-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy pathmodel.py
120 lines (92 loc) · 4.09 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
import timm
import torch
from torch import nn
from timm.models.layers import trunc_normal_
from config import CFG
from utils import create_mask
class Encoder(nn.Module):
def __init__(self, model_name='deit3_small_patch16_384_in21ft1k', pretrained=False, out_dim=256):
super().__init__()
self.model = timm.create_model(
model_name, num_classes=0, global_pool='', pretrained=pretrained)
self.bottleneck = nn.AdaptiveAvgPool1d(out_dim)
def forward(self, x):
features = self.model(x)
return self.bottleneck(features[:, 1:])
class Decoder(nn.Module):
def __init__(self, vocab_size, encoder_length, dim, num_heads, num_layers):
super().__init__()
self.dim = dim
self.embedding = nn.Embedding(vocab_size, dim)
self.decoder_pos_embed = nn.Parameter(torch.randn(1, CFG.max_len-1, dim) * .02)
self.decoder_pos_drop = nn.Dropout(p=0.05)
decoder_layer = nn.TransformerDecoderLayer(d_model=dim, nhead=num_heads)
self.decoder = nn.TransformerDecoder(decoder_layer, num_layers=num_layers)
self.output = nn.Linear(dim, vocab_size)
self.encoder_pos_embed = nn.Parameter(torch.randn(1, encoder_length, dim) * .02)
self.encoder_pos_drop = nn.Dropout(p=0.05)
self.init_weights()
def init_weights(self):
for name, p in self.named_parameters():
if 'encoder_pos_embed' in name or 'decoder_pos_embed' in name:
print("skipping pos_embed...")
continue
if p.dim() > 1:
nn.init.xavier_uniform_(p)
trunc_normal_(self.encoder_pos_embed, std=.02)
trunc_normal_(self.decoder_pos_embed, std=.02)
def forward(self, encoder_out, tgt):
"""
encoder_out: shape(N, L, D)
tgt: shape(N, L)
"""
tgt_mask, tgt_padding_mask = create_mask(tgt)
tgt_embedding = self.embedding(tgt)
tgt_embedding = self.decoder_pos_drop(
tgt_embedding + self.decoder_pos_embed
)
encoder_out = self.encoder_pos_drop(
encoder_out + self.encoder_pos_embed
)
encoder_out = encoder_out.transpose(0, 1)
tgt_embedding = tgt_embedding.transpose(0, 1)
preds = self.decoder(memory=encoder_out,
tgt=tgt_embedding,
tgt_mask=tgt_mask,
tgt_key_padding_mask=tgt_padding_mask)
preds = preds.transpose(0, 1)
return self.output(preds)
def predict(self, encoder_out, tgt):
length = tgt.size(1)
padding = torch.ones(tgt.size(0), CFG.max_len-length-1).fill_(CFG.pad_idx).long().to(tgt.device)
tgt = torch.cat([tgt, padding], dim=1)
tgt_mask, tgt_padding_mask = create_mask(tgt)
# is it necessary to multiply it by math.sqrt(d) ?
tgt_embedding = self.embedding(tgt)
tgt_embedding = self.decoder_pos_drop(
tgt_embedding + self.decoder_pos_embed
)
encoder_out = self.encoder_pos_drop(
encoder_out + self.encoder_pos_embed
)
encoder_out = encoder_out.transpose(0, 1)
tgt_embedding = tgt_embedding.transpose(0, 1)
preds = self.decoder(memory=encoder_out,
tgt=tgt_embedding,
tgt_mask=tgt_mask,
tgt_key_padding_mask=tgt_padding_mask)
preds = preds.transpose(0, 1)
return self.output(preds)[:, length-1, :]
class EncoderDecoder(nn.Module):
def __init__(self, encoder, decoder):
super().__init__()
self.encoder = encoder
self.decoder = decoder
def forward(self, image, tgt):
encoder_out = self.encoder(image)
preds = self.decoder(encoder_out, tgt)
return preds
def predict(self, image, tgt):
encoder_out = self.encoder(image)
preds = self.decoder.predict(encoder_out, tgt)
return preds