-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
288 lines (196 loc) · 9.59 KB
/
train.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
import csv
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import torch.optim as optim
import numpy as np
from tqdm import tqdm
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1 import ImageGrid
class LSLoss(nn.Module):
def __init__(self):
super(LSLoss, self).__init__()
def forward(self, output, labels):
return 0.5 * torch.mean((1 - labels) * (output - 0) ** 2 + labels * (output - 1) ** 2)
class JSD(nn.Module):
def __init__(self):
super(JSD, self).__init__()
def forward(self, net_1_logits, net_2_logits):
net_1_probs = F.softmax(net_1_logits, dim=1)
net_2_probs= F.softmax(net_2_logits, dim=1)
m = 0.5 * (net_1_probs + net_1_probs)
loss = 0.0
loss += F.kl_div(F.log_softmax(net_1_logits, dim=1), m, reduction='batchmean')
loss += F.kl_div(F.log_softmax(net_2_logits, dim=1), m, reduction='batchmean')
return (0.5 * loss)
class Trainer:
def __init__(self, params, setting):
self.device = params['device']
print(self.device)
self.train_loader = params['train_loader']
self.val_loader = params['val_loader']
self.models = params['models']
self.optimizers = params['optimizers']
self.lambdas = params['lambdas'].to(self.device)
self.image_saver = params['image_saver']
self.setting = setting
self.all_cA = []
self.all_cB = []
self.all_dA = []
self.all_dB = []
def disc_loss(self, disc_key, X, X_fake, k=1):
input = torch.concat((X, X_fake), dim=0)
labels = torch.Tensor([1] * X.shape[0] + [0] * X_fake.shape[0]).to(self.device)
flip_labels = torch.Tensor([0] * X.shape[0] + [1] * X_fake.shape[0]).to(self.device)
criterion = LSLoss().to(self.device) #nn.BCELoss().to(self.device)
for i in range(k):
self.optimizers[disc_key].zero_grad()
output = self.models[disc_key](input.detach())
d_loss = criterion(output, labels)
d_loss.backward()
self.optimizers[disc_key].step()
output = self.models[disc_key](input)
g_loss = criterion(output, flip_labels)
return g_loss
def jsd(self, P1, P2):
P1, P2 = P1.softmax(dim=1), P2.softmax(dim=1)
P_mix = (P1 + P2) / 2
KL1 = torch.sum(P1 * torch.log(P1 / P_mix), dim=1)
KL2 = torch.sum(P2 * torch.log(P2 / P_mix), dim=1)
return torch.mean(KL1 + KL2)
def class_loss(self, X, Y_fake, labels=None):
input = torch.concat((X, Y_fake), dim=0)
output = self.models['c'](input)
P1, P2 = output[:X.shape[0],:], output[X.shape[0]:, :]
jsd_loss = JSD()(P1, P2)
criterion = nn.CrossEntropyLoss().to(self.device)
c_loss = torch.Tensor([0]).to(self.device) if labels is None else criterion(P1, labels)
return c_loss, jsd_loss
def rec_loss(self, X, X_fake):
return torch.mean((X - X_fake) ** 2)
def train_loop(self, X, Y, g_XY_key, g_YX_key, d_Y_key, labels=None):
for model in self.models.values():
model.train()
Y_fake = self.models[g_XY_key](X)
g_loss = self.disc_loss(d_Y_key, Y, Y_fake)
c_loss, jsd_loss = self.class_loss(X, Y_fake, labels=labels)
X_fake = self.models[g_YX_key](Y_fake)
r_loss = self.rec_loss(X, X_fake)
losses = torch.concat([loss.reshape(1,-1) for loss in [c_loss, jsd_loss, g_loss, r_loss]], dim=1)
return losses
def train_epoch(self, epoch):
all_losses = None
for (A, labels), (B, _) in tqdm(zip(self.train_loader['A'], self.train_loader['B']), total=len(self.train_loader['A'])):
A, B = A.to(self.device), B.to(self.device)
labels = labels.to(self.device)
for optimizer in self.optimizers.values():
optimizer.zero_grad()
losses = self.train_loop(A, B, 'g_AB', 'g_BA', 'd_B', labels=labels)
losses += self.train_loop(B, A, 'g_BA', 'g_AB', 'd_A', labels=None)
total_loss = torch.sum(losses * self.lambdas)
total_loss.backward()
for name, optimizer in self.optimizers.items():
if name not in ['d_A', 'd_B']:
optimizer.step()
all_losses = losses if all_losses is None else torch.concat((all_losses, losses), dim=0)
avg_losses = torch.mean(all_losses, dim=0)
print('----- Epoch', epoch, '-----')
print('Classifier loss:', float(avg_losses[0]))
print('JSD loss:', float(avg_losses[1]))
print('Generator loss:', float(avg_losses[2]))
