-
Notifications
You must be signed in to change notification settings - Fork 1
/
gan.py
448 lines (374 loc) · 21.8 KB
/
gan.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
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
import sys, os, inspect
import numpy as np
import random
import math
from tqdm import tqdm
import torch
from torch import nn
from torch.utils.data import Dataset, DataLoader
from torch.optim.lr_scheduler import ReduceLROnPlateau
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
from models.complex_transformer import ComplexTransformer
from models.FNNLinear import FNNLinear
from models.FNNSeparated import FNNSeparated
from models.GAN import Generator, Discriminator
from utils import *
from data_utils import *
import argparse
from torch.autograd import Variable
from torch.autograd import grad as torch_grad
###############################################################################
# Function Definition #
###############################################################################
def _gradient_penalty(real_data, generated_data, DNet, mask, num_class, device, args, local=True):
batch_size = real_data.size()[0]
# Calculate interpolation
alpha = torch.rand(batch_size, 1)
alpha = alpha.expand_as(real_data)
alpha = alpha.to(device)
interpolated = alpha * real_data.data + (1 - alpha) * generated_data.data
interpolated = Variable(interpolated, requires_grad=True)
interpolated = interpolated.to(device)
# Calculate probability of interpolated examples
prob_interpolated = DNet(interpolated, mask) # mask = 1 if local=False
# Calculate gradients of probabilities with respect to examples
gradients = torch_grad(outputs=prob_interpolated, inputs=interpolated,
grad_outputs=torch.ones(prob_interpolated.size()).to(device),
create_graph=True, retain_graph=True)[0]
# Gradients have shape (batch_size, num_channels, img_width, img_height),
# so flatten to easily take norm per example in batch
gradients = gradients.view(batch_size, -1)
# Derivatives of the gradient close to 0 can cause problems because of
# the square root, so manually calculate norm and add epsilon
if local:
gradients_norm = torch.sqrt(torch.sum(gradients ** 2, dim=1, keepdim=True) + 1e-12)
gradients_norm = gradients_norm * mask
return args.gpweight * ((gradients_norm - 1) ** 2).mean(dim=0)
else:
gradients_norm = torch.sqrt(torch.sum(gradients ** 2, dim=1) + 1e-12)
return args.gpweight * ((gradients_norm - 1) ** 2).mean()
def train_classifier(labeled_dataloader, CNet, encoder, encoder_MLP, GNet, optimizerCNet, optimizerG, optimizerEncoder, optimizerEncoderMLP, args, target):
acc = 0.0
num_datas = 0.0
CNet.train()
encoder.train()
encoder_MLP.train()
if target:
GNet.train()
# Loop over mini batch for traininig
for batch_id, (x, y) in tqdm(enumerate(labeled_dataloader), total=len(labeled_dataloader)):
# clear optimizer
optimizerCNet.zero_grad()
optimizerG.zero_grad()
optimizerEncoder.zero_grad()
optimizerEncoderMLP.zero_grad()
# load data
x = x.to(device).float()
y = y.to(device)
num_datas += x.size(0)
# encoder inference
x_embedding = encoder_inference(encoder, encoder_MLP, x)
if target:
fake_source_embedding = GNet(x_embedding)
pred = CNet(fake_source_embedding)
else:
pred = CNet(x_embedding)
# acc and loss
acc += (pred.argmax(-1) == y).sum().item()
loss = criterion_classifier(pred, y)
loss.backward()
# update CNet, GNet, Encoder, EncoderMLP
optimizerCNet.step()
optimizerG.step()
optimizerEncoder.step()
optimizerEncoderMLP.step()
acc = acc / num_datas
return acc, loss.item()
def eval_classifier(unlabeled_dataloader, CNet, encoder, encoder_MLP, GNet, target=False):
acc_unlabel = 0.0
num_datas = 0.0
CNet.eval()
encoder.eval()
encoder_MLP.eval()
if target:
GNet.eval()
for batch_id, (x, y) in tqdm(enumerate(unlabeled_dataloader), total=len(unlabeled_dataloader)):
x = x.to(device).float()
y = y.to(device)
num_datas += x.shape[0]
x_embedding = encoder_inference(encoder, encoder_MLP, x)
if target:
fake_source_embedding = GNet(x_embedding)
pred = CNet(fake_source_embedding)
else:
pred = CNet(x_embedding)
acc_unlabel += (pred.argmax(-1) == y).sum().item()
acc_unlabel = acc_unlabel/num_datas
return acc_unlabel
def update_D(source_embedding, fake_source_embedding, DNet, optimizerD, source_mask, target_mask, source_weight_count, target_weight_count, target_weight, num_class, total_error_D, device, args, local=True):
optimizerD.zero_grad()
# adversarial loss
if local:
source_DNet_local = DNet(source_embedding, source_mask)
target_DNet_local = DNet(fake_source_embedding, target_mask)
