-
Notifications
You must be signed in to change notification settings - Fork 0
/
Meta_model_define.py
266 lines (196 loc) · 10.4 KB
/
Meta_model_define.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
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
"""
-------------------------------File info-------------------------
% - File name: Meta_model_define.py
% - Description:
% - Input:
% - Output: None
% - Calls: None
% - usage:
% - Version： V1.0
% - Last update: 2022-10-28
% Copyright (C) PRMI, South China university of technology; 2022
% ------For Educational and Academic Purposes Only ------
% - Author : Chester.Wei.Xie, PRMI, SCUT/ GXU
% - Contact: [email protected]
------------------------------------------------------------------
"""
import torch.nn as nn
import torch.nn.functional as F
from torchvision import models
from utils import *
import math
from torch.nn import Parameter
class MYNET(nn.Module):
def __init__(self, args, mode=None):
super().__init__()
self.mode = mode
self.args = args
self.encoder = models.resnet18(pretrained=False)
self.encoder.conv1 = nn.Conv2d(1, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
self.num_features = self.encoder.fc.in_features
self.encoder = nn.Sequential(*list(self.encoder.children())[:-1])
self.fc = nn.Linear(self.num_features, (self.args.num_class - self.args.base_start_index), bias=False)
hdim = self.num_features
self.Selective_attn = SelectiveAttention(1, hdim, hdim, hdim, dropout=0.5)
def forward(self, input):
if self.mode == 'encoder':
input = self.encode(input)
return input
else:
support_idx, query_idx = input
logits = self._forward(support_idx, query_idx)
return logits
def _forward(self, incremental_support, query, base_support):
emb_dim = incremental_support.size(-1)
incremental_proto = incremental_support.mean(dim=1)
# [1, 30, 64]
base_proto = base_support.mean(dim=1) # [1, 15, 64]
num_batch = incremental_proto.shape[0]
num_incremental_proto = incremental_proto.shape[1]
num_base_proto = base_proto.shape[1]
num_query = query.shape[1] * query.shape[2]
query = query.view(-1, emb_dim).unsqueeze(1)
incremental_proto = incremental_proto.unsqueeze(1).expand(num_batch, num_query,
num_incremental_proto, emb_dim).contiguous()
incremental_proto = incremental_proto.view(num_batch * num_query, num_incremental_proto, emb_dim)
base_proto = base_proto.unsqueeze(1).expand(num_batch, num_query, num_base_proto, emb_dim).contiguous()
base_proto = base_proto.view(num_batch * num_query, num_base_proto, emb_dim)
combined = torch.cat([incremental_proto, query], 1) # Nk x (N + 1) x d, batch_size = NK
combined = self.Selective_attn(combined, combined, combined, combined, base_proto, base_proto)
incremental_proto, query = combined.split(num_incremental_proto, 1)
logits = F.cosine_similarity(query, incremental_proto, dim=-1)
logits = logits * self.args.temperature
return logits
def encode(self, x):
x = self.encoder(x)
x = x.squeeze(-1).squeeze(-1)
return x
def update_fc(self, dataloader, class_list, session):
for batch in dataloader:
data, label = [_.cuda() for _ in batch]
data = self.encode(data).detach()
if self.args.not_data_init:
new_fc = nn.Parameter(
torch.rand(len(class_list), self.num_features, device="cuda"),
requires_grad=True)
nn.init.kaiming_uniform_(new_fc, a=math.sqrt(5))
else:
new_fc = self.update_fc_avg(data, label, class_list)
if 'ft' in self.args.new_mode:
self.update_fc_ft(new_fc, data, label, session)
def update_fc_avg(self, data, label, class_list):
new_fc = []
for class_index in class_list:
class_index = class_index - self.args.base_start_index
data_index = (label == class_index).nonzero().squeeze(-1)
embedding = data[data_index]
proto = embedding.mean(0)
new_fc.append(proto)
self.fc.weight.data[class_index] = proto
new_fc = torch.stack(new_fc, dim=0)
return new_fc
class ScaledDotProductAttention(nn.Module):
""" Scaled Dot-Product Attention """
def __init__(self, temperature, attn_dropout=0.1):
super().__init__()
self.temperature = temperature
self.dropout = nn.Dropout(attn_dropout)
self.softmax = nn.Softmax(dim=2)
def forward(self, q, k, v):
attn = torch.bmm(q, k.transpose(1, 2))
attn = attn / self.temperature
log_attn = F.log_softmax(attn, 2)
attn = self.softmax(attn)
attn = self.dropout(attn)
output = torch.bmm(attn, v)
return output, attn, log_attn
class SelectiveAttention(nn.Module):
""" SelectiveAttention Attention module """
def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
super().__init__()
self.n_head = n_head
self.d_k = d_k
self.d_v = d_v
