rahulvenkk
commited on
Commit
·
8e8833a
1
Parent(s):
19af009
big hard cwm
Browse files- app.py +15 -14
- cwm/model/model_factory.py +5 -0
- cwm/model/model_pretrain.py +17 -0
- cwm/model/modeling_pretrain_cleaned_soft.py +723 -0
app.py
CHANGED
@@ -26,7 +26,7 @@ device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
|
26 |
|
27 |
|
28 |
# Load CWM 3-frame model (automatically download pre-trained checkpoint)
|
29 |
-
model = model_factory.load_model('
|
30 |
|
31 |
model.requires_grad_(False)
|
32 |
model.eval()
|
@@ -89,7 +89,7 @@ import os
|
|
89 |
# preloaded_images = load_preuploaded_images()
|
90 |
#
|
91 |
# print("Preloaded images:", preloaded_images)
|
92 |
-
@spaces.GPU(duration=110)
|
93 |
def get_c(x, points):
|
94 |
x = utils.imagenet_normalize(x)#.to(device)
|
95 |
with torch.no_grad():
|
@@ -107,6 +107,7 @@ with gr.Blocks() as demo:
|
|
107 |
with gr.Column():
|
108 |
# Input image
|
109 |
original_image = gr.State(value=None) # store original image without arrows
|
|
|
110 |
input_image = gr.Image(type="numpy", label="Upload Image")
|
111 |
|
112 |
# Annotate arrows
|
@@ -125,12 +126,12 @@ with gr.Blocks() as demo:
|
|
125 |
output_image = gr.Image(type='numpy')
|
126 |
|
127 |
# Store the original image and resize to square size once uploaded
|
128 |
-
def resize_to_square(img, size=
|
129 |
print("Resizing image to square")
|
130 |
img = Image.fromarray(img)
|
131 |
transform = transforms.Compose([
|
132 |
transforms.Resize((size, size)),
|
133 |
-
transforms.CenterCrop(size)
|
134 |
])
|
135 |
img = transform(img) # .transpose(1, 2, 0)
|
136 |
|
@@ -142,13 +143,13 @@ with gr.Blocks() as demo:
|
|
142 |
img = np.array(Image.open(img_path))
|
143 |
# print(f"Image uploaded with shape: {input.shape}")
|
144 |
resized_img = resize_to_square(img)
|
145 |
-
return resized_img, resized_img, []
|
146 |
|
147 |
|
148 |
def store_img(img):
|
149 |
resized_img = resize_to_square(img) # Resize the uploaded image to a square
|
150 |
print(f"Image uploaded with shape: {resized_img.shape}")
|
151 |
-
return resized_img, resized_img, []
|
152 |
|
153 |
|
154 |
with gr.Row():
|
@@ -165,9 +166,9 @@ with gr.Blocks() as demo:
|
|
165 |
# # run_on_click=True,
|
166 |
# # label="Select an example image to test"
|
167 |
# )
|
168 |
-
gallery.select(load_img, outputs=[input_image, original_image, selected_points])
|
169 |
|
170 |
-
input_image.upload(store_img, [input_image], [input_image, original_image, selected_points])
|
171 |
|
172 |
# Get points and draw arrows or zero-length vectors based on the toggle
|
173 |
def get_point(img, sel_pix, zero_length, evt: gr.SelectData):
|
@@ -251,7 +252,7 @@ with gr.Blocks() as demo:
|
|
251 |
def run_model_on_points(points, input_image, original_image):
|
252 |
H = input_image.shape[0]
|
253 |
W = input_image.shape[1]
|
254 |
-
factor =
|
255 |
# Example: pretend the model processes points and returns a simple transformation on the image
|
256 |
points = torch.from_numpy(np.array(points).reshape(-1, 4)) * factor
|
257 |
|
@@ -260,8 +261,8 @@ with gr.Blocks() as demo:
|
|
260 |
img = Image.fromarray(original_image)
|
261 |
|
262 |
transform = transforms.Compose([
|
263 |
-
transforms.Resize((
|
264 |
-
transforms.CenterCrop(
|
265 |
])
|
266 |
img = np.array(transform(img))
|
267 |
|
@@ -272,7 +273,7 @@ with gr.Blocks() as demo:
|
|
272 |
img = img[None]
|
273 |
|
274 |
# reshape image to [B, C, T, H, W], C = 3, T = 3 (3-frame model), H = W = 224
|
275 |
-
x = img[:, :, None].expand(-1, -1,
|
276 |
|
277 |
# Imagenet-normalize the inputs (standardization)
|
278 |
|
@@ -290,9 +291,9 @@ with gr.Blocks() as demo:
|
|
290 |
return counterfactual
|
291 |
|
292 |
# Run model when the button is clicked
|
293 |
-
run_model_button.click(run_model_on_points, [selected_points, input_image,
|
294 |
|
295 |
|
296 |
|
297 |
# Launch the app
|
298 |
-
demo.queue().launch()
|
|
|
26 |
|
27 |
|
28 |
# Load CWM 3-frame model (automatically download pre-trained checkpoint)
|
29 |
+
model = model_factory.load_model('vitb_8x8patch_2frames_encoder_mask_token')#.to(device)
|
30 |
|
31 |
model.requires_grad_(False)
|
32 |
model.eval()
|
|
|
89 |
# preloaded_images = load_preuploaded_images()
|
90 |
#
|
91 |
# print("Preloaded images:", preloaded_images)
|
92 |
+
# @spaces.GPU(duration=110)
|
93 |
def get_c(x, points):
|
94 |
x = utils.imagenet_normalize(x)#.to(device)
|
95 |
with torch.no_grad():
|
|
|
107 |
with gr.Column():
|
108 |
# Input image
|
109 |
original_image = gr.State(value=None) # store original image without arrows
|
110 |
+
original_image_high_res = gr.State(value=None) # store original image without arrows
|
111 |
input_image = gr.Image(type="numpy", label="Upload Image")
|
112 |
|
113 |
# Annotate arrows
|
|
|
126 |
output_image = gr.Image(type='numpy')
|
127 |
|
128 |
# Store the original image and resize to square size once uploaded
|
129 |
+
def resize_to_square(img, size=512):
|
130 |
print("Resizing image to square")
|
131 |
img = Image.fromarray(img)
|
132 |
transform = transforms.Compose([
|
