forked from IcarusWizard/MAE
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
212 lines (169 loc) · 8.66 KB
/
model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
import re
import torch # type: ignore
import numpy as np # type: ignore
from einops import repeat, rearrange # type: ignore
from einops.layers.torch import Rearrange # type: ignore
from timm.models.layers import trunc_normal_ # type: ignore
from timm.models.vision_transformer import Block # type: ignore
def random_indexes(size : int):
forward_indexes = np.arange(size)
np.random.shuffle(forward_indexes)
if np.random.random() < 0.5:
forward_indexes=[i for i in forward_indexes if i%2==1]+[i for i in forward_indexes if i%2==0]
else:
forward_indexes=[i for i in forward_indexes if i%2==0]+[i for i in forward_indexes if i%2==1]
backward_indexes = np.argsort(forward_indexes)
return forward_indexes, backward_indexes
def take_indexes(sequences, indexes):
#gather reorders repeat is used to give indexes the right size repeting the index in to the c dimention
return torch.gather(sequences, 0, repeat(indexes, 't b -> t b c', c=sequences.shape[-1]))
class PatchShuffle(torch.nn.Module):
def __init__(self, ratio) -> None:
super().__init__()
self.ratio = ratio
def forward(self, patches : torch.Tensor,indexes=None):
T, B, C = patches.shape #what are T, B C (total h*w batch/instances chanels?)
remain_T = int(T * (1 - self.ratio))
#if np.random.random() < 0.5: #randomly add a few extra scanelines
# remain_T=remain_T + 5
#else:
# remain_T=remain_T
indexes = [random_indexes(T) for _ in range(B)] #indexes are just linierindexes into the flaatened h*w image
forward_indexes = torch.as_tensor(np.stack([i[0] for i in indexes], axis=-1), dtype=torch.long).to(patches.device)
backward_indexes = torch.as_tensor(np.stack([i[1] for i in indexes], axis=-1), dtype=torch.long).to(patches.device)
patches = take_indexes(patches, forward_indexes)
#print('patches dim:', patches.shape)
patches = patches[:remain_T]
#print('patches after remain_T :', patches.shape)
return patches, forward_indexes, backward_indexes
class MAE_Encoder(torch.nn.Module):
def __init__(self,
image_width=32,
image_height=32,
image_channels=1,
patch_width=2,
patch_height=2,
emb_dim=192,
num_layer=12,
num_head=4,
mask_ratio=0.75
) -> None:
super().__init__()
self.cls_token = torch.nn.Parameter(torch.zeros(1, 1, emb_dim))
self.pos_embedding = torch.nn.Parameter(torch.zeros((image_width // patch_width * image_height // patch_height , 1, emb_dim)))
self.shuffle = PatchShuffle(mask_ratio)
self.patchify =torch.nn.Conv2d(image_channels, emb_dim, (1, patch_width),(1,patch_width))
self.transformer = torch.nn.Sequential(*[Block(emb_dim, num_head) for _ in range(num_layer)])
self.layer_norm = torch.nn.LayerNorm(emb_dim)
self.init_weight()
def init_weight(self):
trunc_normal_(self.cls_token, std=.02)
trunc_normal_(self.pos_embedding, std=.02)
def set_mask_ratio(self,ratio):
self.shuffle = PatchShuffle(ratio)
def forward(self, img, indexes=None):
patches = self.patchify(img)
#print('patches dim after conv:', patches.shape)
patches = rearrange(patches, 'b c h w -> (h w) b c')
#print('patches dim after rearange:', patches.shape)
patches = patches + self.pos_embedding
patches, forward_indexes, backward_indexes = self.shuffle(patches,indexes)
#print('indexes dim:', forward_indexes.shape,' ', backward_indexes.shape)
#adds a cls token with out a pos_embeding to the end? begining?
