Upload model (#1)

- Upload model (e048792e24082576a734b07c13ca038577dd62f6)

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

adaptor_base.py

@@ -0,0 +1,35 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+from argparse import Namespace
+from typing import NamedTuple
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+
+class AdaptorInput(NamedTuple):
+    images: torch.Tensor
+    summary: torch.Tensor
+    features: torch.Tensor
+
+
+class RadioOutput(NamedTuple):
+    summary: torch.Tensor
+    features: torch.Tensor
+
+    def to(self, *args, **kwargs):
+        return RadioOutput(
+            self.summary.to(*args, **kwargs) if self.summary is not None else None,
+            self.features.to(*args, **kwargs) if self.features is not None else None,
+        )
+
+
+class AdaptorBase(nn.Module):
+    def forward(self, input: AdaptorInput) -> RadioOutput:
+        raise NotImplementedError("Subclasses must implement this!")

adaptor_generic.py

@@ -0,0 +1,29 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+from argparse import Namespace
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from .adaptor_base import AdaptorBase, AdaptorInput, RadioOutput
+from .adaptor_mlp import create_mlp_from_state
+
+
+class GenericAdaptor(AdaptorBase):
+    def __init__(self, main_config: Namespace, adaptor_config, state):
+        super().__init__()
+
+        self.head_mlp = create_mlp_from_state(main_config.mlp_version, state, 'summary.')
+        self.feat_mlp = create_mlp_from_state(main_config.mlp_version, state, 'feature.')
+
+    def forward(self, input: AdaptorInput) -> RadioOutput:
+        summary = self.head_mlp(input.summary)
+        feat = self.feat_mlp(input.features)
+
+        return RadioOutput(summary, feat)

adaptor_mlp.py

@@ -0,0 +1,150 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+import math
+from typing import Dict
+
+import torch
+from torch import nn
+
+from einops import rearrange
+from timm.models.vision_transformer import Block
+
+
+class MLP(nn.Module):
+    def __init__(self, input_size: int, hidden_size: int, output_size: int,
+                 num_inner: int = 0, device: torch.device = None, **kwargs):
+        super(MLP, self).__init__()
+        self.fc1 = nn.Linear(input_size, hidden_size, device=device)
+        self.norm = nn.LayerNorm(hidden_size, device=device)
+        self.relu = nn.ReLU()
+
+        inner = []
+        for _ in range(num_inner):
+            inner.extend([
+                nn.Linear(hidden_size, hidden_size, device=device),
+                nn.LayerNorm(hidden_size, device=device),
+                nn.ReLU(),
+            ])
+        if inner:
+            self.inner = nn.Sequential(*inner)
+        else:
+            self.inner = nn.Identity()
+
+        self.fc2 = nn.Linear(hidden_size, output_size, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.fc1(x)
+        x = self.norm(x)
+        x = self.relu(x)
+        x = self.inner(x)
+        x = self.fc2(x)
+        return x
+
+
+class MLP2(nn.Module):
+    def __init__(self, input_size: int, hidden_size: int, output_size: int,
+                 num_inner: int = 0,
+                 pre_norm: bool = False, device: torch.device = None,
+                 upsample_factor: int = 1,
+                 **kwargs):
+        super().__init__()
+
+        self.pre_norm = nn.Sequential(
+            nn.LayerNorm(input_size),
+            nn.GELU(),
+        ) if pre_norm else nn.Identity()
+
+        self.upsample_factor = upsample_factor
+        self._real_output_dim = output_size
+
+        hidden_size *= upsample_factor
+        output_size *= (upsample_factor ** 2)
+
+        self.fc1 = nn.Linear(input_size, hidden_size, device=device)
+
+        blocks = []
+        for _ in range(num_inner):
+            blocks.append(nn.Sequential(
+                nn.LayerNorm(hidden_size, device=device),
+                nn.GELU(),
+                nn.Linear(hidden_size, hidden_size, device=device),
+            ))
+        self.blocks = nn.ModuleList(blocks)
+
+        self.final = nn.Sequential(
+            nn.LayerNorm(hidden_size, device=device),
+            nn.GELU(),
+            nn.Linear(hidden_size, output_size, device=device),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.pre_norm(x)
+        x = self.fc1(x)
+        for block in self.blocks:
+            x = x + block(x)
+        x = self.final(x)
+
+        if self.upsample_factor > 1:
+            h = w = int(math.sqrt(x.shape[1]))
+            x = rearrange(x, 'b (h w) (u1 u2 c) -> b (u1 h u2 w) c',
+                          h=h, w=w, u1=self.upsample_factor, u2=self.upsample_factor,
+                          c=self._real_output_dim)
+
+        return x
+
+
+MLP_FACTORY = {
+    'v1': MLP,
+    'v2': MLP2,
+}
+
+
+def strip_prefix(state: Dict[str, torch.Tensor], prefix: str):
+    state = {
+        k[len(prefix):]: v
+        for k, v in state.items()
+        if k.startswith(prefix)
+    }
+    return state
+
+
+def get_mlp_info_from_state(version: str, state: Dict[str, torch.Tensor], prefix: str = ''):
+    state = strip_prefix(state, prefix)
+
+    if version == 'v1':
+        hidden_dim, input_dim = state['fc1.weight'].shape
+        output_dim = state['fc2.weight'].shape[0]
+
+        for num_inner in range(1000):
+            k = f'inner.{num_inner}.0.weight'
+            if k not in state:
+                break
+    elif version == 'v2':
+        hidden_dim, input_dim = state['fc1.weight'].shape
+        output_dim = state['final.2.weight'].shape[0]
+
+        for num_inner in range(1000):
+            k = f'blocks.{num_inner}.0.weight'
+            if k not in state:
+                break
+    else:
+        raise ValueError(f'Unsupported MLP version: {version}')
+
+    return input_dim, hidden_dim, output_dim, num_inner
+
+
+def create_mlp_from_state(version: str, state: Dict[str, torch.Tensor], prefix: str = ''):
+    state = strip_prefix(state, prefix)
+
+    input_dim, hidden_dim, output_dim, num_inner = get_mlp_info_from_state(version, state)
+
+    ret: nn.Module = MLP_FACTORY[version](input_dim, hidden_dim, output_dim, num_inner)
+
+    ret.load_state_dict(state)
+
+    return ret

adaptor_registry.py

@@ -0,0 +1,37 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+from argparse import Namespace
+from typing import Dict, Any
+
+import torch
+
+from .adaptor_generic import GenericAdaptor, AdaptorBase
+
+dict_t = Dict[str, Any]
+state_t = Dict[str, torch.Tensor]
+
+
+class AdaptorRegistry:
+    def __init__(self):
+        self._registry = {}
+
+    def register_adaptor(self, name):
+        def decorator(factory_function):
+            if name in self._registry:
+                raise ValueError(f"Model '{name}' already registered")
+            self._registry[name] = factory_function
+            return factory_function
+        return decorator
+
+    def create_adaptor(self, name, main_config: Namespace, adaptor_config: dict_t, state: state_t) -> AdaptorBase:
+        if name not in self._registry:
+            return GenericAdaptor(main_config, adaptor_config, state)
+        return self._registry[name](main_config, adaptor_config, state)
+
+# Creating an instance of the registry
+adaptor_registry = AdaptorRegistry()

cls_token.py

@@ -0,0 +1,55 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import torch
+from torch import nn
+
+
+class ClsToken(nn.Module):
+    def __init__(self, ndim: int,
+                 num_tokens: int = 1,
+                 enabled: bool = True,
+                 register_multiple: int = 0,
+    ):
+        super().__init__()
+
+        self.ndim = ndim
+        self.enabled = enabled
+        self.num_registers = 0
+        self.num_tokens = num_tokens
+        if enabled:
+            if register_multiple > 0:
+                self.num_registers = register_multiple - (num_tokens % register_multiple)
+
+            scale = ndim ** -0.5
+            self.token = nn.Parameter(torch.randn(num_tokens + self.num_registers, ndim) * scale)
+        else:
+            self.token = None
+
+        self.num_patches = self.num_tokens + self.num_registers
+
+    def disable(self):
+        self.token = None
+        self.enabled = False
+
+    def forward(self, x: torch.Tensor):
+        if self.token is None:
+            return x
+
+        token = self.token.unsqueeze(0).expand(x.shape[0], -1, -1)
+        x = torch.cat([
+            token,
+            x,
+        ], dim=1)
+
+        return x
+
+    def no_weight_decay(self):
+        return [
+            'token',
+        ]

common.py

@@ -0,0 +1,71 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from dataclasses import dataclass
+from typing import Optional
+
+from .radio_model import Resolution
+
+
+@dataclass
+class RadioResource:
+    url: str
+    patch_size: int
+    max_resolution: int
+    preferred_resolution: Resolution
+    vitdet_num_windowed: Optional[int] = None
+    vitdet_num_global: Optional[int] = None
+
+
+RESOURCE_MAP = {
+    # RADIOv2.5
+    "radio_v2.5-b": RadioResource(
+        "https://huggingface.co/nvidia/RADIO/resolve/main/radio-v2.5-b_half.pth.tar?download=true",
+        patch_size=16,
+        max_resolution=2048,
+        preferred_resolution=(768, 768),
+        vitdet_num_global=4,
+    ),
+    "radio_v2.5-l": RadioResource(
+        "https://huggingface.co/nvidia/RADIO/resolve/main/radio-v2.5-l_half.pth.tar?download=true",
+        patch_size=16,
+        max_resolution=2048,
+        preferred_resolution=(768, 768),
+        vitdet_num_global=4,
+    ),
+    # RADIO
+    "radio_v2.1": RadioResource(
+        "https://huggingface.co/nvidia/RADIO/resolve/main/radio_v2.1_bf16.pth.tar?download=true",
+        patch_size=16,
+        max_resolution=2048,
+        preferred_resolution=Resolution(432, 432),
+        vitdet_num_windowed=5,
+    ),
+    "radio_v2": RadioResource(
+        "https://huggingface.co/nvidia/RADIO/resolve/main/radio_v2.pth.tar?download=true",
+        patch_size=16,
+        max_resolution=2048,
+        preferred_resolution=Resolution(432, 432),
+        vitdet_num_windowed=5,
+    ),
+    "radio_v1": RadioResource(
+        "https://huggingface.co/nvidia/RADIO/resolve/main/radio_v1.pth.tar?download=true",
+        patch_size=14,
+        max_resolution=1050,
+        preferred_resolution=Resolution(378, 378),
+    ),
+    # E-RADIO
+    "e-radio_v2": RadioResource(
+        "https://huggingface.co/nvidia/RADIO/resolve/main/eradio_v2.pth.tar?download=true",
+        patch_size=16,
+        max_resolution=2048,
+        preferred_resolution=Resolution(512, 512),
+    ),
+}
+
+DEFAULT_VERSION = "radio_v2.5-l"

config.json

@@ -0,0 +1,221 @@
+{
+  "adaptor_names": null,
+  "architectures": [
+    "RADIOModel"
+  ],
+  "args": {
+    "aa": null,
+    "amp": true,
+    "amp_dtype": "bfloat16",
+    "amp_impl": "native",
+    "aug_repeats": 0,
+    "aug_splits": 0,
+    "bn_eps": null,
+    "bn_momentum": null,
+    "cache_dir": null,
+    "channels_last": false,
+    "checkpoint_hist": 10,
+    "chk_keep_forever": 50,
+    "class_map": "",
+    "clip_grad": null,
+    "clip_mode": "norm",
+    "cls_token_per_teacher": true,
+    "coco_annotations_file": "/datasets/coco2017-adlsa/annotations/captions_val2017.json",
+    "coco_image_dir": "/datasets/coco2017-adlsa/val2017",
+    "color_jitter": 0.4,
+    "cooldown_epochs": 0,
+    "cpe_max_size": 2048,
+    "crd_loss": false,
+    "crd_loss_weight": 0.8,
+    "crop_pct": null,
+    "cutmix": 0.0,
+    "cutmix_minmax": null,
+    "dataset_download": false,
+    "debug_full_knn": false,
+    "decay_epochs": 90,
+    "decay_milestones": [
+      90,
+      180,
+      270
+    ],
+    "decay_rate": 0.1,
+    "device": "cuda:0",
+    "dist_bn": "reduce",
+    "dist_norm_weight": 0.0,
+    "distributed": true,
+    "drop": 0.0,
+    "drop_block": null,
+    "drop_connect": null,
+    "drop_path": null,
+    "dtype": "float32",
+    "epoch_repeats": 0.0,
+    "eval": false,
+    "eval_metric": "knn_top1",
+    "eval_teacher": false,
+    "eval_teacher_only": false,
+    "eval_throughput": false,
+    "fast_norm": false,
+    "fd_loss_fn": "MSE",
+    "feature_normalization": "SHIP_NORM",
+    "feature_summarizer": "cls_token",
+    "feature_upscale_factor": null,
+    "force_new_wandb_id": false,
+    "force_spectral_reparam": true,
+    "freeze_bn": false,
+    "fsdp": false,
+    "fuser": "",
+    "gp": null,
+    "grad_accum_steps": 1,
+    "grad_checkpointing": false,
+    "head_init_bias": null,
+    "head_init_scale": null,
+    "head_warmup": 5,
+    "head_weight_decay": 5e-05,
+    "hflip": 0.5,
+    "img_size": null,
+    "in_chans": null,
+    "initial_checkpoint": "",
+    "input_size": null,
+    "interpolation": "",
+    "layer_decay": null,
+    "local_rank": 0,
+    "log_interval": 50,
+    "log_mlflow": false,
+    "log_wandb": true,
+    "loss_auto_balance": false,
+    "lr_base": 0.1,
+    "lr_base_scale": "",
+    "lr_base_size": 256,
+    "lr_cycle_decay": 0.5,
+    "lr_cycle_limit": 1,
+    "lr_cycle_mul": 1.0,
+    "lr_k_decay": 1.0,
+    "lr_noise": null,
+    "lr_noise_pct": 0.67,
+    "lr_noise_std": 1.0,
+    "mean": null,
+    "mesa": false,
+    "min_lr": 0,
+    "mixup": 0.0,
+    "mixup_mode": "batch",
+    "mixup_off_epoch": 0,
+    "mixup_prob": 1.0,
+    "mixup_switch_prob": 0.5,
+    "mlp_hidden_size": 1520,
+    "mlp_num_inner": 1,
+    "mlp_version": "v2",
+    "model": "vit_large_patch16_224",
+    "model_kwargs": {},
+    "model_norm": false,
+    "momentum": 0.9,
+    "no_aug": false,
+    "no_ddp_bb": true,
+    "no_prefetcher": false,
+    "no_resume_opt": false,
+    "num_classes": null,
+    "opt_betas": null,
+    "opt_eps": null,
+    "patience_epochs": 10,
+    "pin_mem": false,
+    "prefetcher": true,
+    "pretrained": false,
+    "rank": 0,
+    "ratio": [
+      0.75,
+      1.3333333333333333
+    ],
+    "recount": 1,
+    "recovery_interval": 0,
+    "register_multiple": 8,
+    "remode": "pixel",
+    "reprob": 0.0,
+    "reset_loss_state": true,
+    "resplit": false,
+    "save_images": false,
+    "scale": [
+      0.5,
+      1.0
+    ],
+    "sched": "cosine",
+    "seed": 42,
+    "smoothing": 0.1,
+    "spectral_reparam": false,
+    "split_bn": false,
+    "start_epoch": null,
+    "std": null,
+    "sync_bn": false,
+    "synchronize_step": false,
+    "teachers": [
+      {
+        "fd_normalize": false,
+        "feature_distillation": true,
+        "input_size": 378,
+        "model": "ViT-H-14-378-quickgelu",
+        "name": "clip",
+        "pretrained": "dfn5b",
+        "type": "open_clip"
+      },
+      {
+        "fd_normalize": false,
+        "feature_distillation": true,
+        "input_size": 378,
+        "model": "ViT-SO400M-14-SigLIP-384",
+        "name": "siglip",
+        "pretrained": "webli",
+        "type": "open_clip",
+        "use_summary": true
+      },
+      {
+        "fd_normalize": false,
+        "feature_distillation": true,
+        "input_size": 378,
+        "model": "dinov2_vitg14_reg",
+        "name": "dino_v2",
+        "type": "dino_v2"
+      },
+      {
+        "fd_normalize": false,
+        "feature_distillation": true,
+        "input_size": 1024,
+        "model": "vit-h",
+        "name": "sam",
+        "type": "sam",
+        "use_summary": false
+      }
+    ],
+    "torchcompile": null,
+    "torchscript": false,
+    "train_interpolation": "random",
+    "train_split": "train",
+    "tta": 0,
+    "use_coco": false,
+    "use_multi_epochs_loader": false,
+    "val_ema_only": false,
+    "val_split": "val",
+    "vflip": 0.0,
+    "vitdet_version": 1,
+    "wandb_entity": "",
+    "wandb_job_type": "",
+    "wandb_name": "",
+    "wandb_project": "",
+    "warmup_lr": 1e-05,
+    "warmup_prefix": false,
+    "worker_seeding": "all",
+    "workers": 8,
+    "world_size": 128
+  },
+  "auto_map": {
+    "AutoConfig": "hf_model.RADIOConfig",
+    "AutoModel": "hf_model.RADIOModel"
+  },
+  "max_resolution": 2048,
+  "patch_size": 16,
+  "preferred_resolution": [
+    768,
+    768
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.40.1",
+  "version": "radio_v2.5-l",
+  "vitdet_window_size": null
+}