133 |
transforms.Resize((size, size)),
|
134 |
+
# transforms.CenterCrop(size)
|
135 |
])
|
136 |
img = transform(img) # .transpose(1, 2, 0)
|
137 |
|
|
|
143 |
img = np.array(Image.open(img_path))
|
144 |
# print(f"Image uploaded with shape: {input.shape}")
|
145 |
resized_img = resize_to_square(img)
|
146 |
+
return resized_img, resized_img, img, []
|
147 |
|
148 |
|
149 |
def store_img(img):
|
150 |
resized_img = resize_to_square(img) # Resize the uploaded image to a square
|
151 |
print(f"Image uploaded with shape: {resized_img.shape}")
|
152 |
+
return resized_img, resized_img, img, []
|
153 |
|
154 |
|
155 |
with gr.Row():
|
|
|
166 |
# # run_on_click=True,
|
167 |
# # label="Select an example image to test"
|
168 |
# )
|
169 |
+
gallery.select(load_img, outputs=[input_image, original_image, original_image_high_res, selected_points])
|
170 |
|
171 |
+
input_image.upload(store_img, [input_image], [input_image, original_image, original_image_high_res, selected_points])
|
172 |
|
173 |
# Get points and draw arrows or zero-length vectors based on the toggle
|
174 |
def get_point(img, sel_pix, zero_length, evt: gr.SelectData):
|
|
|
252 |
def run_model_on_points(points, input_image, original_image):
|
253 |
H = input_image.shape[0]
|
254 |
W = input_image.shape[1]
|
255 |
+
factor = 256/H
|
256 |
# Example: pretend the model processes points and returns a simple transformation on the image
|
257 |
points = torch.from_numpy(np.array(points).reshape(-1, 4)) * factor
|
258 |
|
|
|
261 |
img = Image.fromarray(original_image)
|
262 |
|
263 |
transform = transforms.Compose([
|
264 |
+
transforms.Resize((256, 256)),
|
265 |
+
# transforms.CenterCrop(256)
|
266 |
])
|
267 |
img = np.array(transform(img))
|
268 |
|
|
|
273 |
img = img[None]
|
274 |
|
275 |
# reshape image to [B, C, T, H, W], C = 3, T = 3 (3-frame model), H = W = 224
|
276 |
+
x = img[:, :, None].expand(-1, -1, 2, -1, -1)#.to(torch.float16)
|
277 |
|
278 |
# Imagenet-normalize the inputs (standardization)
|
279 |
|
|
|
291 |
return counterfactual
|
292 |
|
293 |
# Run model when the button is clicked
|
294 |
+
run_model_button.click(run_model_on_points, [selected_points, input_image, original_image_high_res], [output_image])
|
295 |
|
296 |
|
297 |
|
298 |
# Launch the app
|
299 |
+
demo.queue().launch(inbrowser=True, share=True)
|
cwm/model/model_factory.py
CHANGED
@@ -26,6 +26,11 @@ _model_catalogue ={
|
|
26 |
"init_fn": model_pretrain.vitb_8x8patch_2frames,
|
27 |
},
|
28 |
|
|
|
|
|
|
|
|
|
|
|
29 |
}
|
30 |
|
31 |
|
|
|
26 |
"init_fn": model_pretrain.vitb_8x8patch_2frames,
|
27 |
},
|
28 |
|
29 |
+
"vitb_8x8patch_2frames_encoder_mask_token": {
|
30 |
+
"path": "cwm/2frame_cwm_mask_token.pth",
|
31 |
+
"init_fn": model_pretrain.vitb_8x8patch_2frames_encoder_mask_token,
|
32 |
+
},
|
33 |
+
|
34 |
}
|
35 |
|
36 |
|
cwm/model/model_pretrain.py
CHANGED
@@ -818,6 +818,23 @@ def vitb_4x4patch_2frames(**kwargs):
|
|
818 |
**kwargs)
|
819 |
return model
|
820 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
821 |
# def base_8x8patch_2frames_1tube(**kwargs):
|
822 |
# model = pretrain_videomae_base_224_scaffold(
|
823 |
# patch_size=(8, 8),
|
|
|
818 |
**kwargs)
|
819 |
return model
|
820 |
|
821 |
+
from cwm.model.modeling_pretrain_cleaned_soft import pretrain_vit_base_256_scaffold
|
822 |
+
|
823 |
+
def vitb_8x8patch_2frames_encoder_mask_token(
|
824 |
+
use_flash_attention=False, **kwargs):
|
825 |
+
model = pretrain_vit_base_256_scaffold(
|
826 |
+
patch_size=(8, 8),
|
827 |
+
num_frames=2,
|
828 |
+
tubelet_size=1,
|
829 |
+
use_flash_attention=use_flash_attention,
|
830 |
+
interp_noise=False,
|
831 |
+
legacy=False,
|
832 |
+
xla_flash=False,
|
833 |
+
learn_pos_embed=True,
|
834 |
+
**kwargs)
|
835 |
+
return model
|
836 |
+
|
837 |
+
|
838 |
# def base_8x8patch_2frames_1tube(**kwargs):
|
839 |
# model = pretrain_videomae_base_224_scaffold(
|
840 |
# patch_size=(8, 8),
|
cwm/model/modeling_pretrain_cleaned_soft.py
ADDED
@@ -0,0 +1,723 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from functools import partial
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
from timm.models.layers import trunc_normal_ as __call_trunc_normal_
|
7 |
+
from einops import rearrange
|
8 |
+
from cwm.model.model_utils import Block, _cfg, PatchEmbed, get_sinusoid_encoding_table
|
9 |
+
|
10 |
+
from torch import Tensor
|
11 |
+
import cwm.utils as utils
|
12 |
+
|
13 |
+
def trunc_normal_(tensor, mean=0., std=1.):
|
14 |
+
__call_trunc_normal_(tensor, mean=mean, std=std, a=-std, b=std)
|
15 |
+
|
16 |
+
|
17 |
+
def interpolate_pos_encoding(pos_embed, n_frames, h, w):
|
18 |
+
N = pos_embed.shape[1]
|
19 |
+
if N == (h * w * n_frames):
|
20 |
+
return pos_embed
|
21 |
+
old_h = old_w = int((N / n_frames) ** 0.5)
|
22 |
+
patch_pos_embed = pos_embed.view(1, n_frames, old_h, old_w, -1).flatten(0, 1).permute(0, 3, 1, 2)
|
23 |
+
|
24 |
+
patch_pos_embed = F.interpolate(
|
25 |
+
patch_pos_embed,
|
26 |
+
size=(h, w),
|
27 |
+
mode='bicubic',
|
28 |
+
)
|
29 |
+
