-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_meta_transfer777.py
348 lines (293 loc) · 16.1 KB
/
main_meta_transfer777.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
338
339
340
341
342
343
344
345
346
347
348
import torch
import torch.nn as nn
from utlis.models7 import CNNEncoder1d, CNNEncoder2d, RelationNetwork1d, RelationNetwork2d
from torch.autograd import Variable
from torch.optim.lr_scheduler import StepLR
import numpy as np
import os
import utlis.transfer_generator as tg
import argparse
import pickle
from utlis.logger import setlogger
import logging
import matplotlib
#matplotlib.use('TkAgg')
import matplotlib.pyplot as plt
import torch.nn.functional as F
parser = argparse.ArgumentParser(description="meta transfer learning")
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", "--sample_num_per_class", type=int, default=1)
parser.add_argument("-b", "--batch_num_per_class", type=int, default=15)
parser.add_argument("-e", "--episode", type=int, default=100)
parser.add_argument("-t", "--test_episode", type=int, default=1000)
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)
parser.add_argument("-d", "--datatype", type=str, default='fft')
parser.add_argument("-m", "--modeltype", type=str, default='1d')
parser.add_argument("-n", "--snr", type=int, default=-100)
args = parser.parse_args()
# Hyper Parameters
FEATURE_DIM = args.feature_dim
RELATION_DIM = args.relation_dim
CLASS_NUM = args.class_num
SAMPLE_NUM_PER_CLASS = args.sample_num_per_class
BATCH_NUM_PER_CLASS = args.batch_num_per_class
EPISODE = args.episode
TEST_EPISODE = args.test_episode
LEARNING_RATE = args.learning_rate
GPU = args.gpu
HIDDEN_UNIT = args.hidden_unit
DATATYPE = args.datatype
MODELTYPE = args.modeltype
testgap = 5
SNR = args.snr
torch.set_default_tensor_type('torch.cuda.FloatTensor')
tg.set_seed(3)
# -----------------------------path---------------------------
root = '.\\result_transfer_mean'
path1 = DATATYPE '_' MODELTYPE '_' str(CLASS_NUM) 'way' str(SAMPLE_NUM_PER_CLASS) 'shot'
path = os.path.join(root, path1)
if not os.path.exists("%s" % path):
os.makedirs("%s" % path)
# ----------------------------logging-----------------------
setlogger(os.path.join("%s" % path, 'train.log'))
f = open("%s\\train.log" % path, 'w')
f.truncate()
f.close()
def main():
# Step 1: init data folders
logging.info("init data folders")
# init character folders for dataset construction
datapath = '.\\tempdata\\' str(CLASS_NUM) 'way' '.pkl'
if not os.path.exists("%s" % datapath):
metatrain_character_folders, metatest_character_folders = tg.pu_folders(CLASS_NUM)
os.makedirs('.\\tempdata')
output = open(datapath, 'wb')
pickle.dump((metatrain_character_folders, metatest_character_folders), output)
output.close()
else:
with open(datapath, 'rb') as pkl:
metatrain_character_folders, metatest_character_folders = pickle.load(pkl)
# Step 2: init neural networks
logging.info("init neural networks")
if MODELTYPE == '1d':
feature_encoder = CNNEncoder1d(FEATURE_DIM)
relation_network = RelationNetwork1d(FEATURE_DIM, RELATION_DIM)
elif MODELTYPE == '2d':
feature_encoder = CNNEncoder2d(FEATURE_DIM)
relation_network = RelationNetwork2d(FEATURE_DIM, RELATION_DIM)
feature_encoder.cuda(GPU)
relation_network.cuda(GPU)
feature_encoder_optim = torch.optim.Adam(feature_encoder.parameters(), lr=LEARNING_RATE)
feature_encoder_scheduler = StepLR(feature_encoder_optim, step_size=1000, gamma=0.5)
relation_network_optim = torch.optim.Adam(relation_network.parameters(), lr=LEARNING_RATE)
relation_network_scheduler = StepLR(relation_network_optim, step_size=1000, gamma=0.5)
# Step 3: build graph
logging.info("Training...")
last_accuracy = 0.0
accuracys = []
aepochs = []
losses = []
lepochs = []
finalsum = 0
for episode in range(EPISODE):
feature_encoder_scheduler.step(episode)
relation_network_scheduler.step(episode)
# init dataset
# sample_dataloader is to obtain previous samples for compare
# batch_dataloader is to batch samples for training
task = tg.puTask(metatrain_character_folders, CLASS_NUM, SAMPLE_NUM_PER_CLASS, BATCH_NUM_PER_CLASS)
task2 = tg.puTask(metatest_character_folders, CLASS_NUM, SAMPLE_NUM_PER_CLASS, BATCH_NUM_PER_CLASS)
sample_dataloader = tg.get_data_loader(task, num_per_class=SAMPLE_NUM_PER_CLASS, split="train", shuffle=False,
dt=DATATYPE, mt=MODELTYPE)
sample_dataloader2 = tg.get_data_loader(task2, num_per_class=SAMPLE_NUM_PER_CLASS, split="train", shuffle=False,
dt=DATATYPE, mt=MODELTYPE)
batch_dataloader = tg.get_data_loader(task, num_per_class=BATCH_NUM_PER_CLASS, split="test", shuffle=False,
dt=DATATYPE, mt=MODELTYPE)
# sample datas
samples, sample_labels = sample_dataloader.__iter__().next()
samples2, sample_labels2 = sample_dataloader2.__iter__().next()
batches, batch_labels = batch_dataloader.__iter__().next()
# calculate features
sample_features_o = feature_encoder(Variable(samples).cuda(GPU).float())[0] # 5x64*5*5
if MODELTYPE == '1d':
sample_features = sample_features_o.view(CLASS_NUM, SAMPLE_NUM_PER_CLASS, FEATURE_DIM, 5 * 5)
else:
sample_features = sample_features_o.view(CLASS_NUM, SAMPLE_NUM_PER_CLASS, FEATURE_DIM, 5, 5)
sample_features = torch.mean(sample_features, 1).squeeze(1)
batch_features = feature_encoder(Variable(batches.float()).cuda(GPU))[0] # 20x64*5*5
conv_normal1 = feature_encoder(Variable(samples).cuda(GPU).float())[1][0]
conv_noise1 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][0]
logp_x = F.log_softmax(conv_normal1, dim=-1)
p_y = F.softmax(conv_noise1, dim=-1)
kl_sum = F.kl_div(logp_x, p_y, reduction='sum')
kl_mean = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl1 sum is: ",kl_sum)
print("kl1 mean is: ", kl_mean)
conv_normal2 = feature_encoder(Variable(samples).cuda(GPU).float())[1][1]
conv_noise2 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][1]
logp_x = F.log_softmax(conv_normal2, dim=-1)
p_y = F.softmax(conv_noise2, dim=-1)
kl_sum2 = F.kl_div(logp_x, p_y, reduction='sum')
kl_mean2 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl2 sum is: ", kl_sum2)
print("kl2 mean is: ", kl_mean2)
conv_normal3 = feature_encoder(Variable(samples).cuda(GPU).float())[1][2]
conv_noise3 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][2]
logp_x = F.log_softmax(conv_normal3, dim=-1)
p_y = F.softmax(conv_noise3, dim=-1)
kl_sum3 = F.kl_div(logp_x, p_y, reduction='sum')
kl_mean3 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl3 sum is: ", kl_sum3)
print("kl3 mean is: ", kl_mean3)
conv_normal4 = feature_encoder(Variable(samples).cuda(GPU).float())[1][3]
conv_noise4 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][3]
logp_x = F.log_softmax(conv_normal4, dim=-1)
p_y = F.softmax(conv_noise4, dim=-1)
kl_mean4 = F.kl_div(logp_x, p_y, reduction='mean')
print("kl4 mean is: ", kl_mean4)
# conv_normal5 = feature_encoder(Variable(samples).cuda(GPU).float())[1][4]
# conv_noise5 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][4]
# logp_x = F.log_softmax(conv_normal5, dim=-1)
# p_y = F.softmax(conv_noise5, dim=-1)
# kl_mean5 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl5 mean is: ", kl_mean5)
#
# conv_normal6 = feature_encoder(Variable(samples).cuda(GPU).float())[1][5]
# conv_noise6 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][5]
# logp_x = F.log_softmax(conv_normal6, dim=-1)
# p_y = F.softmax(conv_noise6, dim=-1)
# kl_mean6 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl6 mean is: ", kl_mean6)
#
# conv_normal7 = feature_encoder(Variable(samples).cuda(GPU).float())[1][6]
# conv_noise7 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][6]
# logp_x = F.log_softmax(conv_normal7, dim=-1)
# p_y = F.softmax(conv_noise7, dim=-1)
# kl_mean7 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl7 mean is: ", kl_mean7)
#
# conv_normal8 = feature_encoder(Variable(samples).cuda(GPU).float())[1][7]
# conv_noise8 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][7]
# logp_x = F.log_softmax(conv_normal8, dim=-1)
# p_y = F.softmax(conv_noise8, dim=-1)
# kl_mean8 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl8 mean is: ", kl_mean8)
#
# conv_normal9 = feature_encoder(Variable(samples).cuda(GPU).float())[1][8]
# conv_noise9 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][8]
# logp_x = F.log_softmax(conv_normal9, dim=-1)
# p_y = F.softmax(conv_noise9, dim=-1)
# kl_mean9 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl9 mean is: ", kl_mean9)
#
# conv_normal10 = feature_encoder(Variable(samples).cuda(GPU).float())[1][9]
# conv_noise10 = feature_encoder(Variable(samples2).cuda(GPU).float())[1][9]
# logp_x = F.log_softmax(conv_normal10, dim=-1)
# p_y = F.softmax(conv_noise10, dim=-1)
# kl_mean10 = F.kl_div(logp_x, p_y, reduction='mean')
# print("kl10 mean is: ", kl_mean10)
# calculate relations
# each batch sample link to every samples to calculate relations
# to form a 100x128 matrix for relation network
if MODELTYPE == '1d':
sample_features_ext = sample_features.unsqueeze(0).repeat(BATCH_NUM_PER_CLASS * CLASS_NUM, 1, 1, 1)
batch_features_ext = batch_features.unsqueeze(0).repeat(CLASS_NUM, 1, 1, 1)
batch_features_ext = torch.transpose(batch_features_ext, 0, 1)
relation_pairs = torch.cat((sample_features_ext, batch_features_ext), 2).view(-1, FEATURE_DIM * 2, 5 * 5)
else:
sample_features_ext = sample_features.unsqueeze(0).repeat(BATCH_NUM_PER_CLASS * CLASS_NUM, 1, 1, 1, 1)
batch_features_ext = batch_features.unsqueeze(0).repeat(CLASS_NUM, 1, 1, 1, 1)
batch_features_ext = torch.transpose(batch_features_ext, 0, 1)
relation_pairs = torch.cat((sample_features_ext, batch_features_ext), 2).view(-1, FEATURE_DIM * 2, 5, 5)
relations = relation_network(relation_pairs).view(-1, CLASS_NUM)
mse = nn.MSELoss().cuda(GPU)
one_hot_labels = Variable(
torch.zeros(BATCH_NUM_PER_CLASS * CLASS_NUM, CLASS_NUM).scatter_(1, batch_labels.cuda().long().view(-1, 1),
1)).cuda(GPU)
loss = mse(relations, one_hot_labels)
# training
feature_encoder.zero_grad()
relation_network.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(feature_encoder.parameters(), 0.5)
torch.nn.utils.clip_grad_norm_(relation_network.parameters(), 0.5)
feature_encoder_optim.step()
relation_network_optim.step()
logging.info("episode:" str(episode 1) " loss: " str(loss.item()))
losses.append(loss.item())
lepochs.append(episode)
if (episode 1) % testgap == 0:
# test
logging.info("Testing...")
total_rewards = 0
for i in range(TEST_EPISODE):
task = tg.puTask(metatest_character_folders, CLASS_NUM, SAMPLE_NUM_PER_CLASS, SAMPLE_NUM_PER_CLASS, )
sample_dataloader = tg.get_data_loader(task, num_per_class=SAMPLE_NUM_PER_CLASS, split="train",
shuffle=False, dt=DATATYPE, mt=MODELTYPE,snr = SNR)
test_dataloader = tg.get_data_loader(task, num_per_class=SAMPLE_NUM_PER_CLASS, split="test",
shuffle=True, dt=DATATYPE, mt=MODELTYPE,snr = SNR)
sample_images, sample_labels = sample_dataloader.__iter__().next()
test_images, test_labels = test_dataloader.__iter__().next()
# calculate features
sample_features = feature_encoder(Variable(sample_images).cuda(GPU).float())[0] # 5x64
if MODELTYPE == '1d':
sample_features = sample_features.view(CLASS_NUM, SAMPLE_NUM_PER_CLASS, FEATURE_DIM, 5 * 5)
else:
sample_features = sample_features.view(CLASS_NUM, SAMPLE_NUM_PER_CLASS, FEATURE_DIM, 5, 5)
sample_features = torch.mean(sample_features, 1).squeeze(1)
test_features = feature_encoder(Variable(test_images).cuda(GPU).float())[0] # 20x64
# calculate relations
# each batch sample link to every samples to calculate relations
if MODELTYPE == '1d':
sample_features_ext = sample_features.unsqueeze(0).repeat(SAMPLE_NUM_PER_CLASS * CLASS_NUM, 1, 1, 1)
test_features_ext = test_features.unsqueeze(0).repeat(CLASS_NUM, 1, 1, 1)
test_features_ext = torch.transpose(test_features_ext, 0, 1)
relation_pairs = torch.cat((sample_features_ext, test_features_ext), 2).view(-1, FEATURE_DIM * 2,
5 * 5)
else:
sample_features_ext = sample_features.unsqueeze(0).repeat(SAMPLE_NUM_PER_CLASS * CLASS_NUM, 1, 1, 1,
1)
test_features_ext = test_features.unsqueeze(0).repeat(CLASS_NUM, 1, 1, 1, 1)
test_features_ext = torch.transpose(test_features_ext, 0, 1)
relation_pairs = torch.cat((sample_features_ext, test_features_ext), 2).view(-1, FEATURE_DIM * 2, 5,
5)
relations = relation_network(relation_pairs).view(-1, CLASS_NUM)
_, predict_labels = torch.max(relations.data, 1)
predict_labels = predict_labels.int()
rewards = [1 if predict_labels[j] == test_labels[j] else 0 for j in
range(CLASS_NUM * SAMPLE_NUM_PER_CLASS)]
total_rewards = np.sum(rewards)
test_accuracy = total_rewards / 1.0 / CLASS_NUM / SAMPLE_NUM_PER_CLASS / TEST_EPISODE
logging.info("test accuracy:" str(test_accuracy))
if episode 1 > 80:
finalsum = test_accuracy
accuracys.append(test_accuracy)
aepochs.append(episode)
if test_accuracy > last_accuracy:
# save networks
torch.save(feature_encoder.state_dict(), os.path.join(path, "feature_encoder_%f.pkl" % test_accuracy))
torch.save(relation_network.state_dict(), os.path.join(path, "relation_network_%f.pkl" % test_accuracy))
logging.info("save networks for episode:" str(episode))
last_accuracy = test_accuracy
torch.save(feature_encoder.state_dict(), os.path.join(path, "feature_encoder_final.pkl"))
torch.save(relation_network.state_dict(), os.path.join(path, "relation_network_final.pkl"))
logging.info("final accuracy :" str(test_accuracy))
logging.info("final finalsum :" str(finalsum/4))
plt.figure()
plt.suptitle(path1)
plt.subplot(2, 1, 1)
plt.plot(aepochs, accuracys)
plt.title('test accuracy')
plt.subplot(2, 1, 2)
plt.plot(lepochs, losses)
plt.title('train loss')
plt.savefig('%s\\accuracy.png' % path)
output = open('%s\\accuracy.pkl' % path, 'wb')
pickle.dump((aepochs, accuracys, lepochs, losses), output)
output.close()
if __name__ == '__main__':
main()