enable_cpe_support.py

@@ -0,0 +1,158 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from typing import List, Optional, Set, Tuple, Union
+from types import MethodType
+
+import torch
+from torch import nn
+
+from timm.models import VisionTransformer, checkpoint_seq
+
+from .vit_patch_generator import ViTPatchGenerator
+
+
+def _forward_cpe(self: VisionTransformer, x: torch.Tensor) -> torch.Tensor:
+    x = self.patch_generator(x)
+    if self.grad_checkpointing and not torch.jit.is_scripting():
+        x = checkpoint_seq(self.blocks, x)
+    else:
+        x = self.blocks(x)
+    x = self.norm(x)
+    return x
+
+
+def _take_indices(
+        num_blocks: int,
+        n: Optional[Union[int, List[int], Tuple[int]]],
+) -> Tuple[Set[int], int]:
+    if isinstance(n, int):
+        assert n >= 0
+        take_indices = {x for x in range(num_blocks - n, num_blocks)}
+    else:
+        take_indices = {num_blocks + idx if idx < 0 else idx for idx in n}
+    return take_indices, max(take_indices)
+
+
+def _forward_intermediates_cpe(
+        self,
+        x: torch.Tensor,
+        indices: Optional[Union[int, List[int], Tuple[int]]] = None,
+        return_prefix_tokens: bool = False,
+        norm: bool = False,
+        stop_early: bool = False,
+        output_fmt: str = 'NCHW',
+        intermediates_only: bool = False,
+        aggregation: Optional[str] = "sparse",
+) -> Union[List[torch.Tensor], Tuple[torch.Tensor, List[torch.Tensor]]]:
+    """ Forward features that returns intermediates.
+
+    The Dense layer aggregation method is inspired from the paper: "Dense Connector for MLLMs"
+    by Yao, Huanjin et al. (2024). arXiv preprint arXiv:2405.13800}
+
+    Args:
+        x: Input image tensor
+        indices: Take last n blocks if int, select matching indices if sequence
+        return_prefix_tokens: Return both prefix and spatial intermediate tokens
+        norm: Apply norm layer to all intermediates
+        stop_early: Stop iterating over blocks when last desired intermediate hit
+        output_fmt: Shape of intermediate feature outputs
+        intermediates_only: Only return intermediate features
+        aggregation: intermediate layer aggregation method (sparse or dense)
+    Returns:
+    """
+    assert output_fmt in ('NCHW', 'NLC'), 'Output format must be one of NCHW or NLC.'
+    assert aggregation in ('sparse', 'dense'), 'Aggregation must be one of sparse or dense.'
+    reshape = output_fmt == 'NCHW'
+    intermediates = []
+    take_indices, max_index = _take_indices(len(self.blocks), indices)
+    # forward pass
+    B, _, height, width = x.shape
+    x = self.patch_generator(x)
+
+    if not stop_early:  # can't slice blocks in torchscript
+        blocks = self.blocks
+    else:
+        blocks = self.blocks[:max_index + 1]
+
+    accumulator = 0
+    num_accumulated = 0
+
+    for i, blk in enumerate(blocks):
+        x = blk(x)
+        if aggregation == "dense":
+            accumulator = accumulator + x
+            num_accumulated += 1
+        if i in take_indices:
+            if aggregation == "dense":
+                x_ = accumulator / num_accumulated
+                num_accumulated = 0
+                accumulator = 0
+            else:
+                 x_ = x
+            # normalize intermediates with final norm layer if enabled
+            intermediates.append(self.norm(x_) if norm else x_)
+
+    # process intermediates
+
+    # split prefix (e.g. class, distill) and spatial feature tokens
+    prefix_tokens = [y[:, 0:self.patch_generator.num_cls_tokens] for y in intermediates]
+    intermediates = [y[:, self.patch_generator.num_skip:] for y in intermediates]
+
+    if reshape:
+        # reshape to BCHW output format
+        H = height // self.patch_generator.patch_size
+        W = width // self.patch_generator.patch_size
+        intermediates = [y.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() for y in intermediates]
+    if not torch.jit.is_scripting() and return_prefix_tokens:
+        # return_prefix not support in torchscript due to poor type handling
+        intermediates = list(zip(intermediates, prefix_tokens))
+    if intermediates_only:
+        return intermediates
+    x = self.norm(x)
+    return x, intermediates
+
+def enable_cpe(model: nn.Module,
+               max_img_size: Union[int, Tuple[int, int]] = 1024,
+               num_cls_tokens: int = 1,
+               pos_dropout: float = 0.1,
+               register_multiple: int = 0,
+):
+    if not isinstance(model, VisionTransformer):
+        raise ValueError("CPE only support for VisionTransformer models!")
+
+    patch_size = model.patch_embed.patch_size[0]
+    embed_dim = model.embed_dim
+    input_dims = model.patch_embed.img_size
+    normalize_patches = not isinstance(model.patch_embed.norm, nn.Identity)
+    cls_token = model.cls_token is not None
+
+    max_img_size = int(round(max_img_size / patch_size) * patch_size)
+
+    patch_generator = ViTPatchGenerator(
+        patch_size=patch_size,
+        embed_dim=embed_dim,
+        input_dims=input_dims,
+        normalize_patches=normalize_patches,
+        cls_token=cls_token,
+        max_input_dims=max_img_size,
+        pos_dropout=pos_dropout,
+        num_cls_tokens=num_cls_tokens,
+        register_multiple=register_multiple,
+    )
+
+    model.patch_generator = patch_generator
+    model.patch_embed = None
+    model.cls_token = None
+    model.pos_embed = None
+    model.pos_drop = None
+    model.num_cls_tokens = num_cls_tokens
+    model.num_registers = patch_generator.num_registers
+
+    model.forward_features = MethodType(_forward_cpe, model)
+    model.forward_intermediates = MethodType(_forward_intermediates_cpe, model)

enable_spectral_reparam.py

@@ -0,0 +1,227 @@
+from logging import getLogger
+import math
+import os
+from typing import Union, Tuple
+from types import MethodType
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.nn.utils import parametrize
+from torch.nn.utils.parametrizations import _SpectralNorm
+
+from timm.models.vision_transformer import Attention, Mlp
+
+_EPS = 1e-5
+
+
+class _SNReweight(_SpectralNorm):
+    def __init__(self, weight: torch.Tensor, *args, init_norm_to_current: bool = False, alpha: float = 0.05, version: int = 2, **kwargs):
+        super().__init__(weight, *args, **kwargs)
+
+        self.alpha = alpha
+        self.version = version
+        self.register_buffer('_sn_version', torch.tensor(version))
+
+        if init_norm_to_current:
+            # This will set the numerator to match the denominator, which should preserve the original values
+            init_scale = self._get_sigma(weight).item()
+        else:
+            init_scale = 1.0
+
+        if version == 1:
+            init_value = init_scale
+        elif version == 2:
+            t = init_scale - alpha
+            if t < _EPS:
+                getLogger("spectral_reparam").warn(f'The initialized spectral norm {init_scale} is too small to be represented. Setting to {_EPS} instead.')
+                t = _EPS
+
+            init_value = math.log(math.exp(t) - 1)
+        else:
+            raise ValueError(f'Unsupported version: {version}')
+
+        # Make 2D so that weight decay gets applied
+        self.scale = nn.Parameter(torch.tensor([[init_value]], dtype=torch.float32, device=weight.device))
+
+    # Re-implementing this because we need to make division by sigma safe
+    def _get_sigma(self, weight: torch.Tensor) -> torch.Tensor:
+        if weight.ndim == 1:
+            # Faster and more exact path, no need to approximate anything
+            sigma = weight.norm()
+        else:
+            weight_mat = self._reshape_weight_to_matrix(weight)
+            if self.training:
+                self._power_method(weight_mat, self.n_power_iterations)
+            # See above on why we need to clone
+            u = self._u.clone(memory_format=torch.contiguous_format)
+            v = self._v.clone(memory_format=torch.contiguous_format)
+            # The proper way of computing this should be through F.bilinear, but
+            # it seems to have some efficiency issues:
+            # https://github.com/pytorch/pytorch/issues/58093
+            sigma = torch.dot(u, torch.mv(weight_mat, v))
+
+        return sigma + self.eps
+
+    def forward(self, weight: torch.Tensor, *args, **kwargs):
+        dtype = weight.dtype
+        sigma = self._get_sigma(weight, *args, **kwargs)
+
+        if self.version == 1:
+            scale = self.scale
+        elif self.version == 2:
+            scale = F.softplus(self.scale) + self.alpha
+        else:
+            raise ValueError(f'Unsupported version: {self.version}')
+
+        scale = scale.float() / sigma.float()
+
+        y = weight * scale
+
+        if dtype in (torch.float16, torch.bfloat16):
+            y = y.to(dtype)
+        return y
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+        version_key = f'{prefix}_sn_version'
+        if version_key not in state_dict:
+            self.version = 1
+            state_dict[version_key] = torch.tensor(1)
+        return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+
+
+class _AttnSNReweight(nn.Module):
+    def __init__(self, weight: torch.Tensor, *args, init_norm_to_current: bool = False, renorm_values: bool = False, **kwargs):
+        super().__init__()
+
+        parts = weight.split(weight.shape[0] // 3, dim=0)
+
+        ct = 2 if not renorm_values else 3
+
+        self.parts = nn.ModuleList([
+            _SNReweight(p, *args, init_norm_to_current=init_norm_to_current, **kwargs) if i < ct else nn.Identity()
+            for i, p in enumerate(parts)
+        ])
+
+    def forward(self, weight: torch.Tensor, *args, **kwargs):
+        parts = weight.split(weight.shape[0] // 3, dim=0)
+
+        parts = [
+            fn(p)
+            for fn, p in zip(self.parts, parts)
+        ]
+
+        return torch.cat(parts, dim=0)
+
+
+def enable_spectral_reparam(model: nn.Module,
+                            n_power_iterations: int = 1,
+                            eps: float = 1e-6,
+                            init_norm_to_current: bool = False,
+                            renorm_values: bool = True,
+                            renorm_mlp: bool = True):
+    # print('Enabling spectral reparametrization')
+    for mod in model.modules():
+        if isinstance(mod, Attention):
+            parametrize.register_parametrization(
+                mod.qkv,
+                'weight',
+                _AttnSNReweight(mod.qkv.weight, n_power_iterations, dim=0, eps=eps, init_norm_to_current=init_norm_to_current, renorm_values=renorm_values),
+            )
+            pass
+        elif isinstance(mod, Mlp) and renorm_mlp:
+            parametrize.register_parametrization(
+                mod.fc1,
+                'weight',
+                _SNReweight(mod.fc1.weight, n_power_iterations, dim=0, eps=eps, init_norm_to_current=init_norm_to_current),
+            )
+            parametrize.register_parametrization(
+                mod.fc2,
+                'weight',
+                _SNReweight(mod.fc2.weight, n_power_iterations, dim=0, eps=eps, init_norm_to_current=init_norm_to_current),
+            )
+            pass
+
+
+def configure_spectral_reparam_from_args(model: nn.Module, args):
+    spectral_reparam = getattr(args, 'spectral_reparam', False)
+    if isinstance(spectral_reparam, bool) and spectral_reparam:
+        enable_spectral_reparam(model, init_norm_to_current=args.pretrained)
+    elif isinstance(spectral_reparam, dict):
+        enable_spectral_reparam(
+            model,
+            n_power_iterations=spectral_reparam.get('n_power_iterations', 1),
+            eps=spectral_reparam.get('eps', 1e-12),
+            init_norm_to_current=args.pretrained,
+        )
+
+
+def disable_spectral_reparam(model: nn.Module):
+    for mod in model.modules():
+        if isinstance(mod, Attention):
+            parametrize.remove_parametrizations(mod.qkv, 'weight')
+            pass
+        elif isinstance(mod, Mlp):
+            parametrize.remove_parametrizations(mod.fc1, 'weight')
+            parametrize.remove_parametrizations(mod.fc2, 'weight')
+            pass
+
+
+if __name__ == '__main__':
+    import argparse
+    from . import radio_model as create_model
+
+    parser = argparse.ArgumentParser(description='Remove parametrization from state dict')
+    parser.add_argument('--checkpoint', type=str, required=True, help='The checkpoint to load')
+    parser.add_argument('--output', type=str, default='', help='Where to store the checkpoint')
+    parser.add_argument('--release', default=False, action='store_true', help='Prune extraneous checkpoint fields')
+    parser.add_argument('--strict', default=False, action='store_true', help='Strictly load the state dict')
+
+    args = parser.parse_args()
+
+    if not args.output:
+        chk_dir, chk_name = os.path.split(args.checkpoint)
+        args.output = os.path.join(chk_dir, f'clean_{chk_name}')
+        print(f'Set output to "{args.output}"')
+
+    chk = torch.load(args.checkpoint, map_location='cpu', mmap=True)
+
+    model = create_model.create_model_from_args(chk['args'])
+
+    key = 'base_model.'
+    mod_state = dict()
+    extra_state = dict()
+    for k, v in chk['state_dict'].items():
+        if k.startswith(key):
+            mod_state[k[len(key):]] = v
+        else:
+            extra_state[k] = v
+
+    chk_load_info = model.load_state_dict(mod_state, strict=args.strict)
+    if chk_load_info.unexpected_keys or chk_load_info.missing_keys:
+        print(chk_load_info)
+
+    if chk['args'].spectral_reparam:
+        disable_spectral_reparam(model)
+
+    if hasattr(chk['args'], 'dtype'):
+        model.to(dtype=chk['args'].dtype)
+
+    mod_state = model.state_dict()
+    final_state = dict()
+    final_state.update({f'{key}{k}': v for k, v in mod_state.items()})
+    final_state.update(extra_state)
+
+    chk['state_dict'] = final_state
+    chk['args'].spectral_reparam = False
+
+    if args.release:
+        chk = {
+            'arch': chk['arch'],
+            'epoch': chk['epoch'],
+            'state_dict': chk['state_dict'],
+            'args': chk['args'],
+        }
+
+    torch.save(chk, args.output)
+    pass