source_DNet_local_mean = source_DNet_local.sum(dim=0) / source_weight_count
target_DNet_local_mean = (target_DNet_local * target_weight).sum(dim=0) / target_weight_count
# gradient penalty
gp = _gradient_penalty(source_embedding, fake_source_embedding, DNet, source_mask, num_class, device, args)
# loss
loss_D_local = (target_DNet_local_mean - source_DNet_local_mean + gp).sum()
loss_D = loss_D_local * args.dlocal
else:
# loss and gradient penalty
loss_D_global = DNet_global(fake_source_embedding,1).mean() - DNet_global(source_embedding,1).mean()
gradient_penalty = _gradient_penalty(source_embedding, fake_source_embedding, DNet_global, 1, num_class, device, args)
loss_D_global = loss_D_global + gradient_penalty
loss_D = loss_D_global * args.dglobal
# accumulate loss for report
total_error_D += loss_D.item()
# backward and update D
loss_D.backward()
optimizerD.step()
return total_error_D
def update_G(fake_source_embedding, DNet, optimizerG, total_error_G, target_mask, target_weight, target_weight_count, args, local=True):
# clear optimizer
optimizerG.zero_grad()
# calculte loss for Generator
if local:
target_DNet_local = DNet(fake_source_embedding, target_mask)
target_DNet_local_mean = (target_DNet_local * target_weight).sum(dim=0) / target_weight_count
loss_G = -target_DNet_local_mean.sum()
loss_G = loss_G * args.dlocal
else:
loss_G = -DNet(fake_source_embedding,1).mean() * args.dglobal
# accumulate loss for report
total_error_G += loss_G.item()
# backward and update G
loss_G.backward()
optimizerG.step()
return total_error_G
###############################################################################
# parameters #
###############################################################################
"""
Default contains the optimal hyperparameters for conditional (local) GAN (target and source lbl 70%). Use --isglobal and --islocal to choose global or conditional GAN for training. Except for the parameter mentioned in example below, the other parameters between local and conditional gan are shared.
Example of command to run:
$ python gan.py --islocal 1 --n_critic 6 --gpweight 10
$ python gan.py --isglobal 1 --n_critic 12 --gpweight 5
"""
parser = argparse.ArgumentParser(description='WGAN for time series domain adaptation')
parser.add_argument("--data_path", type=str, default="../data_unzip/", help="dataset path")
parser.add_argument("--task", type=str, default='3E', help='3A or 3E')
parser.add_argument('--batch_size', type=int, default=256, help='batch size')
parser.add_argument('--epochs', type=int, default=2000, help='number of epochs')
parser.add_argument('--lr_gan', type=float, default=1e-3, help='learning rate for adversarial')
parser.add_argument('--lr_CNet', type=float, default=1e-3, help='learning rate for classification')
parser.add_argument('--lr_encoder', type=float, default=1e-3, help='learning rate for classification')
parser.add_argument('--n_critic', type=int, default=6, help='gap: Generator train GAP times, discriminator train once')
parser.add_argument('--target_lbl_percentage', type=float, default=0.7, help='percentage of target labeled data')
parser.add_argument('--source_lbl_percentage', type=float, default=0.7, help='percentage of source labeled data')
parser.add_argument('--num_per_class', type=int, default=3, help='Number of sample per class inside each batch when training local discriminator')
parser.add_argument('--save_path', type=str, default="../train_related", help='where to store data')
parser.add_argument('--model_save_period', type=int, default=2, help='period in which the model is saved')
parser.add_argument('--gpweight', type=float, default=10, help='clip_value for WGAN')
parser.add_argument('--sclass', type=float, default=0.7, help='source domain classification weight on loss function')
parser.add_argument('--dlocal', type=float, default=0.01, help='local GAN weight on loss function')
parser.add_argument('--dglobal', type=float, default=0.01, help='global GAN weight on loss function')
parser.add_argument('--isglobal', type=int, default=0, help='if using global DNet')
parser.add_argument('--islocal', type=int, default=0, help='if using local (conditional) DNet')
args = parser.parse_args()
args.isglobal = True if args.isglobal == 1 else False
args.islocal = True if args.islocal == 1 else False
# Check at least one mode for GAN is true
assert args.islocal or args.isglobal, "Please at least use one kind of GAN for training, otherwise please refer to 'naive_apdation.py'."
