-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathtrain.py
302 lines (246 loc) · 14.1 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
289
290
291
292
293
294
295
296
297
298
299
300
301
302
import torch.optim as optim
import glob
import json
import argparse
import os
import random
import math
from time import gmtime, strftime
from models import *
from dataset_GBU import FeatDataLayer, DATA_LOADER
from utils import *
from sklearn.metrics.pairwise import cosine_similarity
import torch.backends.cudnn as cudnn
import classifier
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', default='SUN',help='dataset: CUB, AWA2, APY, FLO, SUN')
parser.add_argument('--dataroot', default='./SDGZSL_data', help='path to dataset')
parser.add_argument('--workers', type=int, help='number of data loading workers', default=4)
parser.add_argument('--image_embedding', default='res101', type=str)
parser.add_argument('--class_embedding', default='att', type=str)
parser.add_argument('--gen_nepoch', type=int, default=400, help='number of epochs to train for')
parser.add_argument('--lr', type=float, default=0.0001, help='learning rate to train generater')
parser.add_argument('--zsl', type=bool, default=False, help='Evaluate ZSL or GZSL')
parser.add_argument('--finetune', type=bool, default=False, help='Use fine-tuned feature')
parser.add_argument('--ga', type=float, default=15, help='relationNet weight')
parser.add_argument('--beta', type=float, default=1, help='tc weight')
parser.add_argument('--weight_decay', type=float, default=1e-6, help='weight_decay')
parser.add_argument('--dis', type=float, default=3, help='Discriminator weight')
parser.add_argument('--dis_step', type=float, default=2, help='Discriminator update interval')
parser.add_argument('--kl_warmup', type=float, default=0.01, help='kl warm-up for VAE')
parser.add_argument('--tc_warmup', type=float, default=0.001, help='tc warm-up')
parser.add_argument('--vae_dec_drop', type=float, default=0.5, help='dropout rate in the VAE decoder')
parser.add_argument('--vae_enc_drop', type=float, default=0.4, help='dropout rate in the VAE encoder')
parser.add_argument('--ae_drop', type=float, default=0.2, help='dropout rate in the auto-encoder')
parser.add_argument('--classifier_lr', type=float, default=0.001, help='learning rate to train softmax classifier')
parser.add_argument('--classifier_steps', type=int, default=50, help='training steps of the classifier')
parser.add_argument('--batchsize', type=int, default=64, help='input batch size')
parser.add_argument('--nSample', type=int, default=1200, help='number features to generate per class')
parser.add_argument('--disp_interval', type=int, default=200)
parser.add_argument('--save_interval', type=int, default=10000)
parser.add_argument('--evl_interval', type=int, default=400)
parser.add_argument('--evl_start', type=int, default=0)
parser.add_argument('--manualSeed', type=int, default=5606, help='manual seed')
parser.add_argument('--latent_dim', type=int, default=20, help='dimention of latent z')
parser.add_argument('--q_z_nn_output_dim', type=int, default=128, help='dimention of hidden layer in encoder')
parser.add_argument('--S_dim', type=int, default=1024)
parser.add_argument('--NS_dim', type=int, default=1024)
parser.add_argument('--gpu', default='0', type=str, help='index of GPU to use')
opt = parser.parse_args()
if opt.manualSeed is None:
opt.manualSeed = random.randint(1, 10000)
print("Random Seed: ", opt.manualSeed)
np.random.seed(opt.manualSeed)
random.seed(opt.manualSeed)
torch.manual_seed(opt.manualSeed)
torch.cuda.manual_seed_all(opt.manualSeed)
cudnn.benchmark = True
print('Running parameters:')
print(json.dumps(vars(opt), indent=4, separators=(',', ': ')))
opt.gpu = torch.device("cuda:"+opt.gpu if torch.cuda.is_available() else "cpu")
def train():
dataset = DATA_LOADER(opt)
opt.C_dim = dataset.att_dim
opt.X_dim = dataset.feature_dim
opt.Z_dim = opt.latent_dim
opt.y_dim = dataset.ntrain_class
out_dir = 'out/{}/wd-{}_b-{}_g-{}_lr-{}_sd-{}_dis-{}_nS-{}_nZ-{}_bs-{}'.format(opt.dataset, opt.weight_decay,
opt.beta, opt.ga, opt.lr,
opt.S_dim, opt.dis, opt.nSample, opt.Z_dim, opt.batchsize)
os.makedirs(out_dir, exist_ok=True)
print("The output dictionary is {}".format(out_dir))
log_dir = out_dir + '/log_{}.txt'.format(opt.dataset)
with open(log_dir, 'w') as f:
f.write('Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
dataset.feature_dim = dataset.train_feature.shape[1]
opt.X_dim = dataset.feature_dim
opt.Z_dim = opt.latent_dim
opt.y_dim = dataset.ntrain_class
data_layer = FeatDataLayer(dataset.train_label.numpy(), dataset.train_feature.cpu().numpy(), opt)
opt.niter = int(dataset.ntrain/opt.batchsize) * opt.gen_nepoch
result_gzsl_soft = Result()
result_zsl_soft = Result()
model = VAE(opt).to(opt.gpu)
relationNet = RelationNet(opt).to(opt.gpu)
discriminator = Discriminator(opt).to(opt.gpu)
ae = AE(opt).to(opt.gpu)
print(model)
with open(log_dir, 'a') as f:
f.write('\n')
f.write('Generative Model Training Start:')
f.write(strftime("%a, %d %b %Y %H:%M:%S +0000", gmtime()) + '\n')
start_step = 0
optimizer = optim.Adam(model.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
relation_optimizer = optim.Adam(relationNet.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
dis_optimizer = optim.Adam(discriminator.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
ae_optimizer = optim.Adam(ae.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
ones = torch.ones(opt.batchsize, dtype=torch.long, device=opt.gpu)
zeros = torch.zeros(opt.batchsize, dtype=torch.long, device=opt.gpu)
mse = nn.MSELoss().to(opt.gpu)
iters = math.ceil(dataset.ntrain/opt.batchsize)
beta = 0.01
coin = 0
gamma = 0
for it in range(start_step, opt.niter+1):
if it % iters == 0:
beta = min(opt.kl_warmup*(it/iters), 1)
gamma = min(opt.tc_warmup * (it / iters), 1)
blobs = data_layer.forward()
feat_data = blobs['data']
labels_numpy = blobs['labels'].astype(int)
labels = torch.from_numpy(labels_numpy.astype('int')).to(opt.gpu)
C = np.array([dataset.train_att[i,:] for i in labels])
C = torch.from_numpy(C.astype('float32')).to(opt.gpu)
X = torch.from_numpy(feat_data).to(opt.gpu)
sample_C = torch.from_numpy(np.array([dataset.train_att[i, :] for i in labels.unique()])).to(opt.gpu)
sample_C_n = labels.unique().shape[0]
sample_label = labels.unique().cpu()
x_mean, z_mu, z_var, z = model(X, C)
loss, ce, kl = multinomial_loss_function(x_mean, X, z_mu, z_var, z, beta=beta)
sample_labels = np.array(sample_label)
re_batch_labels = []
for label in labels_numpy:
index = np.argwhere(sample_labels == label)
re_batch_labels.append(index[0][0])
re_batch_labels = torch.LongTensor(re_batch_labels)
one_hot_labels = torch.zeros(opt.batchsize, sample_C_n).scatter_(1, re_batch_labels.view(-1, 1), 1).to(opt.gpu)
# one_hot_labels = torch.tensor(
# torch.zeros(opt.batchsize, sample_C_n).scatter_(1, re_batch_labels.view(-1, 1), 1)).to(opt.gpu)
x1, h1, hs1, hn1 = ae(x_mean)
relations = relationNet(hs1, sample_C)
relations = relations.view(-1, labels.unique().cpu().shape[0])
p_loss = opt.ga * mse(relations, one_hot_labels)
x2, h2, hs2, hn2 = ae(X)
relations = relationNet(hs2, sample_C)
relations = relations.view(-1, labels.unique().cpu().shape[0])
p_loss = p_loss + opt.ga * mse(relations, one_hot_labels)
rec = mse(x1, X) + mse(x2, X)
if coin > 0:
s_score = discriminator(h1)
tc_loss = opt.beta * gamma *((s_score[:, :1] - s_score[:, 1:]).mean())
s_score = discriminator(h2)
tc_loss = tc_loss + opt.beta * gamma* ((s_score[:, :1] - s_score[:, 1:]).mean())
loss = loss + p_loss + rec + tc_loss
coin -= 1
else:
s, n = permute_dims(hs1, hn1)
b = torch.cat((s, n), 1).detach()
s_score = discriminator(h1)
n_score = discriminator(b)
tc_loss = opt.dis * (F.cross_entropy(s_score, zeros) + F.cross_entropy(n_score, ones))
s, n = permute_dims(hs2, hn2)
b = torch.cat((s, n), 1).detach()
s_score = discriminator(h2)
n_score = discriminator(b)
tc_loss = tc_loss + opt.dis * (F.cross_entropy(s_score, zeros) + F.cross_entropy(n_score, ones))
dis_optimizer.zero_grad()
tc_loss.backward(retain_graph=True)