eradio_model.py

@@ -0,0 +1,1392 @@
+#!/usr/bin/env python3
+
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# E-RADIO model from
+# Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. "AM-RADIO: Agglomerative Model--Reduce All Domains Into One." arXiv preprint arXiv:2312.06709 (2023).
+
+# based on FasterViT, Swin Transformer, YOLOv8
+
+# FasterViT:
+# Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M. Alvarez, Jan Kautz, and Pavlo Molchanov. "FasterViT: Fast Vision Transformers with Hierarchical Attention." arXiv preprint arXiv:2306.06189 (2023).
+
+import timm
+import torch
+import torch.nn as nn
+from timm.models.registry import register_model
+
+from timm.models.layers import trunc_normal_, DropPath, LayerNorm2d
+import numpy as np
+import torch.nn.functional as F
+import math
+import warnings
+
+#######################
+## Codebase from YOLOv8
+## BEGINNING
+#######################
+
+class C2f(nn.Module):
+    """Faster Implementation of CSP Bottleneck with 2 convolutions."""
+    """From YOLOv8 codebase"""
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, drop_path=None):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        if drop_path is None:
+            drop_path = [0.0] * n
+
+        self.c = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, 2 * self.c, 1, 1)
+        self.cv2 = Conv((2 + n) * self.c, c2, 1)  # optional act=FReLU(c2)
+        self.m = nn.ModuleList(Bottleneck(self.c, self.c, shortcut, g, k=((3, 3), (3, 3)), e=1.0, drop_path=drop_path[i]) for i in range(n))
+
+    def forward(self, x):
+        """Forward pass through C2f layer."""
+        y = list(self.cv1(x).chunk(2, 1))
+        y.extend(m(y[-1]) for m in self.m)
+        return self.cv2(torch.cat(y, 1))
+
+    def forward_split(self, x):
+        """Forward pass using split() instead of chunk()."""
+        y = list(self.cv1(x).split((self.c, self.c), 1))
+        y.extend(m(y[-1]) for m in self.m)
+        return self.cv2(torch.cat(y, 1))
+
+class Bottleneck(nn.Module):
+    """Standard bottleneck."""
+
+    def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5, drop_path=0.0):  # ch_in, ch_out, shortcut, groups, kernels, expand
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, k[0], 1)
+        self.cv2 = Conv(c_, c2, k[1], 1, g=g)
+        self.add = shortcut and c1 == c2
+        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+    def forward(self, x):
+        """'forward()' applies the YOLOv5 FPN to input data."""
+        return x + self.drop_path1(self.cv2(self.cv1(x))) if self.add else self.cv2(self.cv1(x))
+
+
+class Conv(nn.Module):
+    """Modified to support layer fusion"""
+    default_act = nn.SiLU()  # default activation
+
+    def __init__(self, a, b, kernel_size=1, stride=1, padding=None, g=1, dilation=1, bn_weight_init=1, bias=False, act=True):
+        super().__init__()
+
+        self.conv = torch.nn.Conv2d(a, b, kernel_size, stride, autopad(kernel_size, padding, dilation), dilation, g, bias=False)
+        if 1:
+            self.bn = torch.nn.BatchNorm2d(b)
+            torch.nn.init.constant_(self.bn.weight, bn_weight_init)
+            torch.nn.init.constant_(self.bn.bias, 0)
+        self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
+
+
+    def forward(self,x):
+        x = self.conv(x)
+        x = self.bn(x)
+        x = self.act(x)
+        return x
+
+    @torch.no_grad()
+    def switch_to_deploy(self):
+        # return 1
+        if not isinstance(self.bn, nn.Identity):
+            c, bn = self.conv, self.bn
+            w = bn.weight / (bn.running_var + bn.eps) ** 0.5
+            w = c.weight * w[:, None, None, None]
+            b = bn.bias - bn.running_mean * bn.weight / \
+                (bn.running_var + bn.eps)**0.5
+
+            self.conv.weight.data.copy_(w)
+            self.conv.bias = nn.Parameter(b)
+
+            self.bn = nn.Identity()
+
+def autopad(k, p=None, d=1):  # kernel, padding, dilation
+    """Pad to 'same' shape outputs."""
+    if d > 1:
+        k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]  # actual kernel-size
+    if p is None:
+        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
+    return p
+
+
+#######################
+## Codebase from YOLOv8
+## END
+#######################
+
+def pixel_unshuffle(data, factor=2):
+    # performs nn.PixelShuffle(factor) in reverse, torch has some bug for ONNX and TRT, so doing it manually
+    B, C, H, W = data.shape
+    return data.view(B, C, factor, H//factor, factor, W//factor).permute(0,1,2,4,3,5).reshape(B, -1, H//factor, W//factor)
+
+class SwiGLU(nn.Module):
+    # should be more advanced, but doesnt improve results so far
+    def forward(self, x):
+        x, gate = x.chunk(2, dim=-1)
+        return F.silu(gate) * x
+
+
+def window_partition(x, window_size):
+    """
+    Function for partitioning image into windows and later do windowed attention
+    Args:
+        x: (B, C, H, W)
+        window_size: window size
+    Returns:
+        windows - local window features (num_windows*B, window_size*window_size, C)
+        (Hp, Wp) -  the size of the padded image
+    """
+    B, C, H, W = x.shape
+
+    if window_size == 0 or (window_size==H and window_size==W):
+        windows = x.flatten(2).transpose(1, 2)
+        Hp, Wp = H, W
+    else:
+        pad_h = (window_size - H % window_size) % window_size
+        pad_w = (window_size - W % window_size) % window_size
+        if pad_h > 0 or pad_w > 0:
+            x = F.pad(x, (0, pad_w, 0, pad_h), mode="reflect")
+        Hp, Wp = H + pad_h, W + pad_w
+
+        x = x.view(B, C, Hp // window_size, window_size, Wp // window_size, window_size)
+        windows = x.permute(0, 2, 4, 3, 5, 1).reshape(-1, window_size*window_size, C)
+
+    return windows, (Hp, Wp)
+
+class Conv2d_BN(nn.Module):
+    '''
+    Conv2d + BN layer with folding capability to speed up inference
+    Can be merged with Conv() function with additional arguments
+    '''
+    def __init__(self, a, b, kernel_size=1, stride=1, padding=0, dilation=1, groups=1, bn_weight_init=1, bias=False):
+        super().__init__()
+        self.conv = torch.nn.Conv2d(a, b, kernel_size, stride, padding, dilation, groups, bias=False)
+        if 1:
+            self.bn = torch.nn.BatchNorm2d(b)
+            torch.nn.init.constant_(self.bn.weight, bn_weight_init)
+            torch.nn.init.constant_(self.bn.bias, 0)
+
+    def forward(self,x):
+        x = self.conv(x)
+        x = self.bn(x)
+        return x
+
+    @torch.no_grad()
+    def switch_to_deploy(self):
+        if not isinstance(self.bn, nn.Identity):
+            c, bn = self.conv, self.bn
+            w = bn.weight / (bn.running_var + bn.eps) ** 0.5
+            w = c.weight * w[:, None, None, None]
+            b = bn.bias - bn.running_mean * bn.weight / \
+                (bn.running_var + bn.eps)**0.5
+            self.conv.weight.data.copy_(w)
+            self.conv.bias = nn.Parameter(b)
+            self.bn = nn.Identity()
+
+
+
+def window_reverse(windows, window_size, H, W, pad_hw):
+    """
+    Windows to the full feature map
+    Args:
+        windows: local window features (num_windows*B, window_size, window_size, C)
+        window_size: Window size
+        H: Height of image
+        W: Width of image
+        pad_w - a tuple of image passing used in windowing step
+    Returns:
+        x: (B, C, H, W)
+
+    """
+    # print(f"window_reverse, windows.shape {windows.shape}")
+    Hp, Wp = pad_hw
+    if window_size == 0 or (window_size==H and window_size==W):
+        B = int(windows.shape[0] / (Hp * Wp / window_size / window_size))
+        x = windows.transpose(1, 2).view(B, -1, H, W)
+    else:
+        B = int(windows.shape[0] / (Hp * Wp / window_size / window_size))
+        x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+        x = x.permute(0, 5, 1, 3, 2, 4).reshape(B,windows.shape[2], Hp, Wp)
+
+        if Hp > H or Wp > W:
+            x = x[:, :, :H, :W, ].contiguous()
+
+    return x
+
+
+
+class PosEmbMLPSwinv2D(nn.Module):
+    """
+    2D positional embedding from Swin Transformer v2
+    Added functionality to store the positional embedding in the model and not recompute it every time
+    """
+    def __init__(
+        self, window_size, pretrained_window_size, num_heads, seq_length, no_log=False, cpb_mlp_hidden=512,
+    ):
+        super().__init__()
+        self.window_size = window_size
+        self.num_heads = num_heads
+        # mlp to generate continuous relative position bias
+        self.cpb_mlp = nn.Sequential(
+            nn.Linear(2, cpb_mlp_hidden, bias=True),
+            nn.ReLU(inplace=True),
+            nn.Linear(cpb_mlp_hidden, num_heads, bias=False),
+        )
+
+        self.grid_exists = False
+        self.seq_length = seq_length
+        self.deploy = False
+        self.num_heads = num_heads
+        self.no_log = no_log
+        self.pretrained_window_size = pretrained_window_size
+        self.relative_bias_window_size = window_size
+
+        relative_coords_table, relative_position_index, relative_bias = self.relative_bias_initialization(window_size, num_heads,
+                                                                                                     pretrained_window_size, seq_length,
+                                                                                                     no_log)
+
+        self.register_buffer("relative_coords_table", relative_coords_table)
+        self.register_buffer("relative_position_index", relative_position_index)
+        self.register_buffer("relative_bias", relative_bias)  # for EMA
+
+    def relative_bias_initialization(self, window_size, num_heads, pretrained_window_size, seq_length, no_log):
+        # as in separate function to support window size chage after model weights loading
+        relative_coords_h = torch.arange(
+            -(window_size[0] - 1), window_size[0], dtype=torch.float32
+        )
+        relative_coords_w = torch.arange(
+            -(window_size[1] - 1), window_size[1], dtype=torch.float32
+        )
+        relative_coords_table = (
+            torch.stack(torch.meshgrid([relative_coords_h, relative_coords_w]))
+            .permute(1, 2, 0)
+            .contiguous()
+            .unsqueeze(0)
+        )  # 1, 2*Wh-1, 2*Ww-1, 2
+        if pretrained_window_size[0] > 0:
+            relative_coords_table[:, :, :, 0] /= pretrained_window_size[0] - 1
+            relative_coords_table[:, :, :, 1] /= pretrained_window_size[1] - 1
+        else:
+            relative_coords_table[:, :, :, 0] /= self.window_size[0] - 1
+            relative_coords_table[:, :, :, 1] /= self.window_size[1] - 1
+
+        if not no_log:
+            relative_coords_table *= 8  # normalize to -8, 8
+            relative_coords_table = (
+                torch.sign(relative_coords_table)
+                * torch.log2(torch.abs(relative_coords_table) + 1.0)
+                / np.log2(8)
+            )
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = (
+            coords_flatten[:, :, None] - coords_flatten[:, None, :]
+        )  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(
+            1, 2, 0
+        ).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+
+        relative_bias = torch.zeros(1, num_heads, seq_length, seq_length)
+
+        self.relative_bias_window_size = window_size
+
+        return relative_coords_table, relative_position_index, relative_bias
+
+
+    def switch_to_deploy(self):
+        self.deploy = True
+        self.grid_exists = True
+
+    def forward(self, input_tensor):
+        # for efficiency, we want this forward to be folded into a single operation (sum)
+        # if resolution stays the same, then we dont need to recompute MLP layers
+
+        if not self.deploy or self.training:
+            self.grid_exists = False
+
+        #compare if all elements in self.window_size list match those in self.relative_bias_window_size
+        if not all([self.window_size[i] == self.relative_bias_window_size[i] for i in range(len(self.window_size))]):
+            relative_coords_table, relative_position_index, relative_bias = self.relative_bias_initialization(self.window_size, self.num_heads,
+                                                                                                        self.pretrained_window_size, self.seq_length,
+                                                                                                        self.no_log)
+
+            self.relative_coords_table = relative_coords_table.to(self.relative_coords_table.device)
+            self.relative_position_index = relative_position_index.to(self.relative_position_index.device)
+            self.relative_bias = relative_bias.to(self.relative_bias.device)
+
+        if self.deploy and self.grid_exists:
+            input_tensor = input_tensor + self.relative_bias
+            return input_tensor
+
+        if 1:
+            self.grid_exists = True
+
+            relative_position_bias_table = self.cpb_mlp(
+                self.relative_coords_table
+            ).view(-1, self.num_heads)
+            relative_position_bias = relative_position_bias_table[
+                self.relative_position_index.view(-1)
+            ].view(
+                self.window_size[0] * self.window_size[1],
+                self.window_size[0] * self.window_size[1],
+                -1,
+            )  # Wh*Ww,Wh*Ww,nH
+
+            relative_position_bias = relative_position_bias.permute(
+                2, 0, 1
+            ).contiguous()  # nH, Wh*Ww, Wh*Ww
+            relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
+
+            self.relative_bias = relative_position_bias.unsqueeze(0)
+
+        input_tensor = input_tensor + self.relative_bias
+        return input_tensor
+
+
+class GRAAttentionBlock(nn.Module):
+    def __init__(self, window_size, dim_in, dim_out,
+                 num_heads, drop_path=0., qk_scale=None, qkv_bias=False,
+                 norm_layer=nn.LayerNorm, layer_scale=None,
+                  use_swiglu=True,
+                  subsample_ratio=1, dim_ratio=1, conv_base=False,
+                  do_windowing=True, multi_query=False, use_shift=0,
+                  cpb_mlp_hidden=512, conv_groups_ratio=0):
+        '''
+        Global Resolution Attention Block , see README for details
+        Attention with subsampling to get a bigger receptive field for attention
+        conv_base - use conv2d instead of avgpool2d for downsample / upsample
+
+
+        '''
+        super().__init__()
+
+        self.shift_size=window_size//2 if use_shift else 0
+
+        self.do_windowing = do_windowing
+        self.subsample_ratio = subsample_ratio
+
+
+
+        if do_windowing:
+            if conv_base:
+                    self.downsample_op = nn.Conv2d(dim_in, dim_out, kernel_size=subsample_ratio, stride=subsample_ratio) if subsample_ratio > 1 else nn.Identity()
+
+
+                    self.downsample_mixer = nn.Identity()
+                    self.upsample_mixer = nn.Identity()
+                    self.upsample_op = nn.ConvTranspose2d(dim_in, dim_out, kernel_size=subsample_ratio, stride=subsample_ratio) if subsample_ratio > 1 else nn.Identity()
+            else:
+                self.downsample_op = nn.AvgPool2d(kernel_size=subsample_ratio, stride=subsample_ratio) if subsample_ratio > 1 else nn.Identity()
+                self.downsample_mixer = Conv2d_BN(dim_in, dim_out, kernel_size=1, stride=1) if subsample_ratio > 1 else nn.Identity()
+                self.upsample_mixer = nn.Upsample(scale_factor=subsample_ratio, mode='nearest') if subsample_ratio > 1 else nn.Identity()
+                self.upsample_op = Conv2d_BN(dim_in, dim_out, kernel_size=1, stride=1, padding=0, bias=False) if subsample_ratio > 1 else nn.Identity()
+
+
+        # in case there is no downsampling conv we want to have it separately
+        # will help with information propagation between windows
+        if subsample_ratio == 1:
+            # conv_groups_ratio=0
+            self.pre_conv = Conv2d_BN(dim_in, dim_in, kernel_size=3, stride=1, padding=1, groups=max(1,int(conv_groups_ratio*dim_in)), bias=False)
+            # self.pre_conv = nn.Conv2d(dim_in, dim_in, kernel_size=3, stride=1, padding=1, groups=max(1,int(conv_groups_ratio*dim_in)), bias=False)
+            # self.pre_conv_act = nn.ReLU6()
+            #for simplicity:
+            self.pre_conv_act = nn.Identity()
+            if conv_groups_ratio == -1:
+                self.pre_conv = nn.Identity()
+                self.pre_conv_act = nn.Identity()
+
+        self.window_size = window_size
+
+        self.norm1 = norm_layer(dim_in)
+
+        self.attn = WindowAttention(
+            dim_in,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            resolution=window_size,
+            seq_length=window_size**2, dim_out=dim_in, multi_query=multi_query,
+            shift_size=self.shift_size, cpb_mlp_hidden=cpb_mlp_hidden)
+
+        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+        use_layer_scale = layer_scale is not None and type(layer_scale) in [int, float]
+        self.gamma1 = nn.Parameter(layer_scale * torch.ones(dim_in))  if use_layer_scale else 1
+
+        ### mlp layer
+        mlp_ratio = 4
+        self.norm2 = norm_layer(dim_in)
+        mlp_hidden_dim = int(dim_in * mlp_ratio)
+
+        activation = nn.GELU if not use_swiglu else SwiGLU
+        mlp_hidden_dim = int((4 * dim_in * 1 / 2) / 64) * 64 if use_swiglu else mlp_hidden_dim
+
+        self.mlp = Mlp(in_features=dim_in, hidden_features=mlp_hidden_dim, act_layer=activation, use_swiglu=use_swiglu)
+
+        self.gamma2 = nn.Parameter(layer_scale * torch.ones(dim_in)) if layer_scale else 1
+        self.drop_path2=DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+
+    def forward(self, x):
+        skip_connection = x
+        attn_mask = None
+
+        # in case there is no downsampling conv we want to have it separately
+        # will help with information propagation
+        if self.subsample_ratio == 1:
+            x = self.pre_conv_act(self.pre_conv(x)) + skip_connection
+
+        if self.do_windowing:
+            # performing windowing if required
+            x = self.downsample_op(x)
+            x = self.downsample_mixer(x)
+
+            if self.window_size>0:
+                H, W = x.shape[2], x.shape[3]
+
+            if self.shift_size > 0 and H>self.window_size and W>self.window_size:
+                # @swin like cyclic shift, doesnt show better performance
+                x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(2, 3))
+
+            x, pad_hw = window_partition(x, self.window_size)
+
+            if self.shift_size > 0 and H>self.window_size and W>self.window_size:
+                # set atten matrix to have -100 and the top right square
+                # attn[:, :, :-self.shift_size, -self.shift_size:] = -100.0
+                # calculate attention mask for SW-MSA
