-
Notifications
You must be signed in to change notification settings - Fork 4
/
resnet_train_relation.py
188 lines (144 loc) · 8.3 KB
/
resnet_train_relation.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
# 学习率下降，将relation改深,先不改权重初始化,用梯度裁剪 # 第二次，加了梯度裁剪，学习率每2k步降低
import torch
import torch.nn as nn
from utils import weight_init
# from RNet_resnet import RelationNetWork
from RNet_b import RelationNetWork
#from embed import CNNEncoder
from MiniImagenet2 import MiniImagenet
from torch.utils.data import DataLoader
from torch.autograd import Variable
import os
import numpy as np
from logger import Logger
from ResNet_feature import resnet18
from torch.optim.lr_scheduler import StepLR
LOG_DIR = './log_re'
logger = Logger(LOG_DIR)
# os.environ['CUDA_VISIBLE_DEVICES'] = '4'
def main():
n_way = 5
k_shot = 1
k_query = 5 # 5-way-5shot
batchsz = 3
best_acc = 0
mdfile1 = './ckpy/zres_feature-%d-way-%d-shot.pkl' %(n_way,k_shot)
mdfile2 = './ckpy/zres_relation-%d-way-%d-shot.pkl' %(n_way,k_shot)
# feature_embed = CNNEncoder().cuda()
# print(torch.cuda.is_available())
feature_embed = resnet18().cuda()
Relation_score = RelationNetWork().cuda() # relation_dim == 8 ??
# Relation_score.apply(weight_init)
feature_optim = torch.optim.Adam(feature_embed.parameters(), lr=0.001)
relation_opim = torch.optim.Adam(Relation_score.parameters(), lr=0.001)
feature_optim_scheduler = StepLR(feature_optim,step_size=10,gamma=0.5) # 1-shot 1w , 5-shot 5k step_size = 1
relation_opim_scheduler = StepLR(relation_opim,step_size=10,gamma=0.5)
loss_fn = torch.nn.MSELoss().cuda()
if os.path.exists(mdfile1):
print("load mdfile1...")
feature_embed.load_state_dict(torch.load(mdfile1))
if os.path.exists(mdfile2):
print("load mdfile2...")
Relation_score.load_state_dict(torch.load(mdfile2))
for epoch in range(100):
feature_optim_scheduler.step(epoch) # 降低学习率
relation_opim_scheduler.step(epoch)
mini = MiniImagenet('./mini-imagenet/', mode='train', n_way=n_way, k_shot=k_shot, k_query=k_query, batchsz=6000, resize=224) #38400
db = DataLoader(mini,batch_size=batchsz,shuffle=True,num_workers=4,pin_memory=True) # batch_size , 5*(1+15) , c, h, w
mini_val = MiniImagenet('./mini-imagenet/', mode='val', n_way=n_way, k_shot=k_shot, k_query=k_query, batchsz=200, resize=224) #9600
db_val = DataLoader(mini_val,batch_size=batchsz,shuffle=True,num_workers=4,pin_memory=True)
for step,batch in enumerate(db):
support_x = Variable(batch[0]).cuda() # [batch_size, n_way*(k_shot+k_query), c , h , w]
support_y = Variable(batch[1]).cuda()
query_x = Variable(batch[2]).cuda()
query_y = Variable(batch[3]).cuda()
bh,set1,c,h,w = support_x.size()
set2 = query_x.size(1)
feature_embed.train()
Relation_score.train()
# support_xf = feature_embed(support_x.view(bh*set1,c,h,w)).view(bh,set1,64,19,19) # 在 test 的 时候 重复
support_xf = feature_embed(support_x.view(bh * set1, c, h, w)).view(bh, set1, 256, 14, 14)
# query_xf = feature_embed(query_x.view(bh*set2,c,h,w)).view(bh,set2,64,19,19)
query_xf = feature_embed(query_x.view(bh*set2,c,h,w)).view(bh,set2,256,14,14)
# print("query_f:", query_xf.size())
# support_xf = support_xf.unsqueeze(1).expand(bh,set2,set1,64,19,19)
support_xf = support_xf.unsqueeze(1).expand(bh,set2,set1,256,14,14)
query_xf = query_xf.unsqueeze(2).expand(bh,set2,set1,256,14,14)
