jadechoghari
commited on
Commit
•
1822fe2
1
Parent(s):
f427d24
Rename modeling.py to pipeline_mar.py
Browse filesdiffusers lib is better than transformers for this model
- modeling.py +0 -183
- pipeline_mar.py +83 -0
modeling.py
DELETED
@@ -1,183 +0,0 @@
|
|
1 |
-
from transformers import PretrainedConfig
|
2 |
-
import torch.nn as nn
|
3 |
-
from transformers import PreTrainedModel
|
4 |
-
import torch
|
5 |
-
from huggingface_hub import hf_hub_download
|
6 |
-
from safetensors.torch import save_file, load_file
|
7 |
-
import os
|
8 |
-
from timm.models.vision_transformer import Block
|
9 |
-
from . import mar
|
10 |
-
from .vae import AutoencoderKL
|
11 |
-
from .mar import MAR
|
12 |
-
import numpy as np
|
13 |
-
|
14 |
-
class MARConfig(PretrainedConfig):
|
15 |
-
model_type = "mar"
|
16 |
-
|
17 |
-
def __init__(self,
|
18 |
-
img_size=256,
|
19 |
-
vae_stride=16,
|
20 |
-
patch_size=1,
|
21 |
-
encoder_embed_dim=1024,
|
22 |
-
encoder_depth=16,
|
23 |
-
encoder_num_heads=16,
|
24 |
-
decoder_embed_dim=1024,
|
25 |
-
decoder_depth=16,
|
26 |
-
decoder_num_heads=16,
|
27 |
-
mlp_ratio=4.,
|
28 |
-
norm_layer="LayerNorm",
|
29 |
-
vae_embed_dim=16,
|
30 |
-
mask_ratio_min=0.7,
|
31 |
-
label_drop_prob=0.1,
|
32 |
-
class_num=1000,
|
33 |
-
attn_dropout=0.1,
|
34 |
-
proj_dropout=0.1,
|
35 |
-
buffer_size=64,
|
36 |
-
diffloss_d=3,
|
37 |
-
diffloss_w=1024,
|
38 |
-
num_sampling_steps='100',
|
39 |
-
diffusion_batch_mul=4,
|
40 |
-
grad_checkpointing=False,
|
41 |
-
**kwargs):
|
42 |
-
super().__init__(**kwargs)
|
43 |
-
|
44 |
-
# store parameters in the config
|
45 |
-
self.img_size = img_size
|
46 |
-
self.vae_stride = vae_stride
|
47 |
-
self.patch_size = patch_size
|
48 |
-
self.encoder_embed_dim = encoder_embed_dim
|
49 |
-
self.encoder_depth = encoder_depth
|
50 |
-
self.encoder_num_heads = encoder_num_heads
|
51 |
-
self.decoder_embed_dim = decoder_embed_dim
|
52 |
-
self.decoder_depth = decoder_depth
|
53 |
-
self.decoder_num_heads = decoder_num_heads
|
54 |
-
self.mlp_ratio = mlp_ratio
|
55 |
-
self.norm_layer = norm_layer
|
56 |
-
self.vae_embed_dim = vae_embed_dim
|
57 |
-
self.mask_ratio_min = mask_ratio_min
|
58 |
-
self.label_drop_prob = label_drop_prob
|
59 |
-
self.class_num = class_num
|
60 |
-
self.attn_dropout = attn_dropout
|
61 |
-
self.proj_dropout = proj_dropout
|
62 |
-
self.buffer_size = buffer_size
|
63 |
-
self.diffloss_d = diffloss_d
|
64 |
-
self.diffloss_w = diffloss_w
|
65 |
-
self.num_sampling_steps = num_sampling_steps
|
66 |
-
self.diffusion_batch_mul = diffusion_batch_mul
|
67 |
-
self.grad_checkpointing = grad_checkpointing
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
class MARModel(PreTrainedModel):
|
72 |
-
# links to MARConfig class
|
73 |
-
config_class = MARConfig
|
74 |
-
|
75 |
-
def __init__(self, config):
|
76 |
-
super().__init__(config)
|
77 |
-
self.config = config
|
78 |
-
|
79 |
-
# convert norm_layer from string to class
|
80 |
-
norm_layer = getattr(nn, config.norm_layer)
|
81 |
-
|
82 |
-
# init the mar model using the parameters from config
|
83 |
-
self.model = MAR(
|
84 |
-
img_size=config.img_size,
|
85 |
-
vae_stride=config.vae_stride,
|
86 |
-
patch_size=config.patch_size,
|
87 |
-
encoder_embed_dim=config.encoder_embed_dim,
|
88 |
-
encoder_depth=config.encoder_depth,
|
89 |
-
encoder_num_heads=config.encoder_num_heads,
|
90 |
-
decoder_embed_dim=config.decoder_embed_dim,
|
91 |
-
decoder_depth=config.decoder_depth,
|
92 |
-
decoder_num_heads=config.decoder_num_heads,
|
93 |
-
mlp_ratio=config.mlp_ratio,
|
94 |
-
norm_layer=norm_layer, # use the actual class for the layer
|
95 |
-
vae_embed_dim=config.vae_embed_dim,
|
96 |
-
mask_ratio_min=config.mask_ratio_min,
|
97 |
-
label_drop_prob=config.label_drop_prob,
|
98 |
-
class_num=config.class_num,
|
99 |
-
attn_dropout=config.attn_dropout,
|
100 |
-
proj_dropout=config.proj_dropout,
|
101 |
-
buffer_size=config.buffer_size,
|
102 |
-
diffloss_d=config.diffloss_d,
|
103 |
-
diffloss_w=config.diffloss_w,
|
104 |
-
num_sampling_steps=config.num_sampling_steps,
|
105 |
-
diffusion_batch_mul=config.diffusion_batch_mul,
|
106 |
-