return patch_pos_embed.permute(0, 2, 3, 1).flatten(0, 2).unsqueeze(0)
|
30 |
+
|
31 |
+
PRINT_PADDING = False
|
32 |
+
|
33 |
+
class PretrainVisionTransformerEncoder(nn.Module):
|
34 |
+
""" Vision Transformer with support for patch or hybrid CNN input stage
|
35 |
+
"""
|
36 |
+
def __init__(self, img_size=224, patch_size=(16, 16), in_chans=3, num_classes=0, embed_dim=768, depth=12,
|
37 |
+
num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
|
38 |
+
drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None, tubelet_size=2,
|
39 |
+
use_learnable_pos_emb=False, num_frames=16, embed_per_frame=False, clumping_factor=None, block_func=Block, k_bias=False, interp_noise=False, block_kwargs={}, legacy=False, xla_flash=False, learn_pos_embed=False):
|
40 |
+
super().__init__()
|
41 |
+
self.num_classes = num_classes
|
42 |
+
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
|
43 |
+
self.patch_size = (tubelet_size,) + patch_size
|
44 |
+
self.pt, self.ph, self.pw = self.patch_size
|
45 |
+
self.h = int(img_size / self.ph)
|
46 |
+
self.w = int(img_size / self.pw)
|
47 |
+
self.hw = self.h * self.w
|
48 |
+
|
49 |
+
self.clumping_factor = clumping_factor
|
50 |
+
self.interp_noise = interp_noise
|
51 |
+
|
52 |
+
self.embed_dim = embed_dim
|
53 |
+
self.num_heads = num_heads
|
54 |
+
|
55 |
+
if self.clumping_factor is not None: # Clump the context frame for memory efficiency
|
56 |
+
self.clumping_embed = nn.Conv3d(in_channels=embed_dim, out_channels=embed_dim,
|
57 |
+
kernel_size=(tubelet_size, clumping_factor, clumping_factor),
|
58 |
+
stride=(tubelet_size, clumping_factor, clumping_factor))
|
59 |
+
|
60 |
+
self._embed_per_frame = embed_per_frame
|
61 |
+
if not self._embed_per_frame:
|
62 |
+
self.patch_embed = PatchEmbed(
|
63 |
+
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,tubelet_size=tubelet_size,num_frames=num_frames)
|
64 |
+
num_patches = self.patch_embed.num_patches
|
65 |
+
elif self._embed_per_frame:
|
66 |
+
assert (num_frames % tubelet_size) == 0
|
67 |
+
num_embeddings = (num_frames // tubelet_size)
|
68 |
+
self.patch_embed = nn.ModuleList([
|
69 |
+
PatchEmbed(
|
70 |
+
img_size=img_size, patch_size=patch_size,
|
71 |
+
in_chans=in_chans, embed_dim=embed_dim,
|
72 |
+
tubelet_size=tubelet_size, num_frames=tubelet_size)
|
73 |
+
for _ in range(num_embeddings)])
|
74 |
+
num_patches = self.patch_embed[0].num_patches * num_embeddings
|
75 |
+
|
76 |
+
self.num_patches = num_patches
|
77 |
+
self.num_frames = num_frames
|
78 |
+
print("NUM PATCHES IN ENCODER", self.num_patches)
|
79 |
+
|
80 |
+
self.pos_embed = get_sinusoid_encoding_table(num_patches, embed_dim)
|
81 |
+
|
82 |
+
if learn_pos_embed:
|
83 |
+
self.pos_embed = nn.Parameter(self.pos_embed)
|
84 |
+
|
85 |
+
self.learn_pos_embed = learn_pos_embed
|
86 |
+
|
87 |
+
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
|
88 |
+
self.blocks = nn.ModuleList([
|
89 |
+
block_func(
|
90 |
+
dim=embed_dim, in_dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
|
91 |
+
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
|
92 |
+
init_values=init_values, **block_kwargs, k_bias=k_bias, legacy=legacy, xla_flash=xla_flash)
|
93 |
+
for i in range(depth)])
|
94 |
+
self.norm = norm_layer(embed_dim)
|
95 |
+
self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
96 |
+
|
97 |
+
if use_learnable_pos_emb:
|
98 |
+
trunc_normal_(self.pos_embed, std=.02)
|
99 |
+
|
100 |
+
self.apply(self._init_weights)
|
101 |
+
|
102 |
+
def _set_pos_embed(self, dim=None):
|
103 |
+
if dim is None:
|
104 |
+
dim = self.embed_dim
|
105 |
+
if self.pos_embed is None:
|
106 |
+
self.pos_embed = get_sinusoid_encoding_table(
|
107 |
+
self.num_patches, dim)
|
108 |
+
|
109 |
+
|
110 |
+
def _init_weights(self, m):
|
111 |
+
if isinstance(m, nn.Linear):
|
112 |
+
nn.init.xavier_uniform_(m.weight)
|
113 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
114 |
+
nn.init.constant_(m.bias, 0)
|
115 |
+
elif isinstance(m, nn.LayerNorm):
|
116 |
+
nn.init.constant_(m.bias, 0)
|
117 |
+
nn.init.constant_(m.weight, 1.0)
|
118 |
+
|
119 |
+
def get_num_layers(self):
|
120 |
+
return len(self.blocks)
|
121 |
+
|
122 |
+
@torch.jit.ignore
|
123 |
+
def no_weight_decay(self):
|
124 |
+
return {'pos_embed', 'cls_token'}
|
125 |
+
|
126 |
+
def get_classifier(self):
|
127 |
+
return self.head
|
128 |
+
|
129 |
+
def reset_classifier(self, num_classes, global_pool=''):
|
130 |
+
self.num_classes = num_classes
|
131 |
+
self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
132 |
+
|
133 |
+
def _get_pos_embed(self):
|
134 |
+
return self.pos_embed
|
135 |
+
|
136 |
+
def forward_block(self, x, idx):
|
137 |
+
return self.blocks[idx](x)
|
138 |
+
|
139 |
+
def interpolate_tensor_with_mask_token(self,
|
140 |
+
x: Tensor, mask: Tensor, mask_token: Tensor, invert: bool = True
|
141 |
+
) -> Tensor:
|
142 |
+
"""
|
143 |
+
Where mask == (0 if invert else 1), return x
|
144 |
+
where mask == (1 if invert else 0), return mask_token
|
145 |
+
Linearly interpolate between these using value of mask.
