
+* Can be used as a library to support [different projects](projects/) on top of it.
+  We'll open source more research projects in this way.
+* It [trains much faster](https://detectron2.readthedocs.io/notes/benchmarks.html).
+
+See our [blog post](https://ai.facebook.com/blog/-detectron2-a-pytorch-based-modular-object-detection-library-/)
+to see more demos and learn about detectron2.
+
+## Installation
+
+See [INSTALL.md](INSTALL.md).
+
+## Quick Start
+
+See [GETTING_STARTED.md](GETTING_STARTED.md),
+or the [Colab Notebook](https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5).
+
+Learn more at our [documentation](https://detectron2.readthedocs.org).
+And see [projects/](projects/) for some projects that are built on top of detectron2.
+
+## Model Zoo and Baselines
+
+We provide a large set of baseline results and trained models available for download in the [Detectron2 Model Zoo](MODEL_ZOO.md).
+
+
+## License
+
+Detectron2 is released under the [Apache 2.0 license](LICENSE).
+
+## Citing Detectron2
+
+If you use Detectron2 in your research or wish to refer to the baseline results published in the [Model Zoo](MODEL_ZOO.md), please use the following BibTeX entry.
+
+```BibTeX
+@misc{wu2019detectron2,
+  author =       {Yuxin Wu and Alexander Kirillov and Francisco Massa and
+                  Wan-Yen Lo and Ross Girshick},
+  title =        {Detectron2},
+  howpublished = {\url{https://github.com/facebookresearch/detectron2}},
+  year =         {2019}
+}
+```

preprocess/humanparsing/mhp_extension/detectron2/configs/Base-RCNN-C4.yaml

@@ -0,0 +1,18 @@
+MODEL:
+  META_ARCHITECTURE: "GeneralizedRCNN"
+  RPN:
+    PRE_NMS_TOPK_TEST: 6000
+    POST_NMS_TOPK_TEST: 1000
+  ROI_HEADS:
+    NAME: "Res5ROIHeads"
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.02
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

preprocess/humanparsing/mhp_extension/detectron2/configs/Base-RCNN-DilatedC5.yaml

@@ -0,0 +1,31 @@
+MODEL:
+  META_ARCHITECTURE: "GeneralizedRCNN"
+  RESNETS:
+    OUT_FEATURES: ["res5"]
+    RES5_DILATION: 2
+  RPN:
+    IN_FEATURES: ["res5"]
+    PRE_NMS_TOPK_TEST: 6000
+    POST_NMS_TOPK_TEST: 1000
+  ROI_HEADS:
+    NAME: "StandardROIHeads"
+    IN_FEATURES: ["res5"]
+  ROI_BOX_HEAD:
+    NAME: "FastRCNNConvFCHead"
+    NUM_FC: 2
+    POOLER_RESOLUTION: 7
+  ROI_MASK_HEAD:
+    NAME: "MaskRCNNConvUpsampleHead"
+    NUM_CONV: 4
+    POOLER_RESOLUTION: 14
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.02
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

preprocess/humanparsing/mhp_extension/detectron2/configs/Base-RCNN-FPN.yaml

@@ -0,0 +1,42 @@
+MODEL:
+  META_ARCHITECTURE: "GeneralizedRCNN"
+  BACKBONE:
+    NAME: "build_resnet_fpn_backbone"
+  RESNETS:
+    OUT_FEATURES: ["res2", "res3", "res4", "res5"]
+  FPN:
+    IN_FEATURES: ["res2", "res3", "res4", "res5"]
+  ANCHOR_GENERATOR:
+    SIZES: [[32], [64], [128], [256], [512]]  # One size for each in feature map
+    ASPECT_RATIOS: [[0.5, 1.0, 2.0]]  # Three aspect ratios (same for all in feature maps)
+  RPN:
+    IN_FEATURES: ["p2", "p3", "p4", "p5", "p6"]
+    PRE_NMS_TOPK_TRAIN: 2000  # Per FPN level
+    PRE_NMS_TOPK_TEST: 1000  # Per FPN level
+    # Detectron1 uses 2000 proposals per-batch,
+    # (See "modeling/rpn/rpn_outputs.py" for details of this legacy issue)
+    # which is approximately 1000 proposals per-image since the default batch size for FPN is 2.
+    POST_NMS_TOPK_TRAIN: 1000
+    POST_NMS_TOPK_TEST: 1000
+  ROI_HEADS:
+    NAME: "StandardROIHeads"
+    IN_FEATURES: ["p2", "p3", "p4", "p5"]
+  ROI_BOX_HEAD:
+    NAME: "FastRCNNConvFCHead"
+    NUM_FC: 2
+    POOLER_RESOLUTION: 7
+  ROI_MASK_HEAD:
+    NAME: "MaskRCNNConvUpsampleHead"
+    NUM_CONV: 4
+    POOLER_RESOLUTION: 14
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.02
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