grad_checkpointing=config.grad_checkpointing,
|
107 |
-
)
|
108 |
-
|
109 |
-
def forward_train(self, imgs, labels):
|
110 |
-
# calls the forward method from the mar class - passing imgs & labels
|
111 |
-
return self.model(imgs, labels)
|
112 |
-
|
113 |
-
def forward(self, num_iter=64, cfg=1.0, cfg_schedule="linear", labels=None, temperature=1.0, progress=False):
|
114 |
-
# call the sample_tokens method from the MAR class
|
115 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
116 |
-
checkpoint_path = hf_hub_download(
|
117 |
-
repo_id=pretrained_model_name_or_path,
|
118 |
-
filename=f"kl16.safetensors"
|
119 |
-
)
|
120 |
-
vae = AutoencoderKL(embed_dim=16, ch_mult=(1, 1, 2, 2, 4), ckpt_path=checkpoint_path)
|
121 |
-
vae = vae.to(device).eval()
|
122 |
-
# can customize more from the user
|
123 |
-
seed = 0
|
124 |
-
torch.manual_seed(seed)
|
125 |
-
np.random.seed(seed)
|
126 |
-
num_ar_steps = 64
|
127 |
-
cfg_scale = 4
|
128 |
-
cfg_schedule = "constant"
|
129 |
-
temperature = 1.0
|
130 |
-
# TODO: this should be defined by the user
|
131 |
-
class_labels = 207, 360, 388, 113, 355, 980, 323, 979 #@param {type:"raw"}
|
132 |
-
samples_per_row = 4
|
133 |
-
|
134 |
-
with torch.cuda.amp.autocast():
|
135 |
-
sampled_tokens = self.model.sample_tokens(
|
136 |
-
bsz=len(class_labels), num_iter=num_ar_steps,
|
137 |
-
cfg=cfg_scale, cfg_schedule=cfg_schedule,
|
138 |
-
labels=torch.Tensor(class_labels).long().to(device),
|
139 |
-
temperature=temperature, progress=True)
|
140 |
-
sampled_images = vae.decode(sampled_tokens / 0.2325)
|
141 |
-
return sampled_images
|
142 |
-
|
143 |
-
@classmethod
|
144 |
-
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
|
145 |
-
# config = MARConfig.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
|
146 |
-
# model = cls(config)
|
147 |
-
buffer_size = kwargs.get('buffer_size', 64)
|
148 |
-
diffloss_d = kwargs.get('diffloss_d', 3)
|
149 |
-
diffloss_w = kwargs.get('diffloss_w', 1024)
|
150 |
-
num_sampling_steps_diffloss = kwargs.get('num_sampling_steps', 100)
|
151 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
152 |
-
model_type = "mar_base"
|
153 |
-
model_architecture = mar.__dict__[model_type](
|
154 |
-
buffer_size=buffer_size,
|
155 |
-
diffloss_d=diffloss_d,
|
156 |
-
diffloss_w=diffloss_w,
|
157 |
-
num_sampling_steps=str(num_sampling_steps_diffloss)
|
158 |
-
).to(device)
|
159 |
-
checkpoint_path = hf_hub_download(
|
160 |
-
repo_id=pretrained_model_name_or_path,
|
161 |
-
filename=f"checkpoint-last.pth"
|
162 |
-
)
|
163 |
-
|
164 |
-
state_dict = torch.load(checkpoint_path, map_location=device)["model_ema"]
|
165 |
-
|
166 |
-
model_architecture.load_state_dict(state_dict, strict=False)
|
167 |
-
|
168 |
-
# update this so the model works on the forward call
|
169 |
-
model = model_architecture
|
170 |
-
model.eval()
|
171 |
-
|
172 |
-
return model
|
173 |
-
|
174 |
-
|
175 |
-
def save_pretrained(self, save_directory):
|
176 |
-
# we will save to safetensors
|
177 |
-
os.makedirs(save_directory, exist_ok=True)
|
178 |
-
state_dict = self.model.state_dict()
|
179 |
-
safetensors_path = os.path.join(save_directory, "pytorch_model.safetensors")
|
180 |
-
save_file(state_dict, safetensors_path)
|
181 |
-
|
182 |
-
# save the configuration as usual
|
183 |
-
self.config.save_pretrained(save_directory)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
pipeline_mar.py
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from diffusers import DiffusionPipeline
|
2 |
+
import torch
|
3 |
+
import numpy as np
|
4 |
+
from huggingface_hub import hf_hub_download
|
5 |
+
from safetensors.torch import load_file
|
6 |
+
import os
|
7 |
+
from mar.vae import AutoencoderKL
|
8 |
+
from mar import mar
|
9 |
+
|
10 |
+
# inheriting from DiffusionPipeline for HF
|
11 |
+
class MARModel(DiffusionPipeline):
|
12 |
+
|
13 |
+
def __init__(self):
|
14 |
+
super().__init__()
|
15 |
+
|
16 |
+
@torch.no_grad()
|
17 |
+
def _call(self, *args, **kwargs):
|
18 |
+
"""
|
19 |
+
This method downloads the model and VAE components,
|
20 |
+
then executes the forward pass based on the user's input.