|
146 |
+
"""
|
147 |
+
# mask_token = mask_token
|
148 |
+
# breakpoint()
|
149 |
+
B, N, C = x.shape
|
150 |
+
assert mask.shape[1] == N, (
|
151 |
+
f"Number of tokens in mask ({mask.shape[1]}) does not match "
|
152 |
+
f"number of tokens in input ({N})"
|
153 |
+
)
|
154 |
+
|
155 |
+
assert mask_token.shape[-1] == C, (
|
156 |
+
f"Dimensionality of mask token ({mask_token.shape[-1]}) does not match "
|
157 |
+
f"dimensionality of tokens in input ({C})"
|
158 |
+
)
|
159 |
+
|
160 |
+
# convert mask to interpolation weights in range [0., 1.]
|
161 |
+
mask = mask.to(x).clip(min=0.0, max=1.0)
|
162 |
+
mask = (1.0 - mask) if invert else mask
|
163 |
+
mask = mask.unsqueeze(-1) # [B, N, 1]
|
164 |
+
|
165 |
+
# expand mask token
|
166 |
+
mask_token = mask_token.view(1, 1, C).expand(B, N, -1)
|
167 |
+
|
168 |
+
# interpolate
|
169 |
+
start = mask_token
|
170 |
+
end = x
|
171 |
+
|
172 |
+
return start + mask * (end - start)
|
173 |
+
|
174 |
+
def interpolate_tensor_with_noise(self,
|
175 |
+
x: Tensor, mask: Tensor, invert: bool = True
|
176 |
+
) -> Tensor:
|
177 |
+
"""
|
178 |
+
Where mask == (0 if invert else 1), return x
|
179 |
+
where mask == (1 if invert else 0), return mask_token
|
180 |
+
Linearly interpolate between these using value of mask.
|
181 |
+
"""
|
182 |
+
# mask_token = mask_token
|
183 |
+
# breakpoint()
|
184 |
+
B, N, C = x.shape
|
185 |
+
assert mask.shape[1] == N, (
|
186 |
+
f"Number of tokens in mask ({mask.shape[1]}) does not match "
|
187 |
+
f"number of tokens in input ({N})"
|
188 |
+
)
|
189 |
+
|
190 |
+
# convert mask to interpolation weights in range [0., 1.]
|
191 |
+
mask = mask.to(x).clip(min=0.0, max=1.0)
|
192 |
+
mask = (1.0 - mask) if invert else mask
|
193 |
+
mask = mask.unsqueeze(-1) # [B, N, 1]
|
194 |
+
|
195 |
+
# ImageNet mean and std
|
196 |
+
mean = torch.tensor([0.485, 0.456, 0.406]).view(3, 1, 1)
|
197 |
+
std = torch.tensor([0.229, 0.224, 0.225]).view(3, 1, 1)
|
198 |
+
|
199 |
+
# Generate a 3x8x8 patch of random numbers from a normal distribution
|
200 |
+
# with the same mean and std as ImageNet images
|
201 |
+
rand_vec = torch.randn(B, N, 3, self.patch_size[-2], self.patch_size[-1]) * std + mean
|
202 |
+
|
203 |
+
rand_vec = rand_vec.to(x.device).to(x.dtype).view(B, N, -1)
|
204 |
+
# interpolate
|
205 |
+
start = rand_vec
|
206 |
+
end = x
|
207 |
+
|
208 |
+
return start + mask * (end - start)
|
209 |
+
|
210 |
+
def tokenize(self, x, mask=None):
|
211 |
+
|
212 |
+
if not self._embed_per_frame:
|
213 |
+
x = self.patch_embed(x)
|
214 |
+
elif self._embed_per_frame:
|
215 |
+
x = torch.cat([
|
216 |
+
self.patch_embed[i](
|
217 |
+
x[:,:,(i*self.pt):((i+1)*self.pt)])
|
218 |
+
for i in range(len(self.patch_embed))], 1)
|
219 |
+
|
220 |
+
pos_embed = self._get_pos_embed().type_as(x).to(x.device).clone()
|
221 |
+
if not self._learnable_pos_embed:
|
222 |
+
pos_embed = pos_embed.detach()
|
223 |
+
x = x + pos_embed
|
224 |
+
return (x, mask)
|
225 |
+
|
226 |
+
def tokenize_and_mask(self, x, mask):
|
227 |
+
|
228 |
+
x, mask = self.tokenize(x, mask)
|
229 |
+
B, _, C = x.shape
|
230 |
+
# breakpoint()
|
231 |
+
x_vis = x[~mask].reshape(B, -1, C)
|
232 |
+
return x_vis
|
233 |
+
|
234 |
+
def tokenize_and_mask_variable_size(self, x, mask):
|
235 |
+
|
236 |
+
x, mask = self.tokenize(x, mask)
|
237 |
+
B, _, C = x.shape
|
238 |
+
all_batches = []
|
239 |
+
max_len = 0
|
240 |
+
all_len = []
|
241 |
+
for i in range(B):
|
242 |
+
x_vis = x[i, ~mask[i]]
|
243 |
+
if x_vis.shape[0] > max_len:
|
244 |
+
max_len = x_vis.shape[0]
|
245 |
+
all_batches.append(x_vis)
|
246 |
+
all_len.append(x_vis.shape[0])
|
247 |
+
|
248 |
+
#pad all batches to max_len in a single line
|
249 |
+
x_vis = torch.stack([F.pad(batch, (0,0,0,max_len-batch.shape[0]), mode='constant', value=0) for batch in all_batches])
|
250 |
+
|
251 |
+
return x_vis, all_len
|
252 |
+
|
253 |
+
def forward_features(self, x, mask, move_patches, static_patches, delta, mask_token, res=1, return_feat_layer=None):
|
254 |
+
_, _, T, H, W = x.shape
|
255 |
+
|
256 |
+
if self.interp_noise:
|
257 |
+
#patchify x with patch size[0], patch size[1]
|
258 |
+
p0 = self.patch_size[-2]
|
259 |
+
p1 = self.patch_size[-1]
|
260 |
+
x = rearrange(x, 'b c t (h p0) (w p1) -> b (t h w) (p0 p1 c)', p0=p0, p1=p1, h=H//p0, w=W//p1) # x: [B, N, C]
|
261 |
+
|
262 |
+
x = self.interpolate_tensor_with_noise(x, mask, invert=True)
|
263 |
+
x = rearrange(x, 'b n (p c) -> b n p c', c=3)