patches = torch.cat([self.cls_token.expand(-1, patches.shape[1], -1), patches], dim=0)
#print('patches after cat dim:', patches.shape)
patches = rearrange(patches, 't b c -> b t c')
#print('patches after rearrange:', patches.shape)
features = self.layer_norm(self.transformer(patches))
features = rearrange(features, 'b t c -> t b c')
return features, backward_indexes
class MAE_Decoder(torch.nn.Module):
def __init__(self,
image_width=32,
image_height=32,
image_channels=1,
patch_width=2,
patch_height=2,
emb_dim=192,
num_layer=4,
num_head=4
) -> None:
super().__init__()
self.mask_token = torch.nn.Parameter(torch.zeros(1, 1, emb_dim))
self.pos_embedding = torch.nn.Parameter(torch.zeros((image_width // patch_width * image_height // patch_height +1 , 1, emb_dim)))
self.transformer = torch.nn.Sequential(*[Block(emb_dim, num_head) for _ in range(num_layer)])
self.head = torch.nn.Linear(emb_dim, image_channels * patch_width *patch_height)
self.patch2img = Rearrange('(h w) b (c p1 p2) -> b c (h p1) (w p2)', c=image_channels, p2=patch_width, p1=patch_height, w=image_width//patch_width, h=image_height//patch_height)
self.init_weight()
def init_weight(self):
trunc_normal_(self.mask_token, std=.02)
trunc_normal_(self.pos_embedding, std=.02)
def forward(self, features, backward_indexes):
T = features.shape[0]
backward_indexes = torch.cat([torch.zeros(1, backward_indexes.shape[1]).to(backward_indexes), backward_indexes + 1], dim=0)
features = torch.cat([features, self.mask_token.expand(backward_indexes.shape[0] - features.shape[0], features.shape[1], -1)], dim=0)
features = take_indexes(features, backward_indexes)
features = features + self.pos_embedding
features = rearrange(features, 't b c -> b t c')
features = self.transformer(features)
features = rearrange(features, 'b t c -> t b c')
features = features[1:] # remove global feature
patches = self.head(features)
mask = torch.zeros_like(patches)
mask[T-1:] = 1
mask = take_indexes(mask, backward_indexes[1:] - 1)
img = self.patch2img(patches)
mask = self.patch2img(mask)
return img, mask
class MAE_ViT(torch.nn.Module):
def __init__(self,
image_width,
image_height,
image_channels,
patch_width,
patch_height,
emb_dim=192,
encoder_layer=12,
encoder_head=3,
decoder_layer=4,
decoder_head=3,
mask_ratio=0.75
) -> None:
super().__init__()
self.encoder = MAE_Encoder(image_width, image_height, image_channels, patch_width, patch_height, emb_dim, encoder_layer, encoder_head, mask_ratio)
self.decoder = MAE_Decoder(image_width, image_height, image_channels, patch_width, patch_height, emb_dim, decoder_layer, decoder_head)
def set_mask_ratio(self,ratio):
self.encoder.set_mask_ratio(ratio)
def forward(self, img):
features, backward_indexes = self.encoder(img)
predicted_img, mask = self.decoder(features, backward_indexes)
return predicted_img, mask
if __name__ == '__main__':
'''Testing/Trial Code'''
#shuffle = PatchShuffle(0.75)
#a = torch.rand(16, 2, 10)
#print('a: ',a.shape)
#b, forward_indexes, backward_indexes = shuffle(a)
#print('b: ',b.shape)
#print('forward index: ',forward_indexes.shape)
#example, chanel(color),w,h
img = torch.rand(1, 1, 32, 64)
sz=img.shape
print('img dim:', sz)
#patch=torch.nn.Conv2d(image_channels, emb_dim, patch_width, patch_height)(img)
#patch=torch.nn.Conv2d(1, 196, 2,2)(img)
#print(patch.shape)
#patch=torch.nn.Conv2d(1, 196, (1, 64),(1,64))(img)
#print(patch.shape)
#patch=torch.nn.Conv2d(1, 196, (2, 2),(2,2))(img)
#print(patch.shape)
encoder = MAE_Encoder( image_width=64,image_height=32,patch_height=1,patch_width=64)
decoder = MAE_Decoder( image_width=64,image_height=32,patch_height=1,patch_width=64)
features, backward_indexes = encoder(img,'bob')
#print('backward index: ',backward_indexes.shape)
#print(backward_indexes)
#predicted_img, mask = decoder(features, backward_indexes)
#print('mask:', mask.shape)
#print('predicted shape:',predicted_img.shape)
#loss = torch.mean((predicted_img - img) ** 2 * mask / 0.75)
#print(loss)