-
Notifications
You must be signed in to change notification settings - Fork 0
/
Main_few_shot_MsKPRN.py
337 lines (274 loc) · 20.1 KB
/
Main_few_shot_MsKPRN.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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import init
from torch.autograd import Variable
from torch.optim.lr_scheduler import StepLR
from torch.optim.lr_scheduler import ReduceLROnPlateau
import numpy as np
import task_generator as tg
import os
import math
import argparse
import scipy as sp
import scipy.stats
import time
import models
parser = argparse.ArgumentParser(description="One Shot Visual Recognition")
parser.add_argument("-f","--feature_dim",type = int, default = 64)
parser.add_argument("-r","--relation_dim",type = int, default = 8)
parser.add_argument("-w","--class_num",type = int, default = 5)
parser.add_argument("-s","--support_num_per_class",type = int, default = 5)
parser.add_argument("-b","--query_num_per_class",type = int, default = 10)
parser.add_argument("-e","--episode",type = int, default= 500000)
parser.add_argument("-t","--test_episode", type = int, default = 600)
parser.add_argument("-v","--val_episode", type = int, default = 300)
parser.add_argument("-l","--learning_rate", type = float, default = 0.001)
parser.add_argument("-g","--gpu",type=int, default=0)
parser.add_argument("-u","--hidden_unit",type=int,default=10)
args = parser.parse_args()
# Hyper Parameters
METHOD = "MsKPRN_Models"
FEATURE_DIM = args.feature_dim
RELATION_DIM = args.relation_dim
CLASS_NUM = args.class_num
SUPPORT_NUM_PER_CLASS = args.support_num_per_class
QUERY_NUM_PER_CLASS = args.query_num_per_class
EPISODES = args.episode
TEST_EPISODE = args.test_episode
VAL_EPISODE = args.val_episode
LEARNING_RATE = args.learning_rate
GPU = args.gpu
HIDDEN_UNIT = args.hidden_unit
def mean_confidence_interval(data, confidence=0.95):
a = 1.0*np.array(data)
n = len(a)
m = np.mean(a)
s = scipy.stats.sem(a)
h = s * sp.stats.t._ppf((1+confidence)/2., n-1)
return m,h
def weights_init_kaiming(m):
classname = m.__class__.__name__
if classname.find('Conv2d') != -1:
init.kaiming_normal_(m.weight.data, a=0, mode='fan_in')
elif classname.find('Linear') != -1:
init.kaiming_normal_(m.weight.data, a=0, mode='fan_in')
elif classname.find('BatchNorm2d') != -1:
init.normal_(m.weight.data, 1.0, 0.02)
init.constant_(m.bias.data, 0.0)
def main():
metatrain_folders,metaval_folders,metatest_folders = tg.mini_imagenet_folders()
best_accuracy = 0.0
best_h = 0.0
EPISODE=0
print("init neural networks")
feature_encoder = models.FeatureEncoder().apply(weights_init_kaiming).cuda(GPU)
kron_rel_net= models.KronRelationNets().cuda(GPU).apply(weights_init_kaiming).cuda(GPU)
feature_encoder_optim = torch.optim.Adam(feature_encoder.parameters(),lr=LEARNING_RATE)
feature_encoder_scheduler = ReduceLROnPlateau(feature_encoder_optim,mode="max",factor=0.5,patience=2,verbose=True)
kron_rel_net_optim = torch.optim.Adam(kron_rel_net.parameters(),lr=LEARNING_RATE)
kron_rel_net_scheduler = ReduceLROnPlateau(kron_rel_net_optim,mode="max",factor=0.5,patience=2,verbose=True)
if os.path.exists(str(METHOD + "/miniImagenet_feature_encoder_" + str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl")):
feature_encoder.load_state_dict(torch.load(str(METHOD + "/miniImagenet_feature_encoder_" + str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl")))
print("load feature encoder success")
print("load feature encoder success", file=F_txt)
if os.path.exists(str(METHOD + "/miniImagenet_kron_rel_net_" + str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl")):
kron_rel_net.load_state_dict(torch.load(str(METHOD + "/miniImagenet_kron_rel_net_" + str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl")))
print("load kron relation network success")
print("load kron relation network success", file=F_txt)
if os.path.exists(METHOD +"/checkpoint_"+ str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl"):
print("resuming training... ")
checkpoint = torch.load(METHOD +"/checkpoint_"+ str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl")
best_accuracy = checkpoint['best_accuracy']
best_h = checkpoint['best_h']
EPISODE = checkpoint['episode']
if os.path.exists(METHOD) == False:
os.system('mkdir ' + METHOD)
F_txt = open(METHOD +"/training_info_"+ str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.txt", 'a')
dim1= 19
dim2= 14
dim3= 10
print_training = False
for episode in range(EPISODE,EPISODES):
if(episode!=0):
if(print_training==True):
print("Training...")
print("Training...", file=F_txt)
print_training=False
# init dataset
task = tg.MiniImagenetTask(metatrain_folders,CLASS_NUM,SUPPORT_NUM_PER_CLASS,QUERY_NUM_PER_CLASS)
support_dataloader = tg.get_mini_imagenet_data_loader(task,num_per_class=SUPPORT_NUM_PER_CLASS,split="train",shuffle=False)
query_dataloader = tg.get_mini_imagenet_data_loader(task,num_per_class=QUERY_NUM_PER_CLASS,split="test",shuffle=True,train_query_argue=True)
# generate support data and query_data
supports_84, supports_64, supports_48, support_labels = support_dataloader.__iter__().next() #25*3*84*84
queries_84, queries_64, queries_48, query_labels = query_dataloader.__iter__().next()
# generate features
support_features_84 = feature_encoder(Variable(supports_84).cuda(GPU)) # 25*64*dim*dim
support_features_84 = support_features_84.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim1,dim1).sum(1)
query_features_84 = feature_encoder(Variable(queries_84).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_84 = support_features_84.unsqueeze(0).repeat(QUERY_NUM_PER_CLASS*CLASS_NUM,1,1,1,1)
support_features_ext_84= support_features_ext_84.view(QUERY_NUM_PER_CLASS*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim1,dim1)
query_features_ext_84 = query_features_84.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_84 = query_features_ext_84.reshape(QUERY_NUM_PER_CLASS*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim1,dim1)
# generate features
support_features_64 = feature_encoder(Variable(supports_64).cuda(GPU)) # 25*64*dim*dim
support_features_64 = support_features_64.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim2,dim2).sum(1)
query_features_64 = feature_encoder(Variable(queries_64).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_64 = support_features_64.unsqueeze(0).repeat(QUERY_NUM_PER_CLASS*CLASS_NUM,1,1,1,1)
support_features_ext_64= support_features_ext_64.view(QUERY_NUM_PER_CLASS*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim2,dim2)
query_features_ext_64 = query_features_64.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_64 = query_features_ext_64.reshape(QUERY_NUM_PER_CLASS*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim2,dim2)
# generate features
support_features_48 = feature_encoder(Variable(supports_48).cuda(GPU)) # 25*64*dim*dim
support_features_48 = support_features_48.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim3,dim3).sum(1)
query_features_48 = feature_encoder(Variable(queries_48).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_48 = support_features_48.unsqueeze(0).repeat(QUERY_NUM_PER_CLASS*CLASS_NUM,1,1,1,1)
support_features_ext_48= support_features_ext_48.view(QUERY_NUM_PER_CLASS*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim3,dim3)
query_features_ext_48 = query_features_48.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_48 = query_features_ext_48.reshape(QUERY_NUM_PER_CLASS*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim3,dim3)
scores_84 = kron_rel_net(query_features_ext_84,support_features_ext_84)
scores_64 = kron_rel_net(query_features_ext_64,support_features_ext_64)
scores_48 = kron_rel_net(query_features_ext_48,support_features_ext_48)
# define the loss function
mse = nn.MSELoss().cuda(GPU)
one_hot_labels = Variable(torch.zeros(QUERY_NUM_PER_CLASS*CLASS_NUM, CLASS_NUM).scatter_(1, query_labels.view(-1,1), 1).cuda(GPU))
loss = mse(scores_84,one_hot_labels) + mse(scores_64,one_hot_labels) + mse(scores_48,one_hot_labels)
feature_encoder.zero_grad()
kron_rel_net.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(feature_encoder.parameters(),0.5)
torch.nn.utils.clip_grad_norm_(kron_rel_net.parameters(),0.5)
feature_encoder_optim.step()
kron_rel_net_optim.step()
if (episode+1)%100 == 0:
print("episode:",episode+1,"loss",loss.item())
print("episode:",episode+1,"loss",loss.item(), file=F_txt)
if (episode+1)%5000 == 0:
print("Validation...")
print("Validation...", file= F_txt)
val_accuracies = []
with torch.no_grad():
for i in range(VAL_EPISODE):
total_rewards = 0
task = tg.MiniImagenetTask(metaval_folders,CLASS_NUM,SUPPORT_NUM_PER_CLASS,15)
support_dataloader = tg.get_mini_imagenet_data_loader(task,num_per_class=SUPPORT_NUM_PER_CLASS,split="train",shuffle=False)
num_per_class = 5
query_dataloader = tg.get_mini_imagenet_data_loader(task,num_per_class=num_per_class,split="test",shuffle=False)
support_images_84,support_images_64,support_images_48,support_labels = support_dataloader.__iter__().next()
for query_images_84,query_images_64,query_images_48,query_labels in query_dataloader:
query_size = query_labels.shape[0]
# generate features
support_features_84 = feature_encoder(Variable(support_images_84).cuda(GPU)) # 25*64*dim*dim
support_features_84 = support_features_84.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim1,dim1).sum(1)
query_features_84 = feature_encoder(Variable(query_images_84).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_84 = support_features_84.unsqueeze(0).repeat(num_per_class*CLASS_NUM,1,1,1,1)
support_features_ext_84= support_features_ext_84.view(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim1,dim1)
query_features_ext_84 = query_features_84.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_84 = query_features_ext_84.reshape(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim1,dim1)
# generate features
support_features_64 = feature_encoder(Variable(support_images_64).cuda(GPU)) # 25*64*dim*dim
support_features_64 = support_features_64.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim2,dim2).sum(1)
query_features_64 = feature_encoder(Variable(query_images_64).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_64 = support_features_64.unsqueeze(0).repeat(num_per_class*CLASS_NUM,1,1,1,1)
support_features_ext_64= support_features_ext_64.view(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim2,dim2)
query_features_ext_64 = query_features_64.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_64 = query_features_ext_64.reshape(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim2,dim2)
# generate features
support_features_48 = feature_encoder(Variable(support_images_48).cuda(GPU)) # 25*64*dim*dim
support_features_48 = support_features_48.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim3,dim3).sum(1)
query_features_48 = feature_encoder(Variable(query_images_48).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_48 = support_features_48.unsqueeze(0).repeat(num_per_class*CLASS_NUM,1,1,1,1)
support_features_ext_48= support_features_ext_48.view(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim3,dim3)
query_features_ext_48 = query_features_48.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_48 = query_features_ext_48.reshape(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim3,dim3)
scores_84 = kron_rel_net(query_features_ext_84,support_features_ext_84)
scores_64 = kron_rel_net(query_features_ext_64,support_features_ext_64)
scores_48 = kron_rel_net(query_features_ext_48,support_features_ext_48)
scores = scores_84+ scores_64 + scores_48
_,predict_labels = torch.max(scores.data,1)
rewards = [1 if predict_labels[j]==query_labels[j].cuda(GPU) else 0 for j in range(query_size)]
total_rewards += np.sum(rewards)
val_accuracy = total_rewards/1.0/CLASS_NUM/15
val_accuracies.append(val_accuracy)
accuracy_val,h = mean_confidence_interval(val_accuracies)
print("val accuracy :",accuracy_val,"h:",h)
print("val accuracy :",accuracy_val,"h:",h, file=F_txt)
feature_encoder_scheduler.step(accuracy_val)
kron_rel_net_scheduler.step(accuracy_val)
if (episode)%2500 == 0:
print("Testing...")
print("Testing...", file=F_txt)
accuracies = []
with torch.no_grad():
for i in range(TEST_EPISODE):
total_rewards = 0
task = tg.MiniImagenetTask(metatest_folders,CLASS_NUM,SUPPORT_NUM_PER_CLASS,15)
support_dataloader = tg.get_mini_imagenet_data_loader(task,num_per_class=SUPPORT_NUM_PER_CLASS,split="train",shuffle=False)
num_per_class = 5
query_dataloader = tg.get_mini_imagenet_data_loader(task,num_per_class=num_per_class,split="test",shuffle=False)
support_images_84,support_images_64,support_images_48,support_labels = support_dataloader.__iter__().next()
for query_images_84,query_images_64,query_images_48,query_labels in query_dataloader:
query_size = query_labels.shape[0]
# generate features
support_features_84 = feature_encoder(Variable(support_images_84).cuda(GPU)) # 25*64*dim*dim
support_features_84 = support_features_84.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim1,dim1).sum(1)
query_features_84 = feature_encoder(Variable(query_images_84).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_84 = support_features_84.unsqueeze(0).repeat(num_per_class*CLASS_NUM,1,1,1,1)
support_features_ext_84= support_features_ext_84.view(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim1,dim1)
query_features_ext_84 = query_features_84.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_84 = query_features_ext_84.reshape(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim1,dim1)
# generate features
support_features_64 = feature_encoder(Variable(support_images_64).cuda(GPU)) # 25*64*dim*dim
support_features_64 = support_features_64.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim2,dim2).sum(1)
query_features_64 = feature_encoder(Variable(query_images_64).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_64 = support_features_64.unsqueeze(0).repeat(num_per_class*CLASS_NUM,1,1,1,1)
support_features_ext_64= support_features_ext_64.view(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim2,dim2)
query_features_ext_64 = query_features_64.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_64 = query_features_ext_64.reshape(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim2,dim2)
# generate features
support_features_48 = feature_encoder(Variable(support_images_48).cuda(GPU)) # 25*64*dim*dim
support_features_48 = support_features_48.view(CLASS_NUM,SUPPORT_NUM_PER_CLASS,FEATURE_DIM,dim3,dim3).sum(1)
query_features_48 = feature_encoder(Variable(query_images_48).cuda(GPU))
# form the QURY_NUM_PER_CLASSxCLASS_NUM relation pairs
support_features_ext_48 = support_features_48.unsqueeze(0).repeat(num_per_class*CLASS_NUM,1,1,1,1)
support_features_ext_48= support_features_ext_48.view(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim3,dim3)
query_features_ext_48 = query_features_48.unsqueeze(1).repeat(1,CLASS_NUM,1,1,1)
query_features_ext_48 = query_features_ext_48.reshape(num_per_class*CLASS_NUM*CLASS_NUM,FEATURE_DIM,dim3,dim3)
scores_84 = kron_rel_net(query_features_ext_84,support_features_ext_84)
scores_64 = kron_rel_net(query_features_ext_64,support_features_ext_64)
scores_48 = kron_rel_net(query_features_ext_48,support_features_ext_48)
scores = scores_84+ scores_64 + scores_48
_,predict_labels = torch.max(scores.data,1)
rewards = [1 if predict_labels[j]==query_labels[j].cuda(GPU) else 0 for j in range(query_size)]
total_rewards += np.sum(rewards)
accuracy = total_rewards/1.0/CLASS_NUM/15
accuracies.append(accuracy)
test_accuracy,h = mean_confidence_interval(accuracies)
print("test accuracy :",test_accuracy,"h:",h)
print("best accuracy :",best_accuracy,"h:",best_h)
print("test accuracy :",test_accuracy,"h:",h, file = F_txt)
print("best accuracy :",best_accuracy,"h:",best_h, file = F_txt)
if test_accuracy > best_accuracy:
# save networks
torch.save(feature_encoder.state_dict(),str(METHOD + "/miniImagenet_feature_encoder_" + str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl"))
torch.save(kron_rel_net.state_dict(),str(METHOD + "/miniImagenet_kron_rel_net_" + str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl"))
torch.save({'best_accuracy': test_accuracy, 'best_h': best_h, 'episode': episode }, METHOD +"/checkpoint_"+ str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl")
print("save networks for episode:",episode)
print("save networks for episode:",episode, file = F_txt)
best_accuracy = test_accuracy
best_h = h
print_training= True
F_txt.flush()
os.fsync(F_txt.fileno())
torch.save({'best_accuracy': test_accuracy, 'best_h': best_h, 'episode': episode }, METHOD +"/checkpoint_"+ str(CLASS_NUM) +"way_" + str(SUPPORT_NUM_PER_CLASS) +"shot.pkl")
F_txt.close()
if __name__ == '__main__':
main()