-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
222 lines (175 loc) · 8.96 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
213
214
215
216
217
218
219
220
221
222
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.models import resnet101,resnet152
from pytorch_pretrained_vit import ViT
from torchvision.models._utils import IntermediateLayerGetter
from torchsummary import summary
import sys
torch.set_printoptions(threshold=sys.maxsize)
class SelfAttention(nn.Module):
def __init__(self, hidden_size, heads):
super(SelfAttention, self).__init__()
self.hidden_size = hidden_size
self.heads = heads
self.head_dim = hidden_size // heads
assert (self.head_dim*heads) == hidden_size, "Hidden size must be divisible by heads"
# Define the three linear transformation of the embeddings
self.linear_query = nn.Linear(self.head_dim, self.head_dim,bias=False)
self.linear_key = nn.Linear(self.head_dim, self.head_dim,bias=False)
self.linear_values = nn.Linear(self.head_dim, self.head_dim,bias=False)
self.fc_out = nn.Linear(hidden_size,hidden_size)
def forward(self,queries,keys,values,mask=None):
# Get information for multi head attention
BATCH_SIZE = queries.shape[0]
value_len,key_len,query_len = values.shape[1],keys.shape[1],queries.shape[1]
values = values.reshape((BATCH_SIZE,value_len,self.heads,self.head_dim))
keys = keys.reshape((BATCH_SIZE,key_len,self.heads,self.head_dim))
queries = queries.reshape((BATCH_SIZE,query_len,self.heads,self.head_dim))
queries = self.linear_query(queries)
values = self.linear_values(values)
keys = self.linear_key(keys)
weights = torch.einsum('nqhd,nkhd->nhqk',[queries,keys])
if mask is not None:
weights = weights.masked_fill(mask==0,float('-1e20'))
weights = F.softmax(weights/(self.hidden_size**(1/2)), dim=3)
# weights [BATCH_SIZE, heads, LEN, LEN]
# values [BATCH_SIZE,LEN,heads,head_dim]
# out [BATCH_SIZE,LEN,heads,head_dim]
out = torch.einsum('nhqk,nkhd->nqhd',[weights,values]).reshape(BATCH_SIZE,query_len,self.heads*self.head_dim)
out = self.fc_out(out)
return out
class Transformer_block(nn.Module):
def __init__(self, hidden_size, heads, dropout, forward_expansion):
super(Transformer_block,self).__init__()
self.attention = SelfAttention(hidden_size,heads)
self.norm1 = nn.LayerNorm(hidden_size)
self.norm2 = nn.LayerNorm(hidden_size)
self.feed_forward = nn.Sequential(
nn.Linear(hidden_size,forward_expansion*hidden_size),
nn.ReLU(),
nn.Linear(forward_expansion*hidden_size,hidden_size)
)
self.dropout = nn.Dropout(dropout)
def forward(self, queries, keys, values, mask=None):
attention = self.attention(queries,keys,values,mask)
x = self.dropout(self.norm1(attention+queries))
forward = self.feed_forward(x)
out = self.dropout(self.norm2(forward+x))
return out
class Encoder(nn.Module):
def __init__(self,hidden_size,num_layers,heads,device,forward_exp,dropout,image_position=64):
super(Encoder,self).__init__()
self.hidden_size = hidden_size
self.device = device
self.position_embed = nn.Embedding(image_position,hidden_size)
self.linear_proj = nn.Linear(2048,hidden_size)
self.layers = nn.ModuleList([
Transformer_block(hidden_size,heads,dropout,forward_exp)
]*num_layers)
self.dropout = nn.Dropout(dropout)
def forward(self,x,mask=None):
N, LEN, _= x.shape
positions = torch.arange(0,LEN).expand(N,LEN).to(self.device) # Crea una matrice di dimensione N,LEN con numeri progressivi in ogni riga
out = self.dropout(self.linear_proj(x) + self.position_embed(positions)) # Need positional embeedding
for layer in self.layers:
out = layer(out, out, out, mask)
return out
class DecoderBlock(nn.Module):
def __init__(self,hidden_size,heads,forward_exp,dropout):
super(DecoderBlock,self).__init__()
self.attention = SelfAttention(hidden_size,heads)
self.norm = nn.LayerNorm(hidden_size)
self.transformer_block = Transformer_block(hidden_size,heads,dropout,forward_exp)