self.w_q = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_k = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_v = nn.Linear(d_model, n_head * d_v, bias=False)
nn.init.normal_(self.w_q.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_k)))
nn.init.normal_(self.w_k.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_k)))
nn.init.normal_(self.w_v.weight, mean=0, std=np.sqrt(2.0 / (d_model + d_v)))
self.self_attention = ScaledDotProductAttention(temperature=np.power(d_k, 0.5))
self.mutual_attention = ScaledDotProductAttention(temperature=np.power(d_k, 0.5))
self.selective_attention = ScaledDotProductAttention(temperature=np.power(d_k, 0.5))
self.fc_select = nn.Linear(n_head * d_v, d_model)
nn.init.xavier_normal_(self.fc_select.weight)
self.dropout_select = nn.Dropout(dropout)
self.layer_norm = nn.LayerNorm(d_model)
def forward(self, q_s, k_s, v_s, q_m, k_m, v_m):
d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
sz_b1, len_q1, _ = q_s.size()
sz_b1, len_k1, _ = k_s.size()
sz_b1, len_v1, _ = v_s.size()
sz_b2, len_q2, _ = q_m.size()
sz_b2, len_k2, _ = k_m.size()
sz_b2, len_v2, _ = v_m.size()
residual = q_s
q_s = self.w_q(q_s).view(sz_b1, len_q1, n_head, d_k)
k_s = self.w_k(k_s).view(sz_b1, len_k1, n_head, d_k)
v_s = self.w_v(v_s).view(sz_b1, len_v1, n_head, d_v)
q_m = self.w_q(q_m).view(sz_b2, len_q2, n_head, d_k)
k_m = self.w_k(k_m).view(sz_b2, len_k2, n_head, d_k)
v_m = self.w_v(v_m).view(sz_b2, len_v2, n_head, d_v)
q_s = q_s.permute(2, 0, 1, 3).contiguous().view(-1, len_q1, d_k) # (n*b) x lq x dk
k_s = k_s.permute(2, 0, 1, 3).contiguous().view(-1, len_k1, d_k) # (n*b) x lk x dk
v_s = v_s.permute(2, 0, 1, 3).contiguous().view(-1, len_v1, d_v) # (n*b) x lv x dv
q_m = q_m.permute(2, 0, 1, 3).contiguous().view(-1, len_q2, d_k) # (n*b) x lq x dk
k_m = k_m.permute(2, 0, 1, 3).contiguous().view(-1, len_k2, d_k) # (n*b) x lk x dk
v_m = v_m.permute(2, 0, 1, 3).contiguous().view(-1, len_v2, d_v) # (n*b) x lv x dv
self_output, _, _ = self.self_attention(q_s, k_s, v_s) # [num_head, num_samples, num_classes, fea_dim]
self_output = self_output.view(n_head, sz_b1, len_q1, d_v)
self_output = self_output.permute(1, 2, 0, 3).contiguous().view(sz_b1, len_q1, -1) # b x lq x (n*dv)
# [num_samples, num_classes, num_head, fea_dim] -> [num_samples, num_classes, num_head*fea_dim]
mutual_output, _, _ = self.mutual_attention(q_m, k_m, v_m) # [num_head, num_samples, num_classes, fea_dim]
mutual_output = mutual_output.view(n_head, sz_b2, len_q2, d_v)
mutual_output = mutual_output.permute(1, 2, 0, 3).contiguous().view(sz_b2, len_q2, -1) # b x lq x (n*dv)
# [num_samples, num_classes, num_head, fea_dim] -> [num_samples, num_classes, num_head*fea_dim]
sz_b3, len_q3, _ = self_output.size()
sz_b3, len_k3, _ = mutual_output.size()
sz_b3, len_v3, _ = self_output.size()
q_select = self.w_q(self_output).view(sz_b3, len_q3, n_head, d_k)
k_select = self.w_k(mutual_output).view(sz_b3, len_k3, n_head, d_k)
v_select = self.w_v(self_output).view(sz_b3, len_v3, n_head, d_v)
q_select = q_select.permute(2, 0, 1, 3).contiguous().view(-1, len_q3, d_k) # (n*b) x lq x dk
k_select = k_select.permute(2, 0, 1, 3).contiguous().view(-1, len_k3, d_k) # (n*b) x lk x dk
v_select = v_select.permute(2, 0, 1, 3).contiguous().view(-1, len_v3, d_v) # (n*b) x lv x dv
selective_output, _, _ = self.selective_attention(q_select, k_select, v_select)
selective_output = self.dropout_select(self.fc_select(selective_output))
output = self.layer_norm(selective_output + residual) # [num_samples, num_classes, fea_dim]
return output
def replace_base_fc(trainset, model, args):
# replace fc.weight with the embedding average of train data
model = model.eval()
trainloader = torch.utils.data.DataLoader(dataset=trainset, batch_size=128,
num_workers=8, pin_memory=True, shuffle=False)
# trainloader.dataset.transform = transform
embedding_list = []
label_list = []
# data_list=[]
with torch.no_grad():
for i, batch in enumerate(trainloader):
data, label = [_.cuda() for _ in batch]
model.mode = 'encoder'
embedding = model(data)
embedding_list.append(embedding.cpu())
label_list.append(label.cpu())
embedding_list = torch.cat(embedding_list, dim=0)
label_list = torch.cat(label_list, dim=0)
proto_list = []
for class_index in range(args.base_class):
class_index = class_index - args.base_start_index
data_index = (label_list == class_index).nonzero()
embedding_this = embedding_list[data_index.squeeze(-1)]
embedding_this = embedding_this.mean(0)
proto_list.append(embedding_this)
proto_list = torch.stack(proto_list, dim=0)
model.fc.weight.data[:args.base_class] = proto_list
return model