|
264 |
+
# Notice: To visualize the reconstruction video, we add the predict and the original mean and var of each patch.
|
265 |
+
x = rearrange(x,
|
266 |
+
'b (t h w) (p0 p1 p2) c -> b c (t p0) (h p1) (w p2)',
|
267 |
+
p0=1,
|
268 |
+
p1=self.patch_size[-2],
|
269 |
+
p2=self.patch_size[-1],
|
270 |
+
h=H//self.patch_size[-2],
|
271 |
+
w=W//self.patch_size[-1])
|
272 |
+
|
273 |
+
x = embed = self.patch_embed(x)
|
274 |
+
|
275 |
+
if res != 1:
|
276 |
+
|
277 |
+
p0 = self.patch_size[-2]
|
278 |
+
p1 = self.patch_size[-1]
|
279 |
+
pos_embed = interpolate_pos_encoding(self.pos_embed, T, int(256 // p0 * res), int(256 // p1 * res))
|
280 |
+
else:
|
281 |
+
|
282 |
+
pos_embed = self._get_pos_embed()
|
283 |
+
|
284 |
+
pos_embed = pos_embed.type_as(x) # .to(x.device).clone()
|
285 |
+
|
286 |
+
if not self.learn_pos_embed:
|
287 |
+
pos_embed = pos_embed.to(x.device).clone().detach()
|
288 |
+
|
289 |
+
x = x + pos_embed
|
290 |
+
B, _, C = x.shape
|
291 |
+
# x_vis = x[~mask].reshape(B, -1, C) # ~mask means visible
|
292 |
+
if not self.interp_noise:
|
293 |
+
x_vis = self.interpolate_tensor_with_mask_token(x, mask, mask_token, invert=True)
|
294 |
+
else:
|
295 |
+
x_vis = x
|
296 |
+
|
297 |
+
if move_patches is not None:
|
298 |
+
|
299 |
+
assert B == 1, "Only support batch size 1 for now"
|
300 |
+
for (px, py) in move_patches:
|
301 |
+
idx = px * self.w + py
|
302 |
+
dx, dy = delta
|
303 |
+
nx, ny = px + dx, py + dy
|
304 |
+
new_idx = nx * self.w + ny + (self.patch_embed.num_frames - 1) * (self.h * self.w)
|
305 |
+
|
306 |
+
emb = embed[:, idx]
|
307 |
+
pos_emb = pos_embed[:, new_idx]
|
308 |
+
emb = emb + pos_emb
|
309 |
+
x_vis = torch.cat([x_vis, emb[None]], 1)
|
310 |
+
|
311 |
+
if static_patches is not None:
|
312 |
+
for (px, py) in static_patches:
|
313 |
+
idx = px * self.w + py
|
314 |
+
new_idx = px * self.w + py + (self.patch_embed.num_frames - 1) * (self.h * self.w)
|
315 |
+
emb = embed[:, idx]
|
316 |
+
pos_emb = pos_embed[:, new_idx]
|
317 |
+
emb = emb + pos_emb
|
318 |
+
x_vis = torch.cat([x_vis, emb[None]], 1)
|
319 |
+
|
320 |
+
for blk_idx, blk in enumerate(self.blocks):
|
321 |
+
x_vis = blk(x_vis)
|
322 |
+
if blk_idx == return_feat_layer:
|
323 |
+
return x_vis
|
324 |
+
|
325 |
+
x_vis = self.norm(x_vis)
|
326 |
+
return x_vis
|
327 |
+
|
328 |
+
def _set_inputs(self, *args, **kwargs):
|
329 |
+
pass
|
330 |
+
|
331 |
+
def forward(self, x, mask, mask_token, return_feat_layer=None, timestamps=None, move_patches=None, static_patches=None, delta=None, res=1):
|
332 |
+
self._set_inputs(x, mask)
|
333 |
+
# pass input through the encoder
|
334 |
+
x = self.forward_features(x, mask, move_patches, static_patches, delta, mask_token, return_feat_layer=return_feat_layer, res=res)
|
335 |
+
# if return_feat_layer is not None and is lesser than the number of blocks it means that we are returning the
|
336 |
+
# features of an intermediate block layer. in this case we do not want to apply the head layer
|
337 |
+
if return_feat_layer is not None and return_feat_layer < len(self.blocks):
|
338 |
+
return x
|
339 |
+
# if we are passing through the entire encoder transformer we apply the head layer
|
340 |
+
x = self.head(x)
|
341 |
+
return x
|
342 |
+
|
343 |
+
class PretrainVisionTransformerDecoder(nn.Module):
|
344 |
+
""" Vision Transformer with support for patch or hybrid CNN input stage
|
345 |
+
"""
|
346 |
+
def __init__(self, patch_size=(16, 16), num_classes=768, embed_dim=768, depth=12,
|
347 |
+
num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
|
348 |
+
drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None, block_func=Block, block_kwargs={}, k_bias=False, legacy=True, xla_flash=False
|
349 |
+
):
|
350 |
+
super().__init__()
|
351 |
+
|
352 |
+
|
353 |
+
self.num_classes = num_classes
|
354 |
+
|
355 |
+
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
|
356 |
+
self.patch_size = patch_size
|
357 |
+
|
358 |
+
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
|
359 |
+
self.blocks = nn.ModuleList([
|
360 |
+
block_func(
|
361 |
+
dim=embed_dim, in_dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
|
362 |
+
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
|
363 |
+
init_values=init_values, **block_kwargs, k_bias=k_bias, legacy=legacy, xla_flash=xla_flash)
|
364 |
+
for i in range(depth)])
|
365 |
+
self.norm = norm_layer(embed_dim)
|
366 |
+
self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
367 |
+
|
368 |
+
self.apply(self._init_weights)
|