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
args.task = '3Av2' if args.task == '3A' else '3E'
num_class = 50 if args.task == "3Av2" else 65
###############################################################################
# Save Path #
###############################################################################
model_sub_folder = 'results/GAN_adaptation/lbl_percent_%f/task_%s_gpweight_%f_ncritic_%i_sclass_%f_global_%i_dglobal_%f_local_%i_dlocal_%f'%(args.target_lbl_percentage, args.task, args.gpweight, args.n_critic, args.sclass, args.isglobal, args.dglobal, args.islocal, args.dlocal)
save_folder = os.path.abspath(os.path.join(args.save_path, model_sub_folder))
if not os.path.exists(save_folder):
os.makedirs(save_folder)
###############################################################################
# Logger #
###############################################################################
logger = get_logger(save_folder, device, args)
###############################################################################
# Data Loading #
###############################################################################
# get target dataset and dataloader
labeled_target_x, labeled_target_y, unlabeled_target_x, unlabeled_target_y = read_data(args.task, 'target', args.target_lbl_percentage, args.data_path)
labeled_target_dataset = SingleDataset(labeled_target_x, labeled_target_y)
unlabeled_target_dataset = SingleDataset(unlabeled_target_x, unlabeled_target_y)
labeled_target_dataloader = DataLoader(labeled_target_dataset, batch_size=args.batch_size, shuffle=True, pin_memory=True, num_workers=4)
unlabeled_target_dataloader = DataLoader(unlabeled_target_dataset, batch_size=args.batch_size, shuffle=False, pin_memory=True, num_workers=4)
# get source dataset and dataloader
labeled_source_x, labeled_source_y, unlabeled_source_x, unlabeled_source_y = read_data(args.task, 'source', args.source_lbl_percentage, args.data_path)
labeled_source_dataset = SingleDataset(labeled_source_x, labeled_source_y)
unlabled_source_dataset = SingleDataset(unlabeled_source_x, unlabeled_source_y)
labeled_source_dataloader = DataLoader(labeled_source_dataset, batch_size=args.batch_size, shuffle=True, pin_memory=True, num_workers=4)
unlabeled_source_dataloader = DataLoader(unlabled_source_dataset, batch_size=args.batch_size, shuffle=False, pin_memory=True, num_workers=4)
# create join dataset and dataloader
join_dataset = JoinDataset(labeled_source_x, labeled_source_y, labeled_target_x, labeled_target_y, random=True)
join_dataloader = DataLoader(join_dataset, batch_size=args.batch_size, shuffle=True, pin_memory=True, num_workers=4)
# create class data dict for local Discrimitor training
source_labeled_dict = get_class_data_dict(labeled_source_x, labeled_source_y, num_class)
target_labeled_dict = get_class_data_dict(labeled_target_x, labeled_target_y, num_class)
###############################################################################
# Model Creation #
###############################################################################
seq_len = 10
feature_dim = 160
# model
encoder = ComplexTransformer(layers=3,
time_step=seq_len,
input_dim=feature_dim,
hidden_size=64,
output_dim=64,
num_heads=8,
out_dropout=0.2,
leaky_slope=0.2)
encoder.to(device)
encoder_MLP = FNNSeparated(d_in=64 * 2 * 1, d_h1=64*4, d_h2=64*2, dp=0.2).to(device)
CNet = FNNLinear(d_h2=64*2, d_out=num_class).to(device)
if args.isglobal:
DNet_global = Discriminator(feature_dim=64*2, d_out=1).to(device)
if args.islocal:
DNet_local = Discriminator(feature_dim=64*2, d_out=num_class).to(device)
GNet = Generator(dim=64*2).to(device)
# loss
criterion_classifier = nn.CrossEntropyLoss().to(device)