preprocess/humanparsing/mhp_extension/detectron2/configs/Base-RetinaNet.yaml

@@ -0,0 +1,24 @@
+MODEL:
+  META_ARCHITECTURE: "RetinaNet"
+  BACKBONE:
+    NAME: "build_retinanet_resnet_fpn_backbone"
+  RESNETS:
+    OUT_FEATURES: ["res3", "res4", "res5"]
+  ANCHOR_GENERATOR:
+    SIZES: !!python/object/apply:eval ["[[x, x * 2**(1.0/3), x * 2**(2.0/3) ] for x in [32, 64, 128, 256, 512 ]]"]
+  FPN:
+    IN_FEATURES: ["res3", "res4", "res5"]
+  RETINANET:
+    IOU_THRESHOLDS: [0.4, 0.5]
+    IOU_LABELS: [0, -1, 1]
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  TEST: ("coco_2017_val",)
+SOLVER:
+  IMS_PER_BATCH: 16
+  BASE_LR: 0.01  # Note that RetinaNet uses a different default learning rate
+  STEPS: (60000, 80000)
+  MAX_ITER: 90000
+INPUT:
+  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
+VERSION: 2

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/fast_rcnn_R_50_FPN_1x.yaml

@@ -0,0 +1,17 @@
+_BASE_: "../Base-RCNN-FPN.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  MASK_ON: False
+  LOAD_PROPOSALS: True
+  RESNETS:
+    DEPTH: 50
+  PROPOSAL_GENERATOR:
+    NAME: "PrecomputedProposals"
+DATASETS:
+  TRAIN: ("coco_2017_train",)
+  PROPOSAL_FILES_TRAIN: ("detectron2://COCO-Detection/rpn_R_50_FPN_1x/137258492/coco_2017_train_box_proposals_21bc3a.pkl", )
+  TEST: ("coco_2017_val",)
+  PROPOSAL_FILES_TEST: ("detectron2://COCO-Detection/rpn_R_50_FPN_1x/137258492/coco_2017_val_box_proposals_ee0dad.pkl", )
+DATALOADER:
+  # proposals are part of the dataset_dicts, and take a lot of RAM
+  NUM_WORKERS: 2

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/faster_rcnn_R_101_C4_3x.yaml

@@ -0,0 +1,9 @@
+_BASE_: "../Base-RCNN-C4.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-101.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 101
+SOLVER:
+  STEPS: (210000, 250000)
+  MAX_ITER: 270000

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/faster_rcnn_R_101_DC5_3x.yaml

@@ -0,0 +1,9 @@
+_BASE_: "../Base-RCNN-DilatedC5.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-101.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 101
+SOLVER:
+  STEPS: (210000, 250000)
+  MAX_ITER: 270000

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml

@@ -0,0 +1,9 @@
+_BASE_: "../Base-RCNN-FPN.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-101.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 101
+SOLVER:
+  STEPS: (210000, 250000)
+  MAX_ITER: 270000

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/faster_rcnn_R_50_C4_1x.yaml

@@ -0,0 +1,6 @@
+_BASE_: "../Base-RCNN-C4.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/faster_rcnn_R_50_C4_3x.yaml

@@ -0,0 +1,9 @@
+_BASE_: "../Base-RCNN-C4.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50
+SOLVER:
+  STEPS: (210000, 250000)
+  MAX_ITER: 270000

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/faster_rcnn_R_50_DC5_1x.yaml

@@ -0,0 +1,6 @@
+_BASE_: "../Base-RCNN-DilatedC5.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50

preprocess/humanparsing/mhp_extension/detectron2/configs/COCO-Detection/faster_rcnn_R_50_DC5_3x.yaml

@@ -0,0 +1,9 @@
+_BASE_: "../Base-RCNN-DilatedC5.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50
+SOLVER:
+  STEPS: (210000, 250000)
+  MAX_ITER: 270000