|
21 |
+
"""
|
22 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
# init the mar model architecture
|
27 |
+
buffer_size = kwargs.get("buffer_size", 64)
|
28 |
+
diffloss_d = kwargs.get("diffloss_d", 3)
|
29 |
+
diffloss_w = kwargs.get("diffloss_w", 1024)
|
30 |
+
num_sampling_steps = kwargs.get("num_sampling_steps", 100)
|
31 |
+
model_type = kwargs.get("model_type", "mar_base")
|
32 |
+
|
33 |
+
|
34 |
+
self.model = mar.__dict__[model_type](
|
35 |
+
buffer_size=buffer_size,
|
36 |
+
diffloss_d=diffloss_d,
|
37 |
+
diffloss_w=diffloss_w,
|
38 |
+
num_sampling_steps=str(num_sampling_steps)
|
39 |
+
).to(device)
|
40 |
+
# download and load the model weights (.safetensors or .pth)
|
41 |
+
model_checkpoint_path = hf_hub_download(
|
42 |
+
repo_id=kwargs.get("repo_id", "jadechoghari/mar"),
|
43 |
+
filename=kwargs.get("model_filename", "checkpoint-last.pth")
|
44 |
+
)
|
45 |
+
|
46 |
+
state_dict = torch.load(model_checkpoint_path, map_location=device)["model_ema"]
|
47 |
+
|
48 |
+
self.model.load_state_dict(state_dict, strict=False)
|
49 |
+
self.model.eval()
|
50 |
+
|
51 |
+
# download and load the vae
|
52 |
+
vae_checkpoint_path = hf_hub_download(
|
53 |
+
repo_id=kwargs.get("repo_id", "jadechoghari/mar"),
|
54 |
+
filename=kwargs.get("vae_filename", "kl16.ckpt")
|
55 |
+
)
|
56 |
+
|
57 |
+
vae = AutoencoderKL(embed_dim=16, ch_mult=(1, 1, 2, 2, 4), ckpt_path=vae_checkpoint_path)
|
58 |
+
vae = vae.to(device).eval()
|
59 |
+
|
60 |
+
# set up user-specified or default values for generation
|
61 |
+
seed = kwargs.get("seed", 0)
|
62 |
+
torch.manual_seed(seed)
|
63 |
+
np.random.seed(seed)
|
64 |
+
|
65 |
+
num_ar_steps = kwargs.get("num_ar_steps", 64)
|
66 |
+
cfg_scale = kwargs.get("cfg_scale", 4)
|
67 |
+
cfg_schedule = kwargs.get("cfg_schedule", "constant")
|
68 |
+
temperature = kwargs.get("temperature", 1.0)
|
69 |
+
class_labels = kwargs.get("class_labels", [207, 360, 388, 113, 355, 980, 323, 979])
|
70 |
+
|
71 |
+
# generate the tokens and images
|
72 |
+
with torch.cuda.amp.autocast():
|
73 |
+
sampled_tokens = self.model.sample_tokens(
|
74 |
+
bsz=len(class_labels), num_iter=num_ar_steps,
|
75 |
+
cfg=cfg_scale, cfg_schedule=cfg_schedule,
|
76 |
+
labels=torch.Tensor(class_labels).long().to(device),
|
77 |
+
temperature=temperature, progress=True
|
78 |
+
)
|
79 |
+
|
80 |
+
sampled_images = vae.decode(sampled_tokens / 0.2325)
|
81 |
+
|
82 |
+
return sampled_images
|
83 |
+
|