# optimizer
if args.isglobal:
optimizerD_global = torch.optim.Adam(DNet_global.parameters(), lr=args.lr_gan)
if args.islocal:
optimizerD_local = torch.optim.Adam(DNet_local.parameters(), lr=args.lr_gan)
optimizerG = torch.optim.Adam(GNet.parameters(), lr=args.lr_gan)
optimizerCNet = torch.optim.Adam(CNet.parameters(), lr=args.lr_CNet)
optimizerEncoder = torch.optim.Adam(encoder.parameters(), lr=args.lr_encoder)
optimizerEncoderMLP = torch.optim.Adam(encoder_MLP.parameters(), lr=args.lr_encoder)
###############################################################################
# Model Initilize #
###############################################################################
if args.isglobal:
DNet_global.apply(weights_init)
if args.islocal:
DNet_local.apply(weights_init)
GNet.apply(weights_init)
encoder.apply(weights_init)
encoder_MLP.apply(weights_init)
CNet.apply(weights_init)
###############################################################################
# Train #
###############################################################################
target_acc_label_ = []
source_acc_label_ = []
source_acc_unlabel_ = []
target_acc_unlabel_ = []
error_D_global = []
error_G_global = []
error_D_local = []
error_G_local = []
print('Started training')
for epoch in range(args.epochs):
# Train classifier
# train on source data
source_acc, _ = train_classifier(labeled_source_dataloader, CNet, encoder, encoder_MLP, None, optimizerCNet, optimizerG, optimizerEncoder, optimizerEncoderMLP, args, target=False)
source_acc_label_.append(source_acc)
# train on target data
target_acc, _ = train_classifier(labeled_target_dataloader, CNet, encoder, encoder_MLP, GNet, optimizerCNet, optimizerG, optimizerEncoder, optimizerEncoderMLP, args, target=True)
target_acc_label_.append(target_acc)
# Eval classifier
# source domain
source_acc_unlabel = eval_classifier(unlabeled_source_dataloader, CNet, encoder, encoder_MLP, GNet, target=False)
source_acc_unlabel_.append(source_acc_unlabel)
# target domain
target_acc_unlabel = eval_classifier(unlabeled_target_dataloader, CNet, encoder, encoder_MLP, GNet, target=True)
target_acc_unlabel_.append(target_acc_unlabel)
logger.info('Epoch: %i, update classifier: source acc: %f; source unlbl acc: %f; target acc: %f; target unlabel acc: %f'%(epoch+1, source_acc, source_acc_unlabel, target_acc, target_acc_unlabel))
# Train GAN
if args.isglobal:
# Train Global GAN
CNet.eval()
encoder.eval()
encoder_MLP.eval()
GNet.train()
DNet_global.train()
total_error_D_global = 0
total_error_G = 0
for batch_id, ((source_x, source_y), (target_x, target_y)) in tqdm(enumerate(join_dataloader), total=len(join_dataloader)):
# get embeddings
source_data = source_x.to(device).float()
source_embedding = encoder_inference(encoder, encoder_MLP, source_data)
target_data = target_x.to(device).float()
target_embedding = encoder_inference(encoder, encoder_MLP, target_data)
fake_source_embedding = GNet(target_embedding).detach()
# Update G Network
total_error_G = update_G(fake_source_embedding, DNet_global, optimizerG, total_error_G, None, None, None, args, local=False)
# Update D_global Network
if batch_id % args.n_critic == 0:
total_error_D_global = update_D(source_embedding, fake_source_embedding, DNet_global, optimizerD_global, None, None, None, None, None, num_class, total_error_D_global, device, args, local=False)
logger.info('Epoch: %i, Global Discrimator Updates: Loss D_global: %f, Loss G: %f; Update_ratio (G/D): %i'%(epoch+1, total_error_D_global, total_error_G, args.n_critic))
# record loss
error_D_global.append(total_error_D_global)
error_G_global.append(total_error_G)
if args.islocal:
# Update local Discriminator
total_error_D_local = 0
total_error_G = 0
CNet.eval()
encoder.eval()
encoder_MLP.eval()
GNet.train()
DNet_local.train()
for batch_id in tqdm(range(math.ceil(labeled_target_x.shape[0]/args.batch_size))):
# load batch data
target_x, target_y, target_weight = get_batch_target_data_on_class(target_labeled_dict, args.num_per_class, None, no_pesudo=True)
source_x, source_y = get_batch_source_data_on_class(source_labeled_dict, args.num_per_class)
source_x = torch.Tensor(source_x).to(device).float()
target_x = torch.Tensor(target_x).to(device).float()
source_y = torch.LongTensor(target_y).to(device)
target_y = torch.LongTensor(target_y).to(device)
target_weight = torch.Tensor(target_weight).to(device)
# create mask
source_mask = torch.zeros(source_x.size(0), num_class).to(device).scatter_(1, source_y.unsqueeze(-1), 1)
target_mask = torch.zeros(target_x.size(0), num_class).to(device).scatter_(1, target_y.unsqueeze(-1), 1)
target_weight = torch.zeros(target_x.size(0), num_class).to(device).scatter_(1, target_y.unsqueeze(-1), target_weight.unsqueeze(-1))
source_weight_count = source_mask.sum(dim=0)
target_weight_count = target_weight.sum(dim=0)
# get embeddings from encoder
source_embedding = encoder_inference(encoder, encoder_MLP, source_x)
target_embedding = encoder_inference(encoder, encoder_MLP, target_x)
fake_source_embedding = GNet(target_embedding).detach()
# Update G Network
total_error_G = update_G(fake_source_embedding, DNet_local, optimizerG, total_error_G, target_mask, target_weight, target_weight_count, args, local=True)
# Update D local Network
if batch_id % args.n_critic == 0:
total_error_D_local = update_D(source_embedding, fake_source_embedding, DNet_local, optimizerD_local, source_mask, target_mask, source_weight_count, target_weight_count, target_weight, num_class, total_error_D_local, device, args, local=True)
logger.info('Epoch: %i, Local Discrimator Updates: Loss D_local: %f, Loss G: %f; Update_ratio (G:D): %i'%(epoch+1, total_error_D_local, total_error_G, args.n_critic))
# record loss
error_D_local.append(total_error_D_local)
error_G_global.append(total_error_G)
# Save models and results
np.save(os.path.join(args.save_path, model_sub_folder, 'target_acc_label_.npy'),target_acc_label_)
np.save(os.path.join(args.save_path, model_sub_folder, 'source_acc_label_.npy'),source_acc_label_)
np.save(os.path.join(args.save_path, model_sub_folder, 'target_acc_unlabel_.npy'),target_acc_unlabel_)
np.save(os.path.join(args.save_path, model_sub_folder, 'source_acc_unlabel_.npy'),source_acc_unlabel_)
if args.isglobal:
np.save(os.path.join(args.save_path, model_sub_folder, 'error_D_global.npy'),error_D_global)
np.save(os.path.join(args.save_path, model_sub_folder, 'error_G_global.npy'),error_G_global)
if args.islocal:
np.save(os.path.join(args.save_path, model_sub_folder, 'error_D_local.npy'),error_D_local)
np.save(os.path.join(args.save_path, model_sub_folder, 'error_G_local.npy'),error_G_local)
if epoch % args.model_save_period == 0:
torch.save(CNet.state_dict(), os.path.join(args.save_path,model_sub_folder, 'CNet_%i.t7'%(epoch+1)))
torch.save(GNet.state_dict(), os.path.join(args.save_path,model_sub_folder, 'GNet_%i.t7'%(epoch+1)))
torch.save(encoder.state_dict(), os.path.join(args.save_path,model_sub_folder,'encoder_%i.t7'%(epoch+1)))
torch.save(encoder_MLP.state_dict(), os.path.join(args.save_path,model_sub_folder,'encoder_MLP_%i.t7'%(epoch+1)))
if args.isglobal:
torch.save(DNet_global.state_dict(), os.path.join(args.save_path, model_sub_folder, 'DNet_global_%i.t7'%(epoch+1)))
if args.islocal:
torch.save(DNet_local.state_dict(), os.path.join(args.save_path, model_sub_folder, 'DNet_local_%i.t7'%(epoch+1)))