+                # not used in final version, can be useful for some cases especially for high res
+                H, W = pad_hw
+                img_mask = torch.zeros((1, H, W, 1), device=x.device)  # 1 H W 1
+                h_slices = (slice(0, -self.window_size),
+                            slice(-self.window_size, -self.shift_size),
+                            slice(-self.shift_size, None))
+                w_slices = (slice(0, -self.window_size),
+                            slice(-self.window_size, -self.shift_size),
+                            slice(-self.shift_size, None))
+                cnt = 0
+                for h in h_slices:
+                    for w in w_slices:
+                        img_mask[:, h, w, :] = cnt
+                        cnt += 1
+                img_mask = img_mask.transpose(1,2).transpose(1,3)
+                mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+
+                mask_windows = mask_windows[0].view(-1, self.window_size * self.window_size)
+                attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+                attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        # window attention
+        x = x + self.drop_path1(self.gamma1*self.attn(self.norm1(x), attn_mask=attn_mask)) # or pass H,W
+        # mlp layer
+        x = x + self.drop_path2(self.gamma2*self.mlp(self.norm2(x)))
+
+        if self.do_windowing:
+            if self.window_size > 0:
+                x = window_reverse(x, self.window_size, H, W, pad_hw)
+
+            # reverse cyclic shift
+            if self.shift_size > 0 and H>self.window_size and W>self.window_size:
+                # @swin like cyclic shift, not tested
+                x = torch.roll(x, shifts=(self.shift_size, self.shift_size), dims=(2, 3))
+
+            x = self.upsample_mixer(x)
+            x = self.upsample_op(x)
+
+
+            if x.shape[2] != skip_connection.shape[2] or x.shape[3] != skip_connection.shape[3]:
+                x = torch.nn.functional.pad(x, ( 0, -x.shape[3] + skip_connection.shape[3], 0, -x.shape[2] + skip_connection.shape[2]), mode="reflect")
+        # need to add skip connection because downsampling and upsampling will break residual connection
+        # 0.5 is needed to make sure that the skip connection is not too strong
+        # in case of no downsample / upsample we can show that 0.5 compensates for the residual connection
+        x = 0.5 * x + 0.5 * skip_connection
+        return x
+
+
+
+
+class MultiResolutionAttention(nn.Module):
+    """
+    MultiResolutionAttention (MRA) module
+    The idea is to use multiple attention blocks with different resolution
+    Feature maps are downsampled / upsampled for each attention block on different blocks
+    Every attention block supports windowing
+    """
+
+    def __init__(self, window_size, sr_ratio,
+                 dim, dim_ratio, num_heads,
+                 do_windowing=True,
+                 layer_scale=1e-5, norm_layer=nn.LayerNorm,
+                 drop_path = 0, qkv_bias=False, qk_scale=1.0,
+                 use_swiglu=True, multi_query=False, conv_base=False,
+                 use_shift=0, cpb_mlp_hidden=512, conv_groups_ratio=0) -> None:
+        """
+        Args:
+            input_resolution: input image resolution
+            window_size: window size
+            compression_ratio: compression ratio
+            max_depth: maximum depth of the GRA module
+            use_shift: do window shifting
+        """
+        super().__init__()
+
+        depth = len(sr_ratio)
+
+        self.attention_blocks = nn.ModuleList()
+
+
+        for i in range(depth):
+            subsample_ratio = sr_ratio[i]
+            if len(window_size) > i:
+                window_size_local = window_size[i]
+            else:
+                window_size_local = window_size[0]
+
+            self.attention_blocks.append(GRAAttentionBlock(window_size=window_size_local,
+                                            dim_in=dim, dim_out=dim, num_heads=num_heads,
+                                            qkv_bias=qkv_bias, qk_scale=qk_scale, norm_layer=norm_layer,
+                                            layer_scale=layer_scale, drop_path=drop_path,
+                                            use_swiglu=use_swiglu, subsample_ratio=subsample_ratio, dim_ratio=dim_ratio,
+                                            do_windowing=do_windowing, multi_query=multi_query, conv_base=conv_base,
+                                            use_shift=use_shift, cpb_mlp_hidden=cpb_mlp_hidden, conv_groups_ratio=conv_groups_ratio),
+                                        )
+
+    def forward(self, x):
+
+        for attention_block in self.attention_blocks:
+            x = attention_block(x)
+
+        return x
+
+
+
+class Mlp(nn.Module):
+    """
+    Multi-Layer Perceptron (MLP) block
+    """
+
+    def __init__(self,
+                 in_features,
+                 hidden_features=None,
+                 out_features=None,
+                 act_layer=nn.GELU,
+                 use_swiglu=True,
+                 drop=0.):
+        """
+        Args:
+            in_features: input features dimension.
+            hidden_features: hidden features dimension.
+            out_features: output features dimension.
+            act_layer: activation function.
+            drop: dropout rate.
+        """
+
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features * (2 if use_swiglu else 1), bias=False)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=False)
+
+    def forward(self, x):
+        x_size = x.size()
+        x = x.view(-1, x_size[-1])
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.fc2(x)
+        x = x.view(x_size)
+        return x
+
+class Downsample(nn.Module):
+    """
+    Down-sampling block
+    Pixel Unshuffle is used for down-sampling, works great accuracy - wise but takes 10% more TRT time
+    """
+
+    def __init__(self,
+                 dim,
+                 shuffle = False,
+                 ):
+        """
+        Args:
+            dim: feature size dimension.
+            shuffle: idea with
+            keep_dim: bool argument for maintaining the resolution.
+        """
+
+        super().__init__()
+        dim_out = 2 * dim
+
+        if shuffle:
+            self.norm = lambda x: pixel_unshuffle(x, factor=2)
+            self.reduction = Conv2d_BN(dim*4, dim_out, 1, 1, 0, bias=False)
+            # pixel unshuffleging works well but doesnt provide any speedup
+        else:
+            # removed layer norm for better, in this formulation we are getting 10% better speed
+            # LayerNorm for high resolution inputs will be a pain as it pools over the entire spatial dimension
+            # therefore we remove it compared to the original implementation in FasterViT
+            self.norm = nn.Identity()
+            self.reduction = Conv2d_BN(dim, dim_out, 3, 2, 1, bias=False)
+
+
+    def forward(self, x):
+        x = self.norm(x)
+        x = self.reduction(x)
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """
+    Patch embedding block
+    Used to convert image into an initial set of feature maps with lower resolution
+    """
+
+    def __init__(self, in_chans=3, in_dim=64, dim=96, shuffle_down=False):
+        """
+        Args:
+            in_chans: number of input channels.
+            in_dim: intermediate feature size dimension to speed up stem.
+            dim: final stem channel number
+            shuffle_down: use PixelUnshuffle for down-sampling, effectively increases the receptive field
+        """
+
+        super().__init__()
+        # shuffle_down = False
+        if not shuffle_down:
+            self.proj = nn.Identity()
+            self.conv_down = nn.Sequential(
+                Conv2d_BN(in_chans, in_dim, 3, 2, 1, bias=False),
+                nn.ReLU(),
+                Conv2d_BN(in_dim, dim, 3, 2, 1, bias=False),
+                nn.ReLU()
+                )
+        else:
+            self.proj = lambda x: pixel_unshuffle(x, factor=4)
+            self.conv_down = nn.Sequential(Conv2d_BN(in_chans*16, dim, 3, 1, 1),
+                                           nn.ReLU(),
+                                           )
+
+    def forward(self, x):
+        x = self.proj(x)
+        x = self.conv_down(x)
+        return x
+
+
+
+class ConvBlock(nn.Module):
+    """
+    Convolutional block, used in first couple of stages
+    Experimented with plan resnet-18 like modules, they are the best in terms of throughput
+    Finally, YOLOv8 idea seem to work fine (resnet-18 like block with squeezed feature dimension, and feature concatendation at the end)
+    """
+    def __init__(self, dim,
+                 drop_path=0.,
+                 layer_scale=None,
+                 kernel_size=3,
+                 ):
+        super().__init__()
+
+        self.conv1 = Conv2d_BN(dim, dim, kernel_size=kernel_size, stride=1, padding=1)
+        self.act1 = nn.GELU()
+
+        self.conv2 = Conv2d_BN(dim, dim, kernel_size=kernel_size, stride=1, padding=1)
+
+        self.layer_scale = layer_scale
+        if layer_scale is not None and type(layer_scale) in [int, float]:
+            self.gamma = nn.Parameter(layer_scale * torch.ones(dim))
+            self.layer_scale = True
+        else:
+            self.layer_scale = False
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+    def forward(self, x):
+        input = x
+
+        x = self.conv1(x)
+        x = self.act1(x)
+        x = self.conv2(x)
+
+        if self.layer_scale:
+            x = x * self.gamma.view(1, -1, 1, 1)
+        x = input + self.drop_path(x)
+        return x
+
+
+class WindowAttention(nn.Module):
+    # Windowed Attention from SwinV2
+    # use a MLP trick to deal with various input image resolutions, then fold it to improve speed
+
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, resolution=0,
+                 seq_length=0, dim_out=None, multi_query=False, shift_size=0, cpb_mlp_hidden=512):
+        # taken from EdgeViT and tweaked with attention bias.
+        super().__init__()
+        if not dim_out: dim_out = dim
+        self.shift_size = shift_size
+        self.multi_query = multi_query
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.head_dim = dim // num_heads
+
+        self.dim_internal = dim
+
+        self.scale = qk_scale or head_dim ** -0.5
+        if not multi_query:
+            self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        else:
+            self.qkv = nn.Linear(dim, dim + 2*self.head_dim, bias=qkv_bias)
+
+        self.proj = nn.Linear(dim, dim_out, bias=False)
+        # attention positional bias
+        self.pos_emb_funct = PosEmbMLPSwinv2D(window_size=[resolution, resolution],
+                                              pretrained_window_size=[resolution, resolution],
+                                              num_heads=num_heads,
+                                              seq_length=seq_length,
+                                              cpb_mlp_hidden=cpb_mlp_hidden)
+
+        self.resolution = resolution
+
+    def forward(self, x, attn_mask = None):
+        B, N, C = x.shape
+
+        if not self.multi_query:
+            qkv = self.qkv(x).reshape(B, -1, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+            q, k, v = qkv[0], qkv[1], qkv[2]
+        else:
+            qkv = self.qkv(x)
+            (q, k, v) = qkv.split([self.dim_internal, self.head_dim, self.head_dim], dim=2)
+
+            q = q.reshape(B, -1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+            k = k.reshape(B, -1, 1, C // self.num_heads).permute(0, 2, 1, 3)
+            v = v.reshape(B, -1, 1, C // self.num_heads).permute(0, 2, 1, 3)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+
+        attn = self.pos_emb_funct(attn)
+
+        #add window shift
+        if attn_mask is not None:
+            nW = attn_mask.shape[0]
+            attn = attn.view(B // nW, nW, self.num_heads, N, N) + attn_mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+
+        attn = attn.softmax(dim=-1)
+        x = (attn @ v).transpose(1, 2).reshape(B, -1, C)
+        x = self.proj(x)
+        return x
+
+
+
+class ERADIOLayer(nn.Module):
+    """
+    E-RADIO Layer
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size,
+                 conv=False,
+                 downsample=True,
+                 mlp_ratio=4.,
+                 qkv_bias=False,
+                 qk_scale=None,
+                 norm_layer=nn.LayerNorm,
+                 drop_path=0.,
+                 layer_scale=None,
+                 layer_scale_conv=None,
+                 sr_dim_ratio=1,
+                 sr_ratio=1,
+                 multi_query=False,
+                 use_swiglu=True,
+                 yolo_arch=False,
+                 downsample_shuffle=False,
+                 conv_base=False,
+                 use_shift=False,
+                 cpb_mlp_hidden=512,
+                 conv_groups_ratio=0,
+                 verbose: bool = True,
+
+    ):
+        """
+        Args:
+            dim: feature size dimension.
+            depth: number of layers in each stage.
+            input_resolution: input image resolution.
+            window_size: window size in each stage.
+            downsample: bool argument for down-sampling.
+            mlp_ratio: MLP ratio.
+            num_heads: number of heads in each stage.
+            qkv_bias: bool argument for query, key, value learnable bias.
+            qk_scale: bool argument to scaling query, key.
+            drop: dropout rate.
+            attn_drop: attention dropout rate.
+            drop_path: drop path rate.
+            norm_layer: normalization layer.
+            layer_scale: layer scaling coefficient.
+            use_shift: SWIN like window shifting for half the window size for every alternating layer (considering multi-resolution)
+            conv_groups_ratio: group ratio for conv when no subsampling in multi-res attention
+        """
+
+        super().__init__()
+        self.conv = conv
+        self.yolo_arch=False
+        self.verbose = verbose
+        if conv:
+            if not yolo_arch:
+                self.blocks = nn.ModuleList([
+                    ConvBlock(dim=dim,
+                            drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                            layer_scale=layer_scale_conv)
+                    for i in range(depth)])
+                self.blocks = nn.Sequential(*self.blocks)
+            else:
+                self.blocks = C2f(dim,dim,n=depth,shortcut=True,e=0.5)
+                self.yolo_arch=True
+        else:
+            if not isinstance(window_size, list): window_size = [window_size]
+            self.window_size = window_size[0]
+            self.do_single_windowing = True
+            if not isinstance(sr_ratio, list): sr_ratio = [sr_ratio]
+            self.sr_ratio = sr_ratio
+            if any([sr!=1 for sr in sr_ratio]) or len(set(window_size))>1:
+                self.do_single_windowing = False
+                do_windowing = True
+            else:
+                self.do_single_windowing = True
+                do_windowing = False
+
+            #for v2_2
+            if conv_groups_ratio != -1:
+                self.do_single_windowing = False
+                do_windowing = True
+
+            self.blocks = nn.ModuleList()
+            for i in range(depth):
+                self.blocks.append(
+                    MultiResolutionAttention(window_size=window_size,
+                                             sr_ratio=sr_ratio,
+                                             dim=dim,
+                                             dim_ratio = sr_dim_ratio,
+                                             num_heads=num_heads,
+                                             norm_layer=norm_layer,
+                                             drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                             layer_scale=layer_scale,
+                                             qkv_bias=qkv_bias,
+                                             qk_scale=qk_scale,
+                                             use_swiglu=use_swiglu,
+                                             do_windowing=do_windowing,
+                                             multi_query=multi_query,
+                                             conv_base=conv_base,
+                                             cpb_mlp_hidden=cpb_mlp_hidden,
+                                             use_shift =0 if ((not use_shift) or ((i) % 2 == 0)) else True    ,
+                                             conv_groups_ratio=conv_groups_ratio,
+                    ))
+            self.blocks = nn.Sequential(*self.blocks)
+
+        self.transformer = not conv
+        self.downsample = None if not downsample else Downsample(dim=dim, shuffle=downsample_shuffle)
+
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+
+        # do padding for transforemr
+        interpolate = True
+        if self.transformer and interpolate:
+            # Windowed Attention will split feature map into windows with the size of window_size x window_size
+            # if the resolution is not divisible by window_size, we need to interpolate the feature map
+            # can be done via padding, but doing so after training hurts the model performance.
+            # interpolation affects the performance as well, but not as much as padding
+            if isinstance(self.window_size, list) or isinstance(self.window_size, tuple):
+                current_max_window_size = max(self.window_size)
+            else:
+                current_max_window_size = self.window_size
+
+            max_window_size = max([res_upsample*current_max_window_size for res_upsample in self.sr_ratio])
+            if H % max_window_size != 0 or W % max_window_size != 0:
+                new_h = int(np.ceil(H/max_window_size)*max_window_size)
+                new_w = int(np.ceil(W/max_window_size)*max_window_size)
+                x = F.interpolate(x, size=(new_h, new_w), mode='nearest')
+                if self.verbose:
+                    warnings.warn(f"Choosen window size is not optimal for given resolution. Interpolation of features maps will be done and it can affect the performance. Max window size is {max_window_size}, feature map size is {H}x{W}, interpolated feature map size is {new_h}x{new_w}.")
+
+
+        if self.transformer and self.do_single_windowing:
+            H, W = x.shape[2], x.shape[3]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        #run main blocks
+        x = self.blocks(x)
+
+        if self.transformer and self.do_single_windowing:
+            x = window_reverse(x, self.window_size, H, W, pad_hw)
+
+        if self.transformer and interpolate:
+            #lets keep original resolution, might be not ideal, but for the upsampling tower we need to keep the expected resolution.
+            x = F.interpolate(x, size=(H, W), mode='nearest')
+
+        if self.downsample is None:
+            return x, x
+
+        return self.downsample(x), x  # changing to output pre downsampled features
+
+
+class InterpolateLayer(nn.Module):
+    def __init__(self, size=None, scale_factor=None, mode='nearest'):
+        super(InterpolateLayer, self).__init__()
+        self.size = size
+        self.scale_factor = scale_factor
+        self.mode = mode
+
+    def forward(self, x):
+        return F.interpolate(x, size=self.size, scale_factor=self.scale_factor, mode=self.mode)
+
+
+class HiResNeck(nn.Module):
+    """
+    The block is used to output dense features from all stages
+    Otherwise, by default, only the last stage features are returned with E-RADIO
+    """
+    def __init__(self, dim, depths, neck_start_stage, full_features_head_dim, downsample_enabled):
+
+        '''
+        Hi Resolution neck to support output of high res features that are useful for dense tasks.
+        depths - total number of layers in the base model
+        neck_start_stage - when to start the neck, 0 - start from the first stage, 1 - start from the second stage etc.
+                            earlier layers result in higher resolution features at the cost of compute
+        full_features_head_dim - number of channels in the dense features head
+        '''
+        super().__init__()
+        # create feature projection layers for segmentation output
+        self.neck_features_proj = nn.ModuleList()
+        self.neck_start_stage = neck_start_stage
+        upsample_ratio = 1
+        for i in range(len(depths)):
+            level_n_features_output = int(dim * 2 ** i)
+
+            if self.neck_start_stage > i: continue
+
+            if (upsample_ratio > 1) or full_features_head_dim!=level_n_features_output:
+                feature_projection = nn.Sequential()
+                if False:
+                    feature_projection.add_module("norm",nn.BatchNorm2d(level_n_features_output)) #fast, but worse
+                    feature_projection.add_module("dconv", nn.ConvTranspose2d(level_n_features_output,
+                                                                            full_features_head_dim, kernel_size=upsample_ratio, stride=upsample_ratio))
+                else:
+                    # B, in_channels, H, W -> B, in_channels, H*upsample_ratio, W*upsample_ratio
+                    # print("upsample ratio", upsample_ratio, level_n_features_output, level_n_features_output)
+                    feature_projection.add_module("upsample", InterpolateLayer(scale_factor=upsample_ratio, mode='nearest'))
+                    feature_projection.add_module("conv1", nn.Conv2d(level_n_features_output, level_n_features_output, kernel_size=3, stride=1, padding=1, groups=level_n_features_output))
+                    feature_projection.add_module("norm",nn.BatchNorm2d(level_n_features_output))
+                    # B, in_channels, H*upsample_ratio, W*upsample_ratio -> B, full_features_head_dim, H*upsample_ratio, W*upsample_ratio
+                    feature_projection.add_module("conv2", nn.Conv2d(level_n_features_output, full_features_head_dim, kernel_size=1, stride=1, padding=0))
+            else:
+                feature_projection = nn.Sequential()
+
+            self.neck_features_proj.append(feature_projection)
+
+            if i>0 and downsample_enabled[i]:
+                upsample_ratio *= 2
+
+    def forward(self, x, il_level=-1, full_features=None):
+        if self.neck_start_stage > il_level:
+            return full_features
+
+        if full_features is None:
+            full_features = self.neck_features_proj[il_level - self.neck_start_stage](x)
+        else:
+            #upsample torch tensor x to match full_features size, and add to full_features
+            feature_projection = self.neck_features_proj[il_level - self.neck_start_stage](x)
+            if feature_projection.shape[2] != full_features.shape[2] or feature_projection.shape[3] != full_features.shape[3]:
+                feature_projection = torch.nn.functional.pad(feature_projection, ( 0, -feature_projection.shape[3] + full_features.shape[3], 0, -feature_projection.shape[2] + full_features.shape[2]))
+            full_features = full_features + feature_projection
+        return full_features
+
+class ERADIO(nn.Module):
+    """
+    Efficient RADIO
+    """
+
+    def __init__(self,
+                 dim,
+                 in_dim,
+                 depths,
+                 window_size,
+                 mlp_ratio,
+                 num_heads,
+                 drop_path_rate=0.2,
+                 in_chans=3,
+                 num_classes=1000,
+                 qkv_bias=False,
+                 qk_scale=None,
+                 layer_scale=None,
+                 layer_scale_conv=None,
+                 layer_norm_last=False,
+                 sr_ratio = [1, 1, 1, 1],
+                 max_depth = -1,
+                 conv_base=False,
+                 use_swiglu=False,
+                 multi_query=False,
+                 norm_layer=nn.LayerNorm,
+                 drop_uniform=False,
+                 yolo_arch=False,
+                 shuffle_down=False,
+                 downsample_shuffle=False,
+                 return_full_features=False,
+                 full_features_head_dim=128,
+                 neck_start_stage=1,
+                 use_neck=False,
+                 use_shift=False,
+                 cpb_mlp_hidden=512,
+                 conv_groups_ratio=0,
+                 verbose: bool = False,
+                 **kwargs):
+        """
+        Args:
+            dim: feature size dimension.
+            depths: number of layers in each stage.
+            window_size: window size in each stage.
+            mlp_ratio: MLP ratio.
+            num_heads: number of heads in each stage.
+            drop_path_rate: drop path rate.
+            in_chans: number of input channels.
+            num_classes: number of classes.
+            qkv_bias: bool argument for query, key, value learnable bias.
+            qk_scale: bool argument to scaling query, key.
+            drop_rate: dropout rate.
+            attn_drop_rate: attention dropout rate.
+            norm_layer: normalization layer.
+            layer_scale: layer scaling coefficient.
+            return_full_features: output dense features as well as logits
+            full_features_head_dim: number of channels in the dense features head
+            neck_start_stage: a stage id to start full feature neck. Model has 4 stages, indix starts with 0
+                                for 224 resolution, the output of the stage before downsample:
+                                stage 0: 56x56, stage 1: 28x28, stage 2: 14x14, stage 3: 7x7
+            use_neck: even for summarization embedding use neck
+            use_shift: SWIN like window shifting but without masking attention
+            conv_groups_ratio: will be used for conv blocks where there is no multires attention,
+                                if 0 then normal conv,
+                                if 1 then channels are independent,
+                                if -1 then no conv at all
+
+        """
+        super().__init__()
+
+        num_features = int(dim * 2 ** (len(depths) - 1))
+        self.num_classes = num_classes
+        self.patch_embed = PatchEmbed(in_chans=in_chans, in_dim=in_dim, dim=dim, shuffle_down=shuffle_down)
+        # set return_full_features true if we want to return full features from all stages
+        self.return_full_features = return_full_features
+        self.use_neck = use_neck
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
+        if drop_uniform:
+            dpr = [drop_path_rate for x in range(sum(depths))]
+
+        if not isinstance(max_depth, list): max_depth = [max_depth] * len(depths)
+
+        self.levels = nn.ModuleList()
+        for i in range(len(depths)):
+            conv = True if (i == 0 or i == 1) else False
+
+            level = ERADIOLayer(dim=int(dim * 2 ** i),
+                                   depth=depths[i],
+                                   num_heads=num_heads[i],
+                                   window_size=window_size[i],
+                                   mlp_ratio=mlp_ratio,
+                                   qkv_bias=qkv_bias,
+                                   qk_scale=qk_scale,
+                                   conv=conv,
+                                   drop_path=dpr[sum(depths[:i]):sum(depths[:i + 1])],
+                                   downsample=(i < len(depths) - 1),
+                                   layer_scale=layer_scale,
+                                   layer_scale_conv=layer_scale_conv,
+                                   sr_ratio=sr_ratio[i],
+                                   use_swiglu=use_swiglu,
+                                   multi_query=multi_query,
+                                   norm_layer=norm_layer,
+                                   yolo_arch=yolo_arch,
+                                   downsample_shuffle=downsample_shuffle,
+                                   conv_base=conv_base,
+                                   cpb_mlp_hidden=cpb_mlp_hidden,
+                                   use_shift=use_shift,
+                                   conv_groups_ratio=conv_groups_ratio,
+                                   verbose=verbose)
+
+            self.levels.append(level)
+
+        if self.return_full_features or self.use_neck:
+            #num_heads
+            downsample_enabled = [self.levels[i-1].downsample is not None for i in range(len(self.levels))]
+            self.high_res_neck = HiResNeck(dim, depths, neck_start_stage, full_features_head_dim, downsample_enabled)
+
+        self.switched_to_deploy = False
+
+        self.norm = LayerNorm2d(num_features) if layer_norm_last else nn.BatchNorm2d(num_features)
+        self.avgpool = nn.AdaptiveAvgPool2d(1)
+        self.head = nn.Linear(num_features, num_classes) if num_classes > 0 else nn.Identity()
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, LayerNorm2d):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.BatchNorm2d):
+            nn.init.ones_(m.weight)
+            nn.init.zeros_(m.bias)
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'rpb'}
+
+    def forward_features(self, x):
+        _, _, H, W = x.shape
+        if H % 32 != 0 or W % 32 != 0:
+            raise ValueError(f"E-RADIO requires input dimensions to be divisible by 32 but got H x W: {H} x {W}")
+        x = self.patch_embed(x)
+        full_features = None
+        for il, level in enumerate(self.levels):
+            x, pre_downsample_x = level(x)
+
+            if self.return_full_features or self.use_neck:
+                full_features = self.high_res_neck(pre_downsample_x, il, full_features)
+
+        # x = self.norm(full_features if (self.return_full_features or self.use_neck) else x)
+        x = self.norm(x) # new version for
+
+        if not self.return_full_features:
+            return x, None
+
+        return x, full_features
+
+    def forward(self, x):
+        x, full_features = self.forward_features(x)
+
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+
+        x = self.head(x)
+        if full_features is not None:
+            return x, full_features
+        return x
+
+    def switch_to_deploy(self):
+        '''
+        A method to perform model self-compression
+        merges BN into conv layers
+        converts MLP relative positional bias into precomputed buffers
+        '''
+        if not self.switched_to_deploy:
+            for level in [self.patch_embed, self.levels, self.head]:
+                for module in level.modules():
+                    if hasattr(module, 'switch_to_deploy'):
+                        module.switch_to_deploy()
+        self.switched_to_deploy = True
+
+
+    def change_window_size(self, new_window_size):
+        """
+        E-RADIO employs windowed attention, which may be sensitive to the choice of this parameter,
+        especially in cases of uneven partitioning of the feature maps.
+        E-RADIO allows for the adjustment of the window size after training,
+        making it adaptable to different input image resolutions.
+        The recommended values for window size based on input resolution are as follows:
+
+        Input Resolution | Window Size
+        224 | 7
+        256 | 8
+        386 | 12
+        512 | 16
+        Ideally, the window size should be a factor of the input resolution. In the third stage, we divide the resolution by 16, so the window size should be
+        img_res/16/2
+        for the third stage and img_res/32 for the last stage. While this can be applied in a brute-force manner, a better way is to do model.change_window_size.
+        Manual way to change resolution -> model.change_window_size(resolution)
+        """
+        window_size = new_window_size
+        print(f"Setting window size to {window_size}")
+        for module in self.modules():
+            if hasattr(module, "window_size"):
+                # check if tuple or a number
+                if isinstance(module.window_size, tuple):
+                    if module.window_size[0] != window_size:
+                        module.window_size = (window_size, window_size)
+                elif isinstance(module.window_size, list):
+                    if module.window_size[0] != window_size:
+                        module.window_size = [window_size, window_size]
+                else:
+                    module.window_size = window_size
+
+
+    def set_optimal_window_size(self, image_dim, max_window_size = 16):
+        """
+        Using hand picked window size for various resolutions.
+
+        E-RADIO employs windowed attention, which may be sensitive to the choice of this parameter,
+        especially in cases of uneven partitioning of the feature maps.
+        E-RADIO allows for the adjustment of the window size after training,
+        making it adaptable to different input image resolutions.
+        The recommended values for window size based on input resolution are as follows:
+
+        Input Resolution | Window Size
+        224 | 7
+        256 | 8
+        386 | 12
+        512 | 16
+        Ideally, the window size should be a factor of the input resolution. In the third stage, we divide the resolution by 16, so the window size should be
+        img_res/16/2
+        for the third stage and img_res/32 for the last stage. While this can be applied in a brute-force manner, a better way is to do model.change_window_size.
+        Manual way to change resolution -> model.change_window_size(resolution)
+
+        """
+        # import math
+
+        def divisorGenerator(n):
+            large_divisors = []
+            for i in range(1, int(math.sqrt(n) + 1)):
+                if n % i == 0:
+                    yield i
+                    if i*i != n:
+                        large_divisors.append(n / i)
+            for divisor in reversed(large_divisors):
+                yield divisor
+
+        if isinstance(image_dim, list) or isinstance(image_dim, tuple):
+            image_dim = min(image_dim)
+
+        # we do windowed attention in the 3rd stage for the first time, therefore //16,
+        # we do subsampled attention with downsample by 2 so need to get //32 actually
+        # ideally we should rewrite this to be dependent on the structure of the model like what if subsampled is removed etc
+        all_divisors = np.array(list(divisorGenerator(image_dim//32)))
+        new_window_size = int(min(all_divisors[all_divisors <= max_window_size][-1], max_window_size))
+
+        # for image_dim in [128, 224, 256, 384, 512, 768, 1024]:
+        #     all_divisors = np.array(list(divisorGenerator(image_dim//32)))
+        #     new_window_size = int(min(all_divisors[all_divisors <= max_window_size][-1], max_window_size))
+        #     print(f"Setting window size to {new_window_size} for image resolution {image_dim}")
+
+        self.change_window_size(new_window_size = new_window_size)
+
+
+@register_model
+def eradio_large_fullres_ws16(pretrained=False, **kwargs):
+    model = ERADIO(
+        depths=[3, 3, 5, 5],
+        num_heads=[2, 4, 8, 16],
+        window_size=[None, None, [16, 16], 16],
+        dim=192,
+        in_dim=64,
+        mlp_ratio=4,
+        drop_path_rate=0.0,
+        sr_ratio=[1, 1, [2, 1], 1],
+        use_swiglu=False,
+        yolo_arch=True,
+        shuffle_down=False,
+        conv_base=True,
+        use_neck=True,
+        full_features_head_dim=1536,
+        neck_start_stage=2,
+        **kwargs,
+    )
+    if pretrained:
+        model.load_state_dict(torch.load(pretrained)["state_dict"])
+    return model
+
+
+@register_model
+def eradio_xxxtiny(pretrained=False, **kwargs):  # ,
+    model = ERADIO(
+        depths=[1, 3, 4, 5],
+        num_heads=[2, 4, 8, 16],
+        window_size=[None, None, [16, 16], 16],
+        dim=32,
+        in_dim=32,
+        mlp_ratio=4,
+        drop_path_rate=0.0,
+        sr_ratio=[1, 1, [2, 1], 1],
+        use_swiglu=False,
+        yolo_arch=True,
+        shuffle_down=False,
+        conv_base=True,
+        use_neck=True,
+        full_features_head_dim=256,
+        neck_start_stage=2,
+        **kwargs,
+    )
+    if pretrained:
+        model.load_state_dict(torch.load(pretrained))
+    return model
+
+@register_model
+def eradio_xxxtiny_8x_ws12(pretrained=False, **kwargs):
+    model = ERADIO(depths=[1, 3, 4, 5],
+        num_heads=[2, 4, 8, 16],
+        window_size=[None, None, [12, 12], 12],
+        dim=32,
+        in_dim=32,
+        mlp_ratio=4,
+        drop_path_rate=0.0,
+        sr_ratio=[1, 1, [2, 1], 1],
+        use_swiglu=False,
+        downsample_shuffle=False,
+        yolo_arch=True,
+        shuffle_down=False,
+        cpb_mlp_hidden=64,
+        use_neck=True,
+        full_features_head_dim=256,
+        neck_start_stage=2,
+        conv_groups_ratio = 1,
+        **kwargs)
+    if pretrained:
+        model.load_state_dict(torch.load(pretrained)["state_dict"])
+    return model
+
+
+@register_model
+def eradio_xxxtiny_8x_ws16(pretrained=False, **kwargs):
+    model = ERADIO(depths=[1, 3, 4, 5],
+        num_heads=[2, 4, 8, 16],
+        window_size=[None, None, [16, 16], 16],
+        dim=32,
+        in_dim=32,
+        mlp_ratio=4,
+        drop_path_rate=0.0,
+        sr_ratio=[1, 1, [2, 1], 1],
+        use_swiglu=False,
+        downsample_shuffle=False,
+        yolo_arch=True,
+        shuffle_down=False,
+        cpb_mlp_hidden=64,
+        use_neck=True,
+        full_features_head_dim=256,
+        neck_start_stage=1,
+        conv_groups_ratio = 1,
+        **kwargs)
+    if pretrained:
+        model.load_state_dict(torch.load(pretrained)["state_dict"])
+    return model
+
+@register_model
+def eradio(pretrained=False, **kwargs):
+    return eradio_large_fullres_ws16(pretrained=pretrained, **kwargs)