369 |
+
|
370 |
+
|
371 |
+
def _init_weights(self, m):
|
372 |
+
if isinstance(m, nn.Linear):
|
373 |
+
nn.init.xavier_uniform_(m.weight)
|
374 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
375 |
+
nn.init.constant_(m.bias, 0)
|
376 |
+
elif isinstance(m, nn.LayerNorm):
|
377 |
+
nn.init.constant_(m.bias, 0)
|
378 |
+
nn.init.constant_(m.weight, 1.0)
|
379 |
+
|
380 |
+
def get_num_layers(self):
|
381 |
+
return len(self.blocks)
|
382 |
+
|
383 |
+
@torch.jit.ignore
|
384 |
+
def no_weight_decay(self):
|
385 |
+
return {'pos_embed', 'cls_token'}
|
386 |
+
|
387 |
+
def get_classifier(self):
|
388 |
+
return self.head
|
389 |
+
|
390 |
+
def reset_classifier(self, num_classes, global_pool=''):
|
391 |
+
self.num_classes = num_classes
|
392 |
+
self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
393 |
+
|
394 |
+
def forward_block(self, x, idx):
|
395 |
+
return self.blocks[idx](x)
|
396 |
+
|
397 |
+
def get_last_tokens(self, x, return_token_num):
|
398 |
+
if return_token_num > 0:
|
399 |
+
return self.head(self.norm(x[:,-return_token_num:]))
|
400 |
+
elif return_token_num == 0:
|
401 |
+
return self.head(self.norm(x))[:,x.size(1):]
|
402 |
+
else:
|
403 |
+
return self.head(self.norm(x))
|
404 |
+
|
405 |
+
def forward(self, x, return_token_num, return_feat_layer=None):
|
406 |
+
|
407 |
+
# pass input through the decoder
|
408 |
+
for blk_idx, blk in enumerate(self.blocks):
|
409 |
+
x = blk(x)
|
410 |
+
# if we are returning the features of an intermediate block
|
411 |
+
# do so and skip the remaining computation
|
412 |
+
if blk_idx == return_feat_layer:
|
413 |
+
return x
|
414 |
+
|
415 |
+
if return_token_num > 0:
|
416 |
+
x = self.head(self.norm(x[:, -return_token_num:])) # only return the mask tokens predict pixels
|
417 |
+
else:
|
418 |
+
x = self.head(self.norm(x))
|
419 |
+
|
420 |
+
return x
|
421 |
+
|
422 |
+
class PretrainVisionTransformer(nn.Module):
|
423 |
+
""" Vision Transformer with support for patch or hybrid CNN input stage
|
424 |
+
"""
|
425 |
+
default_input_kwargs = {'unnormalize': True}
|
426 |
+
def __init__(self,
|
427 |
+
img_size=224,
|
428 |
+
patch_size=(16, 16),
|
429 |
+
main_input=None,
|
430 |
+
main_input_kwargs=default_input_kwargs,
|
431 |
+
encoder_func=PretrainVisionTransformerEncoder,
|
432 |
+
encoder_in_chans=3,
|
433 |
+
encoder_num_classes=0,
|
434 |
+
encoder_embed_dim=768,
|
435 |
+
encoder_depth=12,
|
436 |
+
encoder_num_heads=12,
|
437 |
+
encoder_block_func=Block,
|
438 |
+
encoder_block_kwargs={},
|
439 |
+
decoder_num_classes=None, # For pretraining this parameter isn't relevant but must be set according to tube&patch size
|
440 |
+
decoder_embed_dim=512,
|
441 |
+
decoder_depth=8,
|
442 |
+
decoder_num_heads=8,
|
443 |
+
decoder_block_func=Block,
|
444 |
+
decoder_block_kwargs={},
|
445 |
+
mlp_ratio=4.,
|
446 |
+
qkv_bias=False,
|
447 |
+
k_bias=False,
|
448 |
+
qk_scale=None,
|
449 |
+
num_frames=16,
|
450 |
+
drop_rate=0.,
|
451 |
+
attn_drop_rate=0.,
|
452 |
+
drop_path_rate=0.,
|
453 |
+
norm_layer=nn.LayerNorm,
|
454 |
+
init_values=0.,
|
455 |
+
spacetime_separable_pos_embed=False,
|
456 |
+
tubelet_size=2,
|
457 |
+
num_classes=0, # avoid the error from create_fn in timm
|
458 |
+
in_chans=0, # avoid the error from create_fn in timm
|
459 |
+
embed_per_frame=False,
|
460 |
+
flow_model_ckpt=None,
|
461 |
+
flow_frames=None,
|
462 |
+
random_input=False,
|
463 |
+
use_flash_attention=False,
|
464 |
+
additional_decoder_for_transition=False,
|
465 |
+
additional_decoder_for_x3_hat=False,
|
466 |
+
clumping_factor=None,
|
467 |
+
return_detectron_format=False,
|
468 |
+
out_feature='out_feature',
|
469 |
+
interp_noise=False,
|
470 |
+
legacy=True,
|
471 |
+
xla_flash=False,
|
472 |
+
learn_pos_embed=False,
|
473 |
+
**kwargs
|
474 |
+
):
|
475 |
+
super().__init__()
|
476 |
+
|
477 |
+
encoder_block_kwargs.update({'flash_attention': use_flash_attention})
|
478 |
+
decoder_block_kwargs.update({'flash_attention': use_flash_attention})
|
479 |
+
|
480 |
+
self.clumping_factor = clumping_factor
|
481 |
+
|
482 |
+
self.interp_noise = interp_noise
|
483 |
+
|
484 |
+
self.learn_pos_embed = learn_pos_embed
|
485 |
+
|
486 |
+
if self.clumping_factor is not None:
|
487 |
+
print('Clumping factor = %d' % self.clumping_factor)
|
488 |
+
self.clumping_embed = nn.Conv3d(in_channels=decoder_embed_dim, out_channels=decoder_embed_dim,
|
489 |
+
kernel_size=(1, clumping_factor, clumping_factor),
|
490 |
+
stride=(1, clumping_factor, clumping_factor))
|
491 |
