-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsingle_experiment.py
297 lines (244 loc) · 14.5 KB
/
single_experiment.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
### execute this function to train and test the vae-model
from vaemodel import Model
import numpy as np
import pickle
import torch
import os
import argparse
import datetime
print('\n' + str(datetime.datetime.now()) + '\n')
def str2bool(v):
if v.lower() in ('yes', 'true', 't', 'y', '1'):
return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
parser = argparse.ArgumentParser()
parser.add_argument('--dataset')
#parser.add_argument('--num_shots',type=int)
#parser.add_argument('--generalized', type=str2bool)
parser.add_argument('--pretrain', type=str2bool)
parser.add_argument('--mod_dataset')
parser.add_argument('--device')
parser.add_argument('--num_gen_iter', type=int, default=1)
parser.add_argument('--num_dis_iter', type=int, default=1)
args = parser.parse_args()
########################################
# the basic hyperparameters
########################################
hyperparameters = {
'num_shots': 0,
'device': 'cuda',
'model_specifics': {'cross_reconstruction': True,
'name': 'CADA',
'distance': 'wasserstein',
'warmup': {'beta': {'factor': 0.25,
'end_epoch': 93,
'start_epoch': 0},
'cross_reconstruction': {'factor': 2.37,
'end_epoch': 75,
'start_epoch': 21},
'distance': {'factor': 8.13,
'end_epoch': 22,
'start_epoch': 6}}},
'lr_gen_model': 0.00001,
'generalized': True,
'pretrain': False,
'mod_dataset': '',
'num_gen_iter': 5,
'batch_size': 50,
'xyu_samples_per_class': {'SUN': (200, 0, 400, 0),
'APY': (200, 0, 400, 0),
'CUB': (200, 0, 400, 0),
'AWA2': (200, 0, 400, 0),
'FLO': (200, 0, 400, 0),
'AWA1': (200, 0, 400, 0)},
'epochs': 300,
'loss': 'l1',
'auxiliary_data_source' : 'attributes',
'lr_cls': 0.001,
'dataset': 'CUB',
'hidden_size_rule': {'resnet_features': (1560, 1660),
'attributes': (1450, 665),
'sentences': (1450, 665) },
'latent_size': 64
}
# The training epochs for the final classifier, for early stopping,
# as determined on the validation spit
cls_train_steps = [
{'dataset': 'SUN', 'num_shots': 0, 'generalized': True, 'cls_train_steps': 31},
{'dataset': 'SUN', 'num_shots': 0, 'generalized': False, 'cls_train_steps': 30},
{'dataset': 'SUN', 'num_shots': 1, 'generalized': True, 'cls_train_steps': 22},
{'dataset': 'SUN', 'num_shots': 1, 'generalized': False, 'cls_train_steps': 96},
{'dataset': 'SUN', 'num_shots': 5, 'generalized': True, 'cls_train_steps': 29},
{'dataset': 'SUN', 'num_shots': 5, 'generalized': False, 'cls_train_steps': 78},
{'dataset': 'SUN', 'num_shots': 2, 'generalized': True, 'cls_train_steps': 29},
{'dataset': 'SUN', 'num_shots': 2, 'generalized': False, 'cls_train_steps': 61},
{'dataset': 'SUN', 'num_shots': 10, 'generalized': True, 'cls_train_steps': 79},
{'dataset': 'SUN', 'num_shots': 10, 'generalized': False, