-
Notifications
You must be signed in to change notification settings - Fork 54
/
Copy pathml_decoder.py
178 lines (153 loc) · 7.75 KB
/
ml_decoder.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
from typing import Optional
import torch
from torch import nn, Tensor
from torch.nn.modules.transformer import _get_activation_fn
def add_ml_decoder_head(model, num_classes=-1, num_of_groups=-1, decoder_embedding=768, zsl=0):
if num_classes == -1:
num_classes = model.num_classes
num_features = model.num_features
if hasattr(model, 'global_pool') and hasattr(model, 'fc'): # resnet50
model.global_pool = nn.Identity()
del model.fc
model.fc = MLDecoder(num_classes=num_classes, initial_num_features=num_features, num_of_groups=num_of_groups,
decoder_embedding=decoder_embedding, zsl=zsl)
elif hasattr(model, 'head'): # tresnet
if hasattr(model, 'global_pool'):
model.global_pool = nn.Identity()
del model.head
model.head = MLDecoder(num_classes=num_classes, initial_num_features=num_features, num_of_groups=num_of_groups,
decoder_embedding=decoder_embedding, zsl=zsl)
else:
print("model is not suited for ml-decoder")
exit(-1)
return model
class TransformerDecoderLayerOptimal(nn.Module):
def __init__(self, d_model, nhead=8, dim_feedforward=2048, dropout=0.1, activation="relu",
layer_norm_eps=1e-5) -> None:
super(TransformerDecoderLayerOptimal, self).__init__()
self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.dropout = nn.Dropout(dropout)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
# Implementation of Feedforward model
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.norm3 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.activation = _get_activation_fn(activation)
def __setstate__(self, state):
if 'activation' not in state:
state['activation'] = torch.nn.functional.relu
super(TransformerDecoderLayerOptimal, self).__setstate__(state)
def forward(self, tgt: Tensor, memory: Tensor, tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None) -> Tensor:
tgt = tgt + self.dropout1(tgt)
tgt = self.norm1(tgt)
tgt2 = self.multihead_attn(tgt, memory, memory)[0]
tgt = tgt + self.dropout2(tgt2)
tgt = self.norm2(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
tgt = tgt + self.dropout3(tgt2)
tgt = self.norm3(tgt)
return tgt
# @torch.jit.script
# class ExtrapClasses(object):
# def __init__(self, num_queries: int, group_size: int):
# self.num_queries = num_queries
# self.group_size = group_size
#
# def __call__(self, h: torch.Tensor, class_embed_w: torch.Tensor, class_embed_b: torch.Tensor, out_extrap:
# torch.Tensor):
# # h = h.unsqueeze(-1).expand(-1, -1, -1, self.group_size)
# h = h[..., None].repeat(1, 1, 1, self.group_size) # torch.Size([bs, 5, 768, groups])
# w = class_embed_w.view((self.num_queries, h.shape[2], self.group_size))
# out = (h * w).sum(dim=2) + class_embed_b
# out = out.view((h.shape[0], self.group_size * self.num_queries))
# return out
@torch.jit.script
class GroupFC(object):
def __init__(self, embed_len_decoder: int):
self.embed_len_decoder = embed_len_decoder
def __call__(self, h: torch.Tensor, duplicate_pooling: torch.Tensor, out_extrap: torch.Tensor):
for i in range(h.shape[1]):
h_i = h[:, i, :]
if len(duplicate_pooling.shape)==3:
w_i = duplicate_pooling[i, :, :]
else:
w_i = duplicate_pooling
out_extrap[:, i, :] = torch.matmul(h_i, w_i)
class MLDecoder(nn.Module):
def __init__(self, num_classes, num_of_groups=-1, decoder_embedding=768,
initial_num_features=2048, zsl=0):
super(MLDecoder, self).__init__()
embed_len_decoder = 100 if num_of_groups < 0 else num_of_groups
if embed_len_decoder > num_classes:
embed_len_decoder = num_classes
# switching to 768 initial embeddings
decoder_embedding = 768 if decoder_embedding < 0 else decoder_embedding
embed_standart = nn.Linear(initial_num_features, decoder_embedding)
# non-learnable queries
if not zsl:
query_embed = nn.Embedding(embed_len_decoder, decoder_embedding)
query_embed.requires_grad_(False)
else:
query_embed = None
# decoder
decoder_dropout = 0.1
num_layers_decoder = 1
dim_feedforward = 2048
layer_decode = TransformerDecoderLayerOptimal(d_model=decoder_embedding,
dim_feedforward=dim_feedforward, dropout=decoder_dropout)
self.decoder = nn.TransformerDecoder(layer_decode, num_layers=num_layers_decoder)
self.decoder.embed_standart = embed_standart
self.decoder.query_embed = query_embed
self.zsl = zsl
if self.zsl:
if decoder_embedding != 300:
self.wordvec_proj = nn.Linear(300, decoder_embedding)
else:
self.wordvec_proj = nn.Identity()
self.decoder.duplicate_pooling = torch.nn.Parameter(torch.Tensor(decoder_embedding, 1))
self.decoder.duplicate_pooling_bias = torch.nn.Parameter(torch.Tensor(1))
self.decoder.duplicate_factor = 1
else:
# group fully-connected
self.decoder.num_classes = num_classes
self.decoder.duplicate_factor = int(num_classes / embed_len_decoder + 0.999)
self.decoder.duplicate_pooling = torch.nn.Parameter(
torch.Tensor(embed_len_decoder, decoder_embedding, self.decoder.duplicate_factor))
self.decoder.duplicate_pooling_bias = torch.nn.Parameter(torch.Tensor(num_classes))
torch.nn.init.xavier_normal_(self.decoder.duplicate_pooling)
torch.nn.init.constant_(self.decoder.duplicate_pooling_bias, 0)
self.decoder.group_fc = GroupFC(embed_len_decoder)
self.train_wordvecs = None
self.test_wordvecs = None
def forward(self, x):
if len(x.shape) == 4: # [bs,2048, 7,7]
embedding_spatial = x.flatten(2).transpose(1, 2)
else: # [bs, 197,468]
embedding_spatial = x
embedding_spatial_786 = self.decoder.embed_standart(embedding_spatial)
embedding_spatial_786 = torch.nn.functional.relu(embedding_spatial_786, inplace=True)
bs = embedding_spatial_786.shape[0]
if self.zsl:
query_embed = torch.nn.functional.relu(self.wordvec_proj(self.decoder.query_embed))
else:
query_embed = self.decoder.query_embed.weight
# tgt = query_embed.unsqueeze(1).repeat(1, bs, 1)
tgt = query_embed.unsqueeze(1).expand(-1, bs, -1) # no allocation of memory with expand
h = self.decoder(tgt, embedding_spatial_786.transpose(0, 1)) # [embed_len_decoder, batch, 768]
h = h.transpose(0, 1)
out_extrap = torch.zeros(h.shape[0], h.shape[1], self.decoder.duplicate_factor, device=h.device, dtype=h.dtype)
self.decoder.group_fc(h, self.decoder.duplicate_pooling, out_extrap)
if not self.zsl:
h_out = out_extrap.flatten(1)[:, :self.decoder.num_classes]
else:
h_out = out_extrap.flatten(1)
h_out += self.decoder.duplicate_pooling_bias
logits = h_out
return logits