+
self.clumping_embed.apply(self._init_weights)
|
492 |
+
|
493 |
+
self.up = nn.ConvTranspose2d(decoder_embed_dim, decoder_embed_dim, kernel_size=2, stride=2)
|
494 |
+
self.up.apply(self._init_weights)
|
495 |
+
|
496 |
+
self.encoder = encoder_func(
|
497 |
+
img_size=img_size,
|
498 |
+
patch_size=patch_size,
|
499 |
+
in_chans=encoder_in_chans,
|
500 |
+
num_classes=encoder_num_classes,
|
501 |
+
embed_dim=encoder_embed_dim,
|
502 |
+
depth=encoder_depth,
|
503 |
+
num_heads=encoder_num_heads,
|
504 |
+
mlp_ratio=mlp_ratio,
|
505 |
+
qkv_bias=qkv_bias,
|
506 |
+
qk_scale=qk_scale,
|
507 |
+
drop_rate=drop_rate,
|
508 |
+
attn_drop_rate=attn_drop_rate,
|
509 |
+
drop_path_rate=drop_path_rate,
|
510 |
+
norm_layer=norm_layer,
|
511 |
+
init_values=init_values,
|
512 |
+
tubelet_size=tubelet_size,
|
513 |
+
num_frames=num_frames,
|
514 |
+
embed_per_frame=embed_per_frame,
|
515 |
+
block_func=encoder_block_func,
|
516 |
+
block_kwargs=encoder_block_kwargs,
|
517 |
+
clumping_factor=clumping_factor,
|
518 |
+
k_bias=k_bias,
|
519 |
+
interp_noise = interp_noise,
|
520 |
+
legacy=legacy,
|
521 |
+
xla_flash=xla_flash,
|
522 |
+
learn_pos_embed=learn_pos_embed,
|
523 |
+
**kwargs)
|
524 |
+
|
525 |
+
if not return_detectron_format:
|
526 |
+
self.decoder = PretrainVisionTransformerDecoder(
|
527 |
+
patch_size=patch_size,
|
528 |
+
num_classes= 3*tubelet_size*(patch_size[0]*patch_size[1]) if decoder_num_classes is None else decoder_num_classes,
|
529 |
+
embed_dim=decoder_embed_dim,
|
530 |
+
depth=decoder_depth,
|
531 |
+
num_heads=decoder_num_heads,
|
532 |
+
mlp_ratio=mlp_ratio,
|
533 |
+
qkv_bias=qkv_bias,
|
534 |
+
qk_scale=qk_scale,
|
535 |
+
drop_rate=drop_rate,
|
536 |
+
attn_drop_rate=attn_drop_rate,
|
537 |
+
drop_path_rate=drop_path_rate,
|
538 |
+
norm_layer=norm_layer,
|
539 |
+
init_values=init_values,
|
540 |
+
block_func=decoder_block_func,
|
541 |
+
k_bias=k_bias, xla_flash=xla_flash,
|
542 |
+
block_kwargs=decoder_block_kwargs, legacy=legacy)
|
543 |
+
|
544 |
+
self.encoder_to_decoder = nn.Linear(encoder_embed_dim, decoder_embed_dim, bias=k_bias)
|
545 |
+
|
546 |
+
if not self.interp_noise:
|
547 |
+
self.mask_token = nn.Parameter(torch.zeros(1, 1, encoder_embed_dim))
|
548 |
+
trunc_normal_(self.mask_token, std=.02)
|
549 |
+
else:
|
550 |
+
self.mask_token = None
|
551 |
+
|
552 |
+
self.timestamps = None
|
553 |
+
self.encoder.timestamps = None
|
554 |
+
|
555 |
+
if self.learn_pos_embed:
|
556 |
+
self.pos_embed = nn.Parameter(get_sinusoid_encoding_table(self.encoder.num_patches, decoder_embed_dim))
|
557 |
+
else:
|
558 |
+
self.pos_embed = get_sinusoid_encoding_table(self.encoder.num_patches, decoder_embed_dim)
|
559 |
+
|
560 |
+
self.num_frames = num_frames
|
561 |
+
self.num_patches = self.encoder.num_patches
|
562 |
+
if self.num_frames is not None:
|
563 |
+
self.num_patches_per_frame = self.num_patches // self.num_frames
|
564 |
+
else:
|
565 |
+
self.num_patches_per_frame = self.num_patches
|
566 |
+
self.patch_size = self.encoder.patch_size
|
567 |
+
if isinstance(img_size, int):
|
568 |
+
self.image_size = (img_size, img_size)
|
569 |
+
else:
|
570 |
+
assert hasattr(img_size, '__len__'), img_size
|
571 |
+
self.image_size = img_size
|
572 |
+
|
573 |
+
self.return_detectron_format = return_detectron_format
|
574 |
+
|
575 |
+
@property
|
576 |
+
def mask_size(self):
|
577 |
+
return (self.num_frames // self.patch_size[0],
|
578 |
+
self.image_size[-2] // self.patch_size[-2],
|
579 |
+
self.image_size[-1] // self.patch_size[-1])
|
580 |
+
|
581 |
+
def _init_weights(self, m):
|
582 |
+
if isinstance(m, nn.Linear):
|
583 |
+
nn.init.xavier_uniform_(m.weight)
|
584 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
585 |
+
nn.init.constant_(m.bias, 0)
|
586 |
+
elif isinstance(m, nn.LayerNorm):
|
587 |
+
nn.init.constant_(m.bias, 0)
|
588 |
+
nn.init.constant_(m.weight, 1.0)
|
589 |
+
|
590 |
+
def get_num_layers(self):
|
591 |
+
return len(self.blocks)
|
592 |
+
|
593 |
+
@torch.jit.ignore
|
594 |
+
def no_weight_decay(self):
|
595 |
+
return {'pos_embed', 'cls_token', 'mask_token'}
|
596 |
+
|
597 |
+
|
598 |
+
|
599 |
+
def unpatchify(self, x, mask):
|
600 |
+
# Define the input tensor
|
601 |
+
B, N, C = x.shape # batch size
|
602 |
+
h, w = self.mask_size[-2:]
|
603 |
+
patch_size = self.patch_size[-2:]
|
604 |
+
|
605 |
+
recon = torch.zeros(B, h*w, C).to(x)
|
606 |
+
recon[mask[:, -h*w:]] = x.flatten(0, 1)
|
607 |
+
|
608 |
+
rec_imgs = rearrange(recon, 'b n (p c) -> b n p c', c=3)