comb = torch.cat((support_xf,query_xf),dim=3) # bh,set2,set1,2c,h,w
# print(comb.is_cuda)
# print(comb.view(bh*set2*set1,2*64,19,19).is_cuda)
# print(comb.size())
score = Relation_score(comb.view(bh*set2*set1,2*256,14,14)).view(bh,set2,set1,1).squeeze(3)
support_yf = support_y.unsqueeze(1).expand(bh,set2,set1)
query_yf = query_y.unsqueeze(2).expand(bh,set2,set1)
label = torch.eq(support_yf,query_yf).float()
feature_optim.zero_grad()
relation_opim.zero_grad()
loss = loss_fn(score,label)
loss.backward()
torch.nn.utils.clip_grad_norm(feature_embed.parameters(),0.5) # 梯度裁剪？ 降低学习率？
torch.nn.utils.clip_grad_norm(Relation_score.parameters(),0.5)
feature_optim.step()
relation_opim.step()
# if step%100==0:
# print("step:",epoch+1,"train_loss: ",loss.data[0])
logger.log_value('resnet_{}-way-{}-shot loss：'.format(n_way, k_shot),loss.data[0])
if step%200==0:
print("---------test--------")
total_correct = 0
total_num = 0
accuracy = 0
closs = 0
for j,batch_test in enumerate(db_val):
# if (j%100==0):
# print(j,'-------------')
support_x = Variable(batch_test[0]).cuda()
support_y = Variable(batch_test[1]).cuda()
query_x = Variable(batch_test[2]).cuda()
query_y = Variable(batch_test[3]).cuda()
bh,set1,c,h,w = support_x.size()
set2 = query_x.size(1)
feature_embed.eval()
Relation_score.eval()
support_xf = feature_embed(support_x.view(bh*set1,c,h,w)).view(bh,set1,256,14,14) # 在 test 的 时候 重复
query_xf = feature_embed(query_x.view(bh*set2,c,h,w)).view(bh,set2,256,14,14)
support_xf = support_xf.unsqueeze(1).expand(bh,set2,set1,256,14,14)
query_xf = query_xf.unsqueeze(2).expand(bh,set2,set1,256,14,14)
comb = torch.cat((support_xf,query_xf),dim=3) # bh,set2,set1,2c,h,w
score = Relation_score(comb.view(bh*set2*set1,2*256,14,14)).view(bh,set2,set1,1).squeeze(3)
csupport_yf = support_y.unsqueeze(1).expand(bh, set2, set1)
cquery_yf = query_y.unsqueeze(2).expand(bh, set2, set1)
clabel = torch.eq(csupport_yf, cquery_yf).float()
closs += loss_fn(score, clabel).data[0]
rn_score_np = score.cpu().data.numpy() # 转numpy cpu
pred = []
support_y_np = support_y.cpu().data.numpy()
for ii,tb in enumerate(rn_score_np):
for jj,tset in enumerate(tb):
sim = []
for way in range(n_way):
sim.append(np.sum(tset[way*k_shot:(way+1)*k_shot]))
idx = np.array(sim).argmax()
pred.append(support_y_np[ii,idx*k_shot]) # 同一个类标签相同 ，注意还有batch维度
# ×k_shot是因为，上一个步用sum将k_shot压缩了
#此时的pred.size = [b.set2]
#print("pred.size=", np.array(pred).shape)
pred = Variable(torch.from_numpy(np.array(pred).reshape(bh,set2))).cuda()
correct = torch.eq(pred,query_y).sum()
total_correct += correct.data[0]
total_num += query_y.size(0)*query_y.size(1)
logger.log_value("val_loss{}-way-{}-shot loss:".format(n_way, k_shot),closs)
accuracy = total_correct/total_num
logger.log_value('acc : ', accuracy)
print("epoch:",epoch,"acc:",accuracy)
if accuracy>best_acc:
print("-------------------epoch",epoch,"step:",step,"acc:",accuracy,"---------------------------------------")
best_acc = accuracy
torch.save(feature_embed.state_dict(),mdfile1)
torch.save(Relation_score.state_dict(),mdfile2)
logger.step()
#if step% == 0 and step != 0:
# print("%d-way %d-shot %d batch | epoch:%d step:%d, loss:%f" %(n_way,k_shot,batchsz,epoch,step,loss.cpu().data[0]))
#logger.step()
if __name__=='__main__':
main()