extra_timm_models.py

@@ -0,0 +1,66 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from torch import nn
+
+from timm.models import register_model
+from timm.models.vision_transformer import VisionTransformer, _create_vision_transformer, Mlp
+
+
+@register_model
+def vit_tiny_patch14_224(pretrained=False, **kwargs) -> VisionTransformer:
+    """ ViT-Tiny (Vit-Ti/16)
+    """
+    model_args = dict(patch_size=14, embed_dim=192, depth=12, num_heads=3)
+    model = _create_vision_transformer('vit_tiny_patch14_224', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
+@register_model
+def vit_small_patch14_224(pretrained=False, **kwargs) -> VisionTransformer:
+    """ ViT-Small (ViT-S/16)
+    """
+    model_args = dict(patch_size=14, embed_dim=384, depth=12, num_heads=6)
+    model = _create_vision_transformer('vit_small_patch16_224', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
+@register_model
+def vit_base_patch14_224(pretrained=False, **kwargs) -> VisionTransformer:
+    """ ViT-Base (ViT-B/14) from original paper (https://arxiv.org/abs/2010.11929).
+    ImageNet-1k weights fine-tuned from in21k @ 224x224, source https://github.com/google-research/vision_transformer.
+    """
+    model_args = dict(patch_size=14, embed_dim=768, depth=12, num_heads=12)
+    model = _create_vision_transformer('vit_base_patch14_224', pretrained=pretrained, **dict(model_args, **kwargs))
+    return model
+
+
+@register_model
+def vit_huge_patch16_224(pretrained=False, **kwargs) -> VisionTransformer:
+    """ ViT-Huge model (ViT-H/16) from original paper (https://arxiv.org/abs/2010.11929).
+    """
+    model_args = dict(patch_size=16, embed_dim=1280, depth=32, num_heads=16)
+    if pretrained:
+        # There is no pretrained version of ViT-H/16, but we can adapt a ViT-H/14 for this purpose
+        model = _create_vision_transformer('vit_huge_patch14_clip_336', pretrained=True, **dict(model_args, pre_norm=True, **kwargs))
+    else:
+        model = _create_vision_transformer('vit_huge_patch16_224', pretrained=False, **dict(model_args, **kwargs))
+    return model
+
+
+@register_model
+def vit_huge_patch16_224_mlpnorm(pretrained=False, **kwargs) -> VisionTransformer:
+    """ ViT-Huge model (ViT-H/16) from original paper (https://arxiv.org/abs/2010.11929).
+    """
+    model = vit_huge_patch16_224(pretrained=pretrained, **kwargs)
+
+    for m in model.modules():
+        if isinstance(m, Mlp) and not isinstance(m.norm, nn.LayerNorm):
+            m.norm = nn.LayerNorm(m.fc1.out_features)
+
+    return model

hf_model.py

@@ -0,0 +1,174 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.  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.
+from collections import namedtuple
+from typing import Callable, Optional, List, Union
+
+from timm.models import VisionTransformer
+import torch
+from torch import nn
+from transformers import PretrainedConfig, PreTrainedModel
+
+
+from .common import RESOURCE_MAP, DEFAULT_VERSION
+
+# Import all required modules.
+from .adaptor_base import AdaptorBase, RadioOutput, AdaptorInput
+from .adaptor_generic import GenericAdaptor, AdaptorBase
+from .adaptor_mlp import create_mlp_from_state
+from .adaptor_registry import adaptor_registry
+from .cls_token import ClsToken
+from .enable_cpe_support import enable_cpe
+from .enable_spectral_reparam import configure_spectral_reparam_from_args
+from .eradio_model import eradio
+from .radio_model import create_model_from_args
+from .radio_model import RADIOModel as RADIOModelBase, Resolution
+from .input_conditioner import get_default_conditioner, InputConditioner
+from .open_clip_adaptor import OpenCLIP_RADIO
+from .vit_patch_generator import ViTPatchGenerator
+from .vitdet import apply_vitdet_arch, VitDetArgs
+
+# Register extra models
+from .extra_timm_models import *
+
+
+class RADIOConfig(PretrainedConfig):
+    """Pretrained Hugging Face configuration for RADIO models."""
+
+    def __init__(
+        self,
+        args: Optional[dict] = None,
+        version: Optional[str] = DEFAULT_VERSION,
+        patch_size: Optional[int] = None,
+        max_resolution: Optional[int] = None,
+        preferred_resolution: Optional[Resolution] = None,
+        adaptor_names: Union[str, List[str]] = None,
+        vitdet_window_size: Optional[int] = None,
+        **kwargs,
+    ):
+        self.args = args
+        for field in ["dtype", "amp_dtype"]:
+            if self.args is not None and field in self.args:
+                # Convert to a string in order to make it serializable.
+                # For example for torch.float32 we will store "float32",
+                # for "bfloat16" we will store "bfloat16".
+                self.args[field] = str(args[field]).split(".")[-1]
+        self.version = version
+        resource = RESOURCE_MAP[version]
+        self.patch_size = patch_size or resource.patch_size
+        self.max_resolution = max_resolution or resource.max_resolution
+        self.preferred_resolution = (
+            preferred_resolution or resource.preferred_resolution
+        )
+        self.adaptor_names = adaptor_names
+        self.vitdet_window_size = vitdet_window_size
+        super().__init__(**kwargs)
+
+
+class RADIOModel(PreTrainedModel):
+    """Pretrained Hugging Face model for RADIO.
+
+    This class inherits from PreTrainedModel, which provides
+    HuggingFace's functionality for loading and saving models.
+    """
+
+    config_class = RADIOConfig
+
+    def __init__(self, config: RADIOConfig):
+        super().__init__(config)
+
+        RADIOArgs = namedtuple("RADIOArgs", config.args.keys())
+        args = RADIOArgs(**config.args)
+        self.config = config
+
+        model = create_model_from_args(args)
+        input_conditioner: InputConditioner = get_default_conditioner()
+
+        dtype = getattr(args, "dtype", torch.float32)
+        if isinstance(dtype, str):
+            # Convert the dtype's string representation back to a dtype.
+            dtype = getattr(torch, dtype)
+        model.to(dtype=dtype)
+        input_conditioner.dtype = dtype
+
+        summary_idxs = torch.tensor(
+            [i for i, t in enumerate(args.teachers) if t.get("use_summary", True)],
+            dtype=torch.int64,
+        )
+
+        adaptor_names = config.adaptor_names
+        if adaptor_names is not None:
+            raise NotImplementedError(
+                f"Adaptors are not yet supported in Hugging Face models. Adaptor names: {adaptor_names}"
+            )
+
+        adaptors = dict()
+
+        self.radio_model = RADIOModelBase(
+            model,
+            input_conditioner,
+            summary_idxs=summary_idxs,
+            patch_size=config.patch_size,
+            max_resolution=config.max_resolution,
+            window_size=config.vitdet_window_size,
+            preferred_resolution=config.preferred_resolution,
+            adaptors=adaptors,
+        )
+
+    @property
+    def adaptors(self) -> nn.ModuleDict:
+        return self.radio_model.adaptors
+
+    @property
+    def model(self) -> VisionTransformer:
+        return self.radio_model.model
+
+    @property
+    def input_conditioner(self) -> InputConditioner:
+        return self.radio_model.input_conditioner
+
+    @property
+    def num_summary_tokens(self) -> int:
+        return self.radio_model.num_summary_tokens
+
+    @property
+    def patch_size(self) -> int:
+        return self.radio_model.patch_size
+
+    @property
+    def max_resolution(self) -> int:
+        return self.radio_model.max_resolution
+
+    @property
+    def preferred_resolution(self) -> Resolution:
+        return self.radio_model.preferred_resolution
+
+    @property
+    def window_size(self) -> int:
+        return self.radio_model.window_size
+
+    @property
+    def min_resolution_step(self) -> int:
+        return self.radio_model.min_resolution_step
+
+    def make_preprocessor_external(self) -> Callable[[torch.Tensor], torch.Tensor]:
+        return self.radio_model.make_preprocessor_external()
+
+    def get_nearest_supported_resolution(self, height: int, width: int) -> Resolution:
+        return self.radio_model.get_nearest_supported_resolution(height, width)
+
+    def switch_to_deploy(self):
+        return self.radio_model.switch_to_deploy()
+
+    def forward(self, x: torch.Tensor):
+        return self.radio_model.forward(x)