|
609 |
+
# Notice: To visualize the reconstruction video, we add the predict and the original mean and var of each patch.
|
610 |
+
rec_imgs = rearrange(rec_imgs,
|
611 |
+
'b (t h w) (p0 p1 p2) c -> b c (t p0) (h p1) (w p2)',
|
612 |
+
p0=1,
|
613 |
+
p1=patch_size[0],
|
614 |
+
p2=patch_size[1],
|
615 |
+
h=h,
|
616 |
+
w=w)
|
617 |
+
|
618 |
+
# MEAN = torch.from_numpy(np.array((0.485, 0.456, 0.406))[None, :, None, None, None]).cuda().half()
|
619 |
+
# STD = torch.from_numpy(np.array((0.229, 0.224, 0.225))[None, :, None, None, None]).cuda().half()
|
620 |
+
#
|
621 |
+
# rec_imgs = (rec_imgs - MEAN) / STD
|
622 |
+
|
623 |
+
return rec_imgs
|
624 |
+
|
625 |
+
|
626 |
+
def forward(self, x, mask, timestamps=None, return_feat_layer=None, res=1, *args, get_encoder_out=False, **kwargs):
|
627 |
+
|
628 |
+
_, _, T, _, _ = x.shape
|
629 |
+
|
630 |
+
self.device = x.device
|
631 |
+
|
632 |
+
enc_out = self.encoder(x, mask, self.mask_token, timestamps=timestamps, return_feat_layer=return_feat_layer, res=res, *args, **kwargs) # [B, N_vis, C_e]
|
633 |
+
|
634 |
+
x_vis = self.encoder_to_decoder(enc_out)
|
635 |
+
|
636 |
+
# check if we are returning the features of an intermediate block layer
|
637 |
+
if return_feat_layer is not None:
|
638 |
+
# if the returned layer is one of the encoder layers (the first N_enc layers) we return the features
|
639 |
+
# if the return feat layer is exactly N_enc then we are returning the layer after the entire encoder block
|
640 |
+
# in both cases this manifests as returning x_vis, since self.encoder will return either the final block embedding
|
641 |
+
# or the final head embedding depending on the return_feat_layer
|
642 |
+
# in either case we subtract the number of encoder blocks + 1 (for the intermediate embedding layer)
|
643 |
+
# from the return_feat_layer to get the correct index for the decoder block
|
644 |
+
return_feat_layer = return_feat_layer - len(self.encoder.blocks) - 1
|
645 |
+
if return_feat_layer < 0:
|
646 |
+
return x_vis
|
647 |
+
|
648 |
+
# add pos embedding
|
649 |
+
if res != 1:
|
650 |
+
p0 = self.patch_size[-2]
|
651 |
+
p1 = self.patch_size[-1]
|
652 |
+
pos_embed = interpolate_pos_encoding(self.pos_embed, T, int(256 // p0 * res), int(256 // p1 * res))
|
653 |
+
else:
|
654 |
+
pos_embed = self.pos_embed
|
655 |
+
dec_pos_embed = pos_embed.expand(x_vis.size(0), -1, -1).type_as(x)
|
656 |
+
|
657 |
+
if not self.learn_pos_embed:
|
658 |
+
dec_pos_embed = dec_pos_embed.to(x.device).clone().detach()
|
659 |
+
|
660 |
+
x_vis = x_vis + dec_pos_embed
|
661 |
+
|
662 |
+
# pass input through the decoder, this will automatically return an intermediate layer if return_feat_layer is set
|
663 |
+
x_all = self.decoder(x_vis, 0, return_feat_layer=return_feat_layer)
|
664 |
+
|
665 |
+
if get_encoder_out:
|
666 |
+
return x_all, enc_out
|
667 |
+
|
668 |
+
return x_all
|
669 |
+
|
670 |
+
def get_counterfactual(self, x, move_patches):
|
671 |
+
'''
|
672 |
+
:param x: input tensor [1, C, T, H, W]: support only batch size 1 for now
|
673 |
+
:param move_patches: torch tensor [N, 4] sized array where each row contains patch motion [x1, y1, x2, y2] in pixel coordinates
|
674 |
+
:return:
|
675 |
+
'''
|
676 |
+
B, _, T, H, H = x.shape
|
677 |
+
|
678 |
+
mask = torch.ones(B, self.encoder.hw * self.encoder.num_frames).to(x.device).bool()
|
679 |
+
mask[:, :self.encoder.hw * (self.encoder.num_frames - 1)] = False
|
680 |
+
|
681 |
+
move_patches = (move_patches / H) * self.encoder.h
|
682 |
+
move_patches = move_patches.to(torch.int64)
|
683 |
+
|
684 |
+
for x1, y1, x2, y2 in move_patches:
|
685 |
+
idx2 = x2 * self.encoder.w + y2 + (self.encoder.num_frames - 1) * (self.encoder.h * self.encoder.w)
|
686 |
+
mask[:, idx2] = False
|
687 |
+
im_x1 = x1 * self.encoder.ph
|
688 |
+
im_y1 = y1 * self.encoder.pw
|
689 |
+
im_x2 = x2 * self.encoder.ph
|
690 |
+
im_y2 = y2 * self.encoder.pw
|
691 |
+
x[:, :, -1, im_x2:im_x2 + self.encoder.ph, im_y2:im_y2 + self.encoder.pw] = x[:, :, -2,
|
692 |
+
im_x1:im_x1 + self.encoder.ph,
|
693 |
+
im_y1:im_y1 + self.encoder.pw]
|
694 |
+
|
695 |
+
prediction = self.forward(x, mask)[:, -self.encoder.hw:]
|
696 |
+
|
697 |
+
prediction = utils.unpatchify_cwm(
|
698 |
+
prediction,
|
699 |
+
patch_size=self.encoder.patch_size[-1],
|
700 |
+
) # reshape the output to an image
|
701 |
+
|
702 |
+
return prediction
|
703 |
+
|
704 |
+
|
705 |
+
def pretrain_vit_base_256_scaffold(**kwargs):
|
706 |
+
model = PretrainVisionTransformer(
|
707 |
+
img_size=256,
|
708 |
+
encoder_embed_dim=768,
|
709 |
+
encoder_depth=12,
|
710 |
+
encoder_num_heads=12,
|
711 |
+
encoder_num_classes=0,
|
712 |
+
decoder_embed_dim=768,
|
713 |
+
decoder_num_heads=12,
|
714 |
+
decoder_depth=12,
|
715 |
+
mlp_ratio=4,
|
716 |
+
|
717 |
+
qkv_bias=True,
|
718 |
+
k_bias=True,
|
719 |
+
norm_layer=partial(nn.LayerNorm, eps=1e-6),
|
720 |
+
**kwargs)
|
721 |
+
model.default_cfg = _cfg()
|
722 |
+
return model
|
723 |
+
|