self.dropout = nn.Dropout(dropout)
def forward(self,x, keys, values,trg_mask):
attention = self.attention(x,x,x,trg_mask)
queries = self.dropout(self.norm(attention+x))
out = self.transformer_block(queries,keys,values)
return out
class Decoder(nn.Module):
def __init__(self,trg_vocab_size,hidden_size,num_layers,heads,forward_exp,dropout,device,max_len):
super(Decoder,self).__init__()
self.device = device
self.word_embedding = nn.Embedding(trg_vocab_size,hidden_size,padding_idx=0)
self.position_embedding = nn.Embedding(max_len,hidden_size)
self.layers = nn.ModuleList(
[DecoderBlock(hidden_size,heads,forward_exp,dropout)
]*num_layers
)
self.fc = nn.Linear(hidden_size,trg_vocab_size)
self.dropout = nn.Dropout(dropout)
def forward(self,x,enc_out,trg_mask):
N, LEN = x.shape
positions = torch.arange(0,LEN).expand(N,LEN).to(self.device)
out = self.dropout(self.word_embedding(x) + self.position_embedding(positions))
for layer in self.layers:
out = layer(out,enc_out,enc_out,trg_mask)
out = self.fc(out)
return out
class ResnetBackbone(nn.Module):
def __init__(self,pretrained: bool,trainable: bool):
super(ResnetBackbone,self).__init__()
backbone = resnet152(pretrained=pretrained)
return_layers = {'layer4': "0"}
# Pretraining as in catr!
for name , parameter in backbone.named_parameters():
if not trainable or 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
parameter.requires_grad_(False)
self.body = IntermediateLayerGetter(backbone, return_layers=return_layers)
def forward(self, x):
# x is a tensor of shape [N, C, H, W]
xs = self.body(x)
out = xs['0']
out = out.view(out.shape[0],-1,out.shape[1])
return out
class ViTBackbone(nn.Module):
def __init__(self,pretrained: bool,trainable: bool):
super(ViTBackbone,self).__init__()
self.backbone = ViT('B_16_imagenet1k',pretrained=pretrained,image_size=256)
#summary(backbone, input_size=(3,256,256),device='cpu')
# for name, param in backbone.named_parameters():
# print(name, param.shape)
return_layers = {'norm': "0"}
# Pretraining as in catr!
for _, parameter in self.backbone.named_parameters():
# if not train_backbone or 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
if not trainable:
parameter.requires_grad_(False)
self.body = IntermediateLayerGetter(self.backbone, return_layers=return_layers)
def forward(self, x):
# x is a tensor of shape [N, C, H, W]
x = self.backbone(x)
# xs = self.body(x)
# out = xs['0']
# out = out.reshape(out.shape[0],-1,out.shape[1])
# return out
return x[:,1:,:]
class Transformer(nn.Module):
def __init__(self,trg_vocab_size,trg_pad_idx,hidden_size=256,
num_layers=6,forward_exp=4,heads=8,dropout=0,device='cuda',max_length=100, backbone='resnet',pretrained = True, train_backbone=True):
super(Transformer,self).__init__()
self.backbone_type = backbone
self.decoder = Decoder(trg_vocab_size,hidden_size,num_layers,heads,forward_exp,dropout,device,max_length)
if(self.backbone_type == 'resnet'):
self.backbone = ResnetBackbone(pretrained=pretrained,trainable=train_backbone)
self.encoder = Encoder(hidden_size,num_layers,heads,device,forward_exp,dropout)
elif(self.backbone_type == 'vit'):
self.backbone = ViTBackbone(pretrained=pretrained,trainable=train_backbone)
self.linear_proj = nn.Linear(768,hidden_size,bias=False)
self.trg_pad_idx = trg_pad_idx
self.device = device
self.max_len = max_length
self.hidden_size = hidden_size
def forward(self,src,trg,trg_mask):
if(self.backbone_type=='resnet'):
src = self.backbone(src) # Backbone prende una batch di dimensione (N,C,W,H) in ingresso e restituisce un tensore di dimensione (N,LEN,HIDDEN_SIZE)
enc_src = self.encoder(src,None)
elif(self.backbone_type=='vit'):
enc_src = self.backbone(src)
enc_src = self.linear_proj(enc_src)
else:
print('Backbone not supported')
out = self.decoder(trg,enc_src,trg_mask)
return out
# CHECK OUT THIS https://github.com/wtliao/ImageTransformer