input_conditioner.py

@@ -0,0 +1,49 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from typing import Union, Tuple
+
+import torch
+from torch import nn
+
+
+norm_t = Union[Tuple[float, float, float], torch.Tensor]
+
+class InputConditioner(nn.Module):
+    def __init__(self,
+                 input_scale: float,
+                 norm_mean: norm_t,
+                 norm_std: norm_t,
+                 dtype: torch.dtype = None,
+    ):
+        super().__init__()
+
+        self.dtype = dtype
+
+        self.register_buffer("norm_mean", _to_tensor(norm_mean) / input_scale)
+        self.register_buffer("norm_std", _to_tensor(norm_std) / input_scale)
+
+    def forward(self, x: torch.Tensor):
+        y = (x - self.norm_mean) / self.norm_std
+        if self.dtype is not None:
+            y = y.to(self.dtype)
+        return y
+
+
+def get_default_conditioner():
+    from timm.data.constants import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
+
+    return InputConditioner(
+        input_scale=1.0,
+        norm_mean=OPENAI_CLIP_MEAN,
+        norm_std=OPENAI_CLIP_STD,
+    )
+
+
+def _to_tensor(v: norm_t):
+    return torch.as_tensor(v, dtype=torch.float32).view(-1, 1, 1)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9153ddb6c061bcf4ec299adcac9e83eda290bf56349138918714cc5e9b6417d
+size 1279558216

open_clip_adaptor.py

@@ -0,0 +1,41 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+from argparse import Namespace
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from .adaptor_registry import adaptor_registry, dict_t, state_t
+
+from .adaptor_generic import GenericAdaptor
+
+
+class OpenCLIP_RADIO(GenericAdaptor):
+    def __init__(self, main_config: Namespace, adaptor_config: dict_t, state: state_t):
+        super().__init__(main_config, adaptor_config, state)
+
+        import open_clip
+
+        self.oc_model = open_clip.create_model_from_pretrained(
+            model_name=adaptor_config['model'],
+            pretrained=adaptor_config['pretrained'],
+            return_transform=False,
+        )
+        # Unload these parameters
+        self.oc_model.visual = None
+
+        self.tokenizer = open_clip.get_tokenizer(model_name=adaptor_config['model'])
+
+    def encode_text(self, text, normalize: bool = False):
+        return self.oc_model.encode_text(text, normalize=normalize)
+
+
+@adaptor_registry.register_adaptor("open_clip")
+def create_open_clip_adaptor(main_config: Namespace, adaptor_config: dict_t, state: state_t):
+    return OpenCLIP_RADIO(main_config, adaptor_config, state)

radio_model.py

@@ -0,0 +1,257 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+from typing import Callable, Dict, List, NamedTuple, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from timm.models import create_model, VisionTransformer
+
+from .enable_cpe_support import enable_cpe
+from .input_conditioner import InputConditioner
+# Register extra models
+from . import extra_timm_models
+from .adaptor_base import AdaptorBase, RadioOutput, AdaptorInput
+from . import eradio_model
+from .enable_spectral_reparam import configure_spectral_reparam_from_args
+
+
+class Resolution(NamedTuple):
+    height: int
+    width: int
+
+
+class RADIOModel(nn.Module):
+    def __init__(
+        self,
+        model: nn.Module,
+        input_conditioner: InputConditioner,
+        patch_size: int,
+        max_resolution: int,
+        preferred_resolution: Resolution,
+        summary_idxs: Optional[torch.Tensor] = None,
+        window_size: int = None,
+        adaptors: Dict[str, AdaptorBase] = None,
+    ):
+        super().__init__()
+
+        self.model = model
+        self.input_conditioner = input_conditioner
+        if summary_idxs is not None:
+            self.register_buffer('summary_idxs', summary_idxs)
+        else:
+            self.summary_idxs = None
+
+        self._preferred_resolution = preferred_resolution
+        self._patch_size = patch_size
+        self._max_resolution = max_resolution
+        self._window_size = window_size
+
+        adaptors = adaptors or dict()
+        self.adaptors = nn.ModuleDict(adaptors)
+
+    @property
+    def num_summary_tokens(self) -> int:
+        patch_gen = getattr(self.model, "patch_generator", None)
+        if patch_gen is not None:
+            return patch_gen.num_skip
+        elif self.model.global_pool == 'avg':
+            return 0
+        return 1
+
+    @property
+    def patch_size(self) -> int:
+        if self._patch_size is not None:
+            return self._patch_size
+        patch_gen = getattr(self.model, "patch_generator", None)
+        if patch_gen is not None:
+            return patch_gen.patch_size
+        return None
+
+    @property
+    def max_resolution(self) -> int:
+        return self._max_resolution
+
+    @property
+    def preferred_resolution(self) -> Resolution:
+        return self._preferred_resolution
+
+    @property
+    def window_size(self) -> int:
+        return self._window_size
+
+    @property
+    def min_resolution_step(self) -> int:
+        res = self.patch_size
+        if self.window_size is not None:
+            res *= self.window_size
+        return res
+
+    def make_preprocessor_external(self) -> Callable[[torch.Tensor], torch.Tensor]:
+        ret = self.input_conditioner
+        self.input_conditioner = nn.Identity()
+        return ret
+
+    def get_nearest_supported_resolution(self, height: int, width: int) -> Resolution:
+        height = int(round(height / self.min_resolution_step) * self.min_resolution_step)
+        width = int(round(width / self.min_resolution_step) * self.min_resolution_step)
+
+        height = max(height, self.min_resolution_step)
+        width = max(width, self.min_resolution_step)
+
+        return Resolution(height=height, width=width)
+
+    def switch_to_deploy(self):
+        fn = getattr(self.model, 'switch_to_deploy', None)
+        if fn is not None:
+            fn()
+
+    def forward(self, x: torch.Tensor) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        res_step = self.min_resolution_step
+        if res_step is not None and (x.shape[-2] % res_step != 0 or x.shape[-1] % res_step != 0):
+            raise ValueError('The input resolution must be a multiple of `self.min_resolution_step`. '
+                             '`self.get_nearest_supported_resolution(<height>, <width>) is provided as a convenience API. '
+                             f'Input: {x.shape[-2:]}, Nearest: {self.get_nearest_supported_resolution(*x.shape[-2:])}')
+
+        x = self.input_conditioner(x)
+        y = self.model.forward_features(x)
+
+        if isinstance(self.model, VisionTransformer):
+            patch_gen = getattr(self.model, "patch_generator", None)
+            if patch_gen is not None:
+                all_summary = y[:, : patch_gen.num_cls_tokens]
+                if self.summary_idxs is not None:
+                    bb_summary = all_summary[:, self.summary_idxs]
+                else:
+                    bb_summary = all_summary
+                all_feat = y[:, patch_gen.num_skip :]
+            elif self.model.global_pool == "avg":
+                all_summary = y[:, self.model.num_prefix_tokens :].mean(dim=1)
+                bb_summary = all_summary
+                all_feat = y
+            else:
+                all_summary = y[:, 0]
+                bb_summary = all_summary
+                all_feat = y[:, 1:]
+        elif isinstance(self.model, eradio_model.ERADIO):
+            _, f = y
+            all_feat = f.flatten(2).transpose(1, 2)
+            all_summary = all_feat.mean(dim=1)
+            bb_summary = all_summary
+        elif isinstance(y, (list, tuple)):
+            all_summary, all_feat = y
+            bb_summary = all_summary
+        else:
+            raise ValueError("Unsupported model type")
+
+        all_feat = all_feat.float()
+        ret = RadioOutput(bb_summary.flatten(1), all_feat).to(torch.float32)
+        if self.adaptors:
+            ret = dict(backbone=ret)
+            for name, adaptor in self.adaptors.items():
+                if all_summary.ndim == 3:
+                    summary = all_summary[:, adaptor.head_idx]
+                else:
+                    summary = all_summary
+                ada_input = AdaptorInput(images=x, summary=summary.float(), features=all_feat)
+                v = adaptor(ada_input).to(torch.float32)
+                ret[name] = v
+
+        return ret
+
+    def forward_intermediates(
+            self,
+            x: torch.Tensor,
+            indices: Optional[Union[int, List[int], Tuple[int]]] = None,
+            return_prefix_tokens: bool = False,
+            norm: bool = False,
+            stop_early: bool = False,
+            output_fmt: str = 'NCHW',
+            intermediates_only: bool = False,
+            aggregation: Optional[str] = "sparse",
+    ) -> List[RadioOutput]:
+        """ Forward features that returns intermediates.
+        Args:
+            x: Input image tensor
+            indices: Take last n blocks if int, select matching indices if sequence
+            return_prefix_tokens: Return both prefix and spatial intermediate tokens
+            norm: Apply norm layer to all intermediates
+            stop_early: Stop iterating over blocks when last desired intermediate hit
+            output_fmt: Shape of intermediate feature outputs
+            intermediates_only: Only return intermediate features
+            aggregation: intermediate layer aggregation method (sparse or dense).
+                Dense accumulation is done by averaging the features in each group.
+        Returns:
+            List of RadioOutput objects.
+        """
+        outputs = self.model.forward_intermediates(
+            x,
+            indices=indices,
+            return_prefix_tokens=return_prefix_tokens,
+            norm=norm,
+            stop_early=stop_early,
+            output_fmt=output_fmt,
+            intermediates_only=intermediates_only,
+            aggregation=aggregation,
+        )
+        if return_prefix_tokens:
+            radio_outputs = [RadioOutput(summary, features) for (summary, features) in outputs]
+        else:
+            radio_outputs = [RadioOutput(None, features) for features in outputs]
+        return radio_outputs
+
+
+def create_model_from_args(args) -> nn.Module:
+    in_chans = 3
+    if args.in_chans is not None:
+        in_chans = args.in_chans
+    elif args.input_size is not None:
+        in_chans = args.input_size[0]
+
+    # Skip weight initialization unless it's explicitly requested.
+    weight_init = args.model_kwargs.pop("weight_init", "skip")
+
+    model = create_model(
+        args.model,
+        pretrained=args.pretrained,
+        in_chans=in_chans,
+        num_classes=args.num_classes,
+        drop_rate=args.drop,
+        drop_path_rate=args.drop_path,
+        drop_block_rate=args.drop_block,
+        global_pool=args.gp,
+        bn_momentum=args.bn_momentum,
+        bn_eps=args.bn_eps,
+        scriptable=args.torchscript,
+        checkpoint_path=args.initial_checkpoint,
+        weight_init=weight_init,
+        **args.model_kwargs,
+    )
+
+    if hasattr(model, 'norm') and not getattr(args, 'model_norm', False):
+        model.norm = nn.Identity()
+
+    model.head = nn.Identity()
+
+    assert (
+        not args.cls_token_per_teacher or args.cpe_max_size is not None
+    ), "CPE must be enabled for multiple CLS tokens!"
+
+    if args.cpe_max_size is not None:
+        uq_teachers = set(t['name'] for t in args.teachers)
+        enable_cpe(
+            model,
+            args.cpe_max_size,
+            num_cls_tokens=len(uq_teachers) if args.cls_token_per_teacher else 1,
+            register_multiple=args.register_multiple,
+        )
+
+    if args.spectral_reparam:
+        configure_spectral_reparam_from_args(model, args)
+
+    return model