dis_optimizer.step()
loss = loss + p_loss + rec
coin += opt.dis_step
optimizer.zero_grad()
relation_optimizer.zero_grad()
ae_optimizer.zero_grad()
loss.backward()
optimizer.step()
relation_optimizer.step()
ae_optimizer.step()
if it % opt.disp_interval == 0 and it:
log_text = 'Iter-[{}/{}]; loss: {:.3f}; kl:{:.3f}; p_loss:{:.3f}; rec:{:.3f}; tc:{:.3f}; gamma:{:.3f};'.format(it,
opt.niter, loss.item(),kl.item(),p_loss.item(),rec.item(), tc_loss.item(), gamma)
log_print(log_text, log_dir)
if it % opt.evl_interval == 0 and it > opt.evl_start:
model.eval()
ae.eval()
gen_feat, gen_label = synthesize_feature_test(model, ae, dataset, opt)
with torch.no_grad():
train_feature = ae.encoder(dataset.train_feature.to(opt.gpu))[:,:opt.S_dim].cpu()
test_unseen_feature = ae.encoder(dataset.test_unseen_feature.to(opt.gpu))[:,:opt.S_dim].cpu()
test_seen_feature = ae.encoder(dataset.test_seen_feature.to(opt.gpu))[:,:opt.S_dim].cpu()
train_X = torch.cat((train_feature, gen_feat), 0)
train_Y = torch.cat((dataset.train_label, gen_label + dataset.ntrain_class), 0)
if opt.zsl:
"""ZSL"""
cls = classifier.CLASSIFIER(opt, gen_feat, gen_label, dataset, test_seen_feature, test_unseen_feature,
dataset.ntrain_class + dataset.ntest_class, True, opt.classifier_lr, 0.5, 20,
opt.nSample, False)
result_zsl_soft.update(it, cls.acc)
log_print("ZSL Softmax:", log_dir)
log_print("Acc {:.2f}% Best_acc [{:.2f}% | Iter-{}]".format(
cls.acc, result_zsl_soft.best_acc, result_zsl_soft.best_iter), log_dir)
else:
""" GZSL"""
cls = classifier.CLASSIFIER(opt, train_X, train_Y, dataset, test_seen_feature, test_unseen_feature,
dataset.ntrain_class + dataset.ntest_class, True, opt.classifier_lr, 0.5,
opt.classifier_steps, opt.nSample, True)
result_gzsl_soft.update_gzsl(it, cls.acc_seen, cls.acc_unseen, cls.H)
log_print("GZSL Softmax:", log_dir)
log_print("U->T {:.2f}% S->T {:.2f}% H {:.2f}% Best_H [{:.2f}% {:.2f}% {:.2f}% | Iter-{}]".format(
cls.acc_unseen, cls.acc_seen, cls.H, result_gzsl_soft.best_acc_U_T, result_gzsl_soft.best_acc_S_T,
result_gzsl_soft.best_acc, result_gzsl_soft.best_iter), log_dir)
if result_gzsl_soft.save_model:
files2remove = glob.glob(out_dir + '/Best_model_GZSL_*')
for _i in files2remove:
os.remove(_i)
save_model(it, model, opt.manualSeed, log_text,
out_dir + '/Best_model_GZSL_H_{:.2f}_S_{:.2f}_U_{:.2f}.tar'.format(result_gzsl_soft.best_acc,
result_gzsl_soft.best_acc_S_T,
result_gzsl_soft.best_acc_U_T))
###############################################################################################################
# retrieval code
cls_centrild = np.zeros((dataset.ntest_class, opt.S_dim))
for i in range(dataset.ntest_class):
cls_centrild[i] = torch.mean(gen_feat[gen_label == i,], dim=0)
dist = cosine_similarity(cls_centrild, test_unseen_feature)
precision_100 = torch.zeros(dataset.ntest_class)
precision_50 = torch.zeros(dataset.ntest_class)
precision_25 = torch.zeros(dataset.ntest_class)
dist = torch.from_numpy(-dist)
for i in range(dataset.ntest_class):
is_class = dataset.test_unseen_label == i
# print(is_class.sum())
cls_num = int(is_class.sum())
# 100%
_, idx = torch.topk(dist[i, :], cls_num, largest=False)
precision_100[i] = (is_class[idx]).sum().float() / cls_num
# 50%
cls_num_50 = int(cls_num / 2)
_, idx = torch.topk(dist[i, :], cls_num_50, largest=False)
precision_50[i] = (is_class[idx]).sum().float() / cls_num_50
# 25%
cls_num_25 = int(cls_num / 4)
_, idx = torch.topk(dist[i, :], cls_num_25, largest=False)
precision_25[i] = (is_class[idx]).sum().float() / cls_num_25
print("retrieval results 100%%: %.3f 50%%: %.3f 25%%: %.3f" % (precision_100.mean().item(),
precision_50.mean().item(),
precision_25.mean().item()))
###############################################################################################################
model.train()
ae.train()
if it % opt.save_interval == 0 and it:
save_model(it, model, opt.manualSeed, log_text,
out_dir + '/Iter_{:d}.tar'.format(it))
print('Save model to ' + out_dir + '/Iter_{:d}.tar'.format(it))
if __name__ == "__main__":
train()