print('Reconstruction loss:', float(avg_losses[3]))
print('Total loss:', float(torch.sum(avg_losses * self.lambdas)))
print('------------------------')
print()
def evaluate(self, epoch):
for model in self.models.values():
model.eval()
class_preds = { 'A' : [], 'B' : [] }
class_labels = { 'A' : [], 'B' : [] }
disc_preds = { 'A' : [], 'B' : [] }
disc_labels = { 'A' : [], 'B' : [] }
with torch.no_grad():
for i, ((A, A_labels), (B, B_labels)) in enumerate(zip(self.val_loader['A'], self.val_loader['B'])):
A, B = A.to(self.device), B.to(self.device)
class_preds['A'] += self.models['c'](A).cpu().detach().argmax(dim=1).tolist()
class_labels['A'] += A_labels.tolist()
class_preds['B'] += self.models['c'](B).cpu().detach().argmax(dim=1).tolist()
class_labels['B'] += B_labels.tolist()
B_fake = self.models['g_AB'](A)
B_all = torch.concat((B, B_fake), dim=0)
disc_preds['B'] += self.models['d_B'](B_all).cpu().round().int().tolist()
disc_labels['B'] += [1] * B.shape[0] + [0] * B_fake.shape[0]
A_fake = self.models['g_BA'](B)
A_all = torch.concat((A, A_fake), dim=0)
disc_preds['A'] += self.models['d_A'](A_all).cpu().round().int().tolist()
disc_labels['A'] += [1] * A.shape[0] + [0] * A_fake.shape[0]
A_rec = self.models['g_BA'](B_fake)
B_rec = self.models['g_AB'](A_fake)
if i == 0 and self.image_saver is not None:
A1, B1 = A, B
#self.image_saver(A.cpu(), B.cpu(), A_fake.cpu(), B_fake.cpu(),
#A_rec.cpu(), B_rec.cpu(), epoch, self.setting)
advA, advB = self.explainability(A1, B1)
fig = plt.figure(figsize=(4,8))
grid = ImageGrid(fig, 111, nrows_ncols=(4,2), axes_pad=0.1)
for j in range(2):
grid[2*j].imshow(A1[j,:].cpu().reshape(28,28), cmap='gray')
grid[2*j+4].imshow(B1[j,:].cpu().reshape(28,28), cmap='gray')
grid[2*j+1].imshow(advA[j,:].cpu().permute(1,2,0))
grid[2*j+5].imshow(advB[j,:].cpu().permute(1,2,0))
plt.title('Epoch ' + str(epoch))
plt.savefig('results/data/test/epoch-' + str(epoch) + '.png')
plt.clf()
plt.close()
self.all_cA.append(np.mean(np.array(class_preds['A']) == np.array(class_labels['A'])))
self.all_cB.append(np.mean(np.array(class_preds['B']) == np.array(class_labels['B'])))
self.all_dA.append(np.mean(np.array(disc_preds['A']) == np.array(disc_labels['A'])))
self.all_dB.append(np.mean(np.array(disc_preds['B']) == np.array(disc_labels['B'])))
print('===== Classification =====')
print('(A) acc:', self.all_cA[-1])
print('(B) acc:', self.all_cB[-1])
print('==========================')
print()
print('===== Discrimination =====')
print('(A) acc:', self.all_dA[-1])
print('(B) acc:', self.all_dB[-1])
print('==========================')
print()
def to_advX(self, X, X_eps, eps=50):
X_abs_eps = torch.abs(X_eps)
advX = X - eps * (X_abs_eps / X_abs_eps.sum(dim=2).sum(dim=2).reshape(-1,1,1,1))
return advX.clip(min=0,max=1)
def explainability(self, A, B, eta=1e10, n_iters=1000):
X = torch.concat((A, B), dim=0)
X_eps = Variable(torch.randn(X.shape).to(self.device), requires_grad=True)
optimizer = optim.SGD([X_eps], lr=eta)
for i in range(n_iters):
optimizer.zero_grad()
advX = self.to_advX(X, X_eps)
output = F.softmax(self.models['c'](advX), dim=1)
entropy = torch.sum(output * torch.log(output), dim=1).mean()
entropy.backward(retain_graph=True)
if i == n_iters - 1:
print('Entropy:', float(entropy))
optimizer.step()
advX = self.to_advX(X, X_eps)
advX = self.adv_images(X, advX)
advA, advB = advX[:A.shape[0], :, :, :], advX[A.shape[0]:, :, :, :]
return advA.detach(), advB.detach()
def adv_images(self, X, advX):
images = torch.zeros((X.shape[0], 3, X.shape[2], X.shape[3])).to(self.device)
images[:,:,:,:] = X
images[:,1:,:,:] -= X - advX
return images
def train(self, n_epochs=500, epoch_step=1):
for epoch in range(1, n_epochs+1):
self.train_epoch(epoch)
if epoch % epoch_step == 0:
self.evaluate(epoch)
def write_results(self):
with open('results/metrics.csv', 'a') as f:
writer = csv.writer(f)
writer.writerow([self.setting, self.all_cA[-1], self.all_cB[-1],
self.all_dA[-1], self.all_dB[-1], self.all_cA,
self.all_cB, self.all_dA, self.all_dB])