vit_patch_generator.py

@@ -0,0 +1,299 @@
+# Copyright (c) 2023-2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import math
+from typing import Union, Tuple, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from einops import rearrange
+
+from .cls_token import ClsToken
+
+input_dim_t = Union[int, Tuple[int, int]]
+
+try:
+    # raise ImportError()
+    from indirect_grid_sample import indirect_grid_sample
+except ImportError:
+    indirect_grid_sample = None
+
+class ViTPatchGenerator(nn.Module):
+    def __init__(self,
+                 patch_size: int,
+                 embed_dim: int,
+                 input_dims: input_dim_t,
+                 abs_pos: bool = True,
+                 normalize_patches: bool = False,
+                 cls_token: bool = False,
+                 max_input_dims: Optional[input_dim_t] = None,
+                 pos_dropout: float = 0.0,
+                 return_pos_enc: bool = False,
+                 num_cls_tokens: int = 1,
+                 register_multiple: int = 0,
+                 device=None, dtype=None,
+    ):
+        super().__init__()
+
+        if isinstance(input_dims, int):
+            input_dims = (input_dims, input_dims)
+
+        if max_input_dims is None:
+            max_input_dims = input_dims
+        if isinstance(max_input_dims, int):
+            max_input_dims = (max_input_dims, max_input_dims)
+
+        max_input_dims = tuple(
+            int(math.ceil(d / patch_size) * patch_size)
+            for d in max_input_dims
+        )
+
+        self.cpe_mode = max_input_dims != input_dims
+        self.pos_dropout = pos_dropout
+        self.return_pos_enc = return_pos_enc
+
+        factory = dict(device=device, dtype=dtype)
+
+        self.patch_size = patch_size
+        self.abs_pos = abs_pos
+        self.embed_dim = embed_dim
+
+        self.num_rows = max_input_dims[0] // patch_size
+        self.num_cols = max_input_dims[1] // patch_size
+        self.input_dims = tuple(d // patch_size for d in input_dims)
+        self.num_patches = self.num_rows * self.num_cols
+        self.max_input_dims = max_input_dims
+
+        self.im_to_patches = Im2Patches(patch_size)
+        self.embedder = ViTPatchLinear(patch_size, embed_dim, **factory)
+
+        if abs_pos:
+            scale = embed_dim ** -0.5
+            self.pos_embed = nn.Parameter(torch.randn(1, self.num_patches, embed_dim, **factory) * scale)
+
+        self.cls_token = ClsToken(
+            embed_dim,
+            num_tokens=num_cls_tokens,
+            enabled=cls_token,
+            register_multiple=register_multiple,
+        )
+
+        self.patch_normalizer = nn.LayerNorm(embed_dim) if normalize_patches else nn.Identity()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        patches = self.embed_patches(x)
+        patches, pos_enc = self.apply_pos_enc(patches, input_size=x.shape[2:])
+        patches = self.cls_token(patches)
+        patches = self.patch_normalizer(patches)
+        if self.return_pos_enc:
+            return patches, pos_enc
+        return patches
+
+    @property
+    def apply_cls_token(self):
+        return self.cls_token.enabled
+
+    @property
+    def num_cls_tokens(self):
+        return self.cls_token.num_tokens
+
+    @property
+    def num_registers(self):
+        return self.cls_token.num_registers
+
+    @property
+    def num_skip(self):
+        return self.num_cls_tokens + self.num_registers
+
+    def no_weight_decay(self):
+        return [
+            'pos_embed',
+        ]
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+        if self.abs_pos:
+            self._load_embed(state_dict[f'{prefix}pos_embed'], self.pos_embed)
+
+    def _load_embed(self, src_embed: torch.Tensor, targ_embed: nn.Parameter):
+        if src_embed.shape != targ_embed.shape:
+            src_size = int(math.sqrt(src_embed.shape[1]))
+
+            assert src_size ** 2 == src_embed.shape[1], 'Unable to interpolate non-square embedding'
+
+            src_embed = rearrange(src_embed, 'b (h w) c -> b c h w', h=src_size, w=src_size)
+            src_embed = F.interpolate(src_embed, size=(self.num_rows, self.num_cols), mode='bicubic', align_corners=True, antialias=False)
+            src_embed = rearrange(src_embed, 'b c h w -> b (h w) c')
+        targ_embed.data.copy_(src_embed)
+
+    def _load_projection(self, src_proj_weight: torch.Tensor, targ_proj_weight: torch.Tensor):
+        if src_proj_weight.shape != targ_proj_weight.shape:
+            src_patch_size = int(math.sqrt(src_proj_weight.shape[1] // 3))
+
+            assert (src_patch_size ** 2) * 3 == src_proj_weight.shape[1], 'Unable to interpolate non-square patch size'
+
+            src_proj_weight = rearrange(src_proj_weight, 'b (c h w) -> b c h w', c=3, h=src_patch_size, w=src_patch_size)
+            src_proj_weight = F.interpolate(src_proj_weight, size=(self.patch_size, self.patch_size), mode='bicubic', align_corners=True, antialias=False)
+            src_proj_weight = rearrange(src_proj_weight, 'b c h w -> b (c h w)')
+        targ_proj_weight.data.copy_(src_proj_weight)
+
+    def embed_patches(self, x: torch.Tensor) -> torch.Tensor:
+        patches = self.im_to_patches(x)
+        patches = self.embedder(patches)
+        return patches
+
+    def apply_pos_enc(self,
+                      patches: torch.Tensor,
+                      patch_idxs: Optional[torch.Tensor] = None,
+                      input_size: Optional[Tuple[int, int]] = None,
+    ) -> torch.Tensor:
+        if not self.abs_pos:
+            return patches
+
+        pos_enc = self.get_pos_enc(patches.shape[0], patch_idxs, input_size)
+
+        if self.training and self.pos_dropout > 0:
+            keeps = torch.rand(patches.shape[0], 1, 1, dtype=pos_enc.dtype, device=pos_enc.device) > self.pos_dropout
+            pos_enc_drop = torch.where(keeps, pos_enc, 0)
+        else:
+            pos_enc_drop = pos_enc
+
+        return patches + pos_enc_drop, pos_enc
+
+    def get_pos_enc(self,
+                    batch_size: int,
+                    patch_idxs: Optional[torch.Tensor] = None,
+                    input_size: Optional[Tuple[int, int]] = None,
+    ) -> torch.Tensor:
+        if input_size is None:
+            input_dims = self.input_dims
+        else:
+            input_dims = tuple(d // self.patch_size for d in input_size)
+
+        pos_embed = self._get_pos_embeddings(batch_size, input_dims)
+
+        if patch_idxs is None:
+            return pos_embed
+
+        exp_patch_idxs = patch_idxs.unsqueeze(-1).expand(-1, -1, pos_embed.shape[-1])
+
+        pos_embed = torch.gather(pos_embed.expand(patch_idxs.shape[0], -1, -1), dim=1, index=exp_patch_idxs)
+        return pos_embed
+
+
+    def _get_pos_embeddings(self, batch_size: int, input_dims: Tuple[int, int]):
+        if (self.num_rows, self.num_cols) == input_dims:
+            return self.pos_embed
+
+        pos_embed = self.pos_embed.reshape(1, self.num_rows, self.num_cols, -1).permute(0, 3, 1, 2)
+
+        def window_select(pos_embed):
+            if input_dims[0] < pos_embed.shape[-2]:
+                pos_embed = pos_embed[..., :input_dims[0], :]
+            if input_dims[1] < pos_embed.shape[-1]:
+                pos_embed = pos_embed[..., :, :input_dims[1]]
+            return pos_embed
+
+        if self.cpe_mode:
+            if self.training:
+                min_scale = math.sqrt(0.1)
+                scale = torch.rand(batch_size, 1, 1, device=pos_embed.device) * (1 - min_scale) + min_scale
+                aspect_min = math.log(3 / 4)
+                aspect_max = -aspect_min
+                aspect = torch.exp(torch.rand(batch_size, 1, 1, device=pos_embed.device) * (aspect_max - aspect_min) + aspect_min)
+
+                scale_x = scale * aspect
+                scale_y = scale * (1 / aspect)
+                scale_xy = torch.stack([scale_x, scale_y], dim=-1).clamp_(0, 1)
+
+                pos_xy = torch.rand(batch_size, 1, 1, 2, device=pos_embed.device) * (1 - scale_xy)
+
+                lin_x = torch.linspace(0, 1, steps=input_dims[1], device=pos_embed.device)[None, None].expand(batch_size, input_dims[0], -1)
+                lin_y = torch.linspace(0, 1, steps=input_dims[0], device=pos_embed.device)[None, :, None].expand(batch_size, -1, input_dims[1])
+
+                lin_xy = torch.stack([lin_x, lin_y], dim=-1)
+
+                grid_xy = lin_xy * scale_xy + pos_xy
+
+                # Convert to [-1, 1] range
+                grid_xy.mul_(2).sub_(1)
+
+                pos_embed = F.grid_sample(
+                    pos_embed.float().expand(batch_size, -1, -1, -1),
+                    grid=grid_xy,
+                    mode='bilinear',
+                    padding_mode='zeros',
+                    align_corners=True,
+                ).to(pos_embed.dtype)
+            else:
+                # i_rows, i_cols = input_dims
+                # p_rows, p_cols = pos_embed.shape[2:]
+                # if i_rows <= p_rows and i_cols <= p_cols:
+                #     left = (p_cols - i_cols) // 2
+                #     top = (p_rows - i_rows) // 2
+                #     pos_embed = pos_embed[..., top:top+i_rows, left:left+i_cols]
+                # else:
+                max_dim = max(input_dims)
+                pos_embed = F.interpolate(pos_embed.float(), size=(max_dim, max_dim), align_corners=True, mode='bilinear').to(pos_embed.dtype)
+
+                pos_embed = window_select(pos_embed)
+        else:
+            pos_embed = window_select(pos_embed)
+
+        if pos_embed.shape[-2:] != input_dims:
+            pos_embed = F.interpolate(pos_embed.float(), size=input_dims, align_corners=True, mode='bilinear').to(pos_embed.dtype)
+
+        pos_embed = pos_embed.flatten(2).permute(0, 2, 1)
+
+        return pos_embed
+
+
+class Im2Patches(nn.Module):
+    def __init__(self, patch_size: int):
+        super().__init__()
+        self.patch_size = patch_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.patch_size == 1:
+            patches = x.flatten(2)
+            patches = patches.permute(0, 2, 1)
+            return patches
+
+        py = x.shape[-2] // self.patch_size
+        px = x.shape[-1] // self.patch_size
+        patches = rearrange(x, 'b c (py yy) (px xx) -> b (py px) (c yy xx)',
+                            py=py, yy=self.patch_size,
+                            px=px, xx=self.patch_size,
+        )
+        return patches
+
+
+class ViTPatchLinear(nn.Linear):
+    def __init__(self, patch_size: int, embed_dim: int, **factory):
+        super().__init__(
+            3 * (patch_size ** 2),
+            embed_dim,
+            bias=False,
+            **factory
+        )
+        self.patch_size = patch_size
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+        if self.bias is not None:
+            self.bias.data.copy_(state_dict[f'{prefix}bias'])
+
+        chk_weight = state_dict[f'{prefix}weight']
+        if chk_weight.shape != self.weight.shape:
+            src_patch_size = int(math.sqrt(chk_weight.shape[1] // 3))
+
+            assert (src_patch_size ** 2) * 3 == chk_weight.shape[1], 'Unable to interpolate non-square patch size'
+
+            chk_weight = rearrange(chk_weight, 'b (c h w) -> b c h w', c=3, h=src_patch_size, w=src_patch_size)
+            chk_weight = F.interpolate(chk_weight, size=(self.patch_size, self.patch_size), mode='bicubic', align_corners=True, antialias=False)
+            chk_weight = rearrange(chk_weight, 'b c h w -> b (c h w)')
+        self.weight.data.copy_(chk_weight)

vitdet.py

@@ -0,0 +1,181 @@
+from collections import defaultdict
+from contextlib import contextmanager
+from logging import getLogger
+import math
+import sys
+from typing import List, Union, Iterable
+
+import numpy as np
+import torch
+from torch import nn
+
+from timm.models import VisionTransformer
+from einops import rearrange
+
+DEFAULT_NUM_WINDOWED = 5
+DEFAULT_NUM_GLOBAL = 4
+
+
+class VitDetArgs:
+    def __init__(self,
+                 window_size: int,
+                 num_summary_tokens: int,
+                 num_windowed: int = None,
+                 num_global: int = None,
+    ):
+        self.window_size = window_size
+        self.num_summary_tokens = num_summary_tokens
+        self.num_windowed = num_windowed
+        self.num_global = num_global
+
+
+def apply_vitdet_arch(model: VisionTransformer, args: VitDetArgs):
+    if isinstance(model, VisionTransformer):
+        patch_embed = getattr(model, 'patch_generator', model.patch_embed)
+
+        return ViTDetHook(patch_embed, model.blocks, args)
+    else:
+        print(f'Warning: Unable to apply VitDet aug!', file=sys.stderr)
+
+
+class ViTDetHook:
+    def __init__(self,
+                 embedder: nn.Module,
+                 blocks: nn.Sequential,
+                 args: VitDetArgs,
+    ):
+        self.blocks = blocks
+        self.num_summary_tokens = args.num_summary_tokens
+        self.window_size = args.window_size
+
+        self._input_resolution = None
+        self._num_windows = None
+        self._cls_patch = None
+        self._order_cache = dict()
+
+        embedder.register_forward_pre_hook(self._enter_model)
+
+        # This will decide if we window-fy the patches
+        # and enable vit-det for this iteration, and if so,
+        # rearrange the patches for efficient mode switching
+        blocks.register_forward_pre_hook(self._enter_blocks)
+
+        is_global = True
+        if args.num_windowed is not None:
+            period = args.num_windowed + 1
+        else:
+            num_global = args.num_global or DEFAULT_NUM_GLOBAL
+            period = max(len(blocks) // num_global, 1)
+
+        for i, layer in enumerate(blocks[:-1]):
+            ctr = i % period
+            if ctr == 0:
+                layer.register_forward_pre_hook(self._to_windows)
+                is_global = False
+            elif ctr == period - 1:
+                layer.register_forward_pre_hook(self._to_global)
+                is_global = True
+
+        # Always ensure the final layer is a global layer
+        if not is_global:
+            blocks[-1].register_forward_pre_hook(self._to_global)
+
+        blocks.register_forward_hook(self._exit_model)
+
+    def _enter_model(self, _, input: List[torch.Tensor]):
+        self._input_resolution = input[0].shape[-2:]
+
+    def _enter_blocks(self, _, input: List[torch.Tensor]):
+        # print(f'{get_rank()} - ViTDet Window Size: {self._window_size}', file=sys.stderr)
+
+        patches = input[0]
+        patches = self._rearrange_patches(patches)
+
+        return (patches,) + input[1:]
+
+    def _to_windows(self, _, input: List[torch.Tensor]):
+        patches = input[0]
+
+        if self.num_summary_tokens:
+            self._cls_patch = patches[:, :self.num_summary_tokens]
+            patches = patches[:, self.num_summary_tokens:]
+
+        patches = rearrange(
+            patches, 'b (p t) c -> (b p) t c',
+            p=self._num_windows, t=self.window_size ** 2,
+        )
+
+        return (patches,) + input[1:]
+
+    def _to_global(self, _, input: List[torch.Tensor]):
+        patches = input[0]
+
+        patches = rearrange(
+            patches, '(b p) t c -> b (p t) c',
+            p=self._num_windows, t=self.window_size ** 2,
+            b=patches.shape[0] // self._num_windows,
+        )
+
+        if self.num_summary_tokens:
+            patches = torch.cat([
+                self._cls_patch,
+                patches,
+            ], dim=1)
+
+        return (patches,) + input[1:]
+
+    def _exit_model(self, _, inputs: List[torch.Tensor], patches: torch.Tensor):
+        # Return patches to their original order
+        patch_order = self._order_cache[self._input_resolution][0]
+        patch_order = patch_order.reshape(1, -1, 1).expand_as(patches)
+
+        ret_patches = torch.empty_like(patches)
+        ret_patches = torch.scatter(
+            ret_patches,
+            dim=1,
+            index=patch_order,
+            src=patches,
+        )
+
+        return ret_patches
+
+    def _rearrange_patches(self, patches: torch.Tensor):
+        # We rearrange the patches so that we can efficiently
+        # switch between windowed and global mode by just
+        # reshaping the tensor
+
+        patch_order, self._num_windows = self._order_cache.get(self._input_resolution, (None, None))
+        if patch_order is None:
+            num_feat_patches = patches.shape[1] - self.num_summary_tokens
+            num_pixels = self._input_resolution[0] * self._input_resolution[1]
+
+            patch_size = int(round(math.sqrt(num_pixels / num_feat_patches)))
+            rows = self._input_resolution[-2] // patch_size
+            cols = self._input_resolution[-1] // patch_size
+
+            w_rows = rows // self.window_size
+            w_cols = cols // self.window_size
+
+            patch_order = torch.arange(0, num_feat_patches, device=patches.device)
+
+            patch_order = rearrange(
+                patch_order, '(wy py wx px) -> (wy wx py px)',
+                wy=w_rows, wx=w_cols,
+                py=self.window_size, px=self.window_size,
+            )
+
+            if self.num_summary_tokens:
+                patch_order = torch.cat([
+                    torch.arange(self.num_summary_tokens, dtype=patch_order.dtype, device=patch_order.device),
+                    patch_order + self.num_summary_tokens,
+                ])
+
+            self._num_windows = w_rows * w_cols
+            self._order_cache[self._input_resolution] = (
+                patch_order,
+                self._num_windows,
+            )
+
+        patch_order = patch_order.reshape(1, -1, 1).expand_as(patches)
+        patches = torch.gather(patches, dim=1, index=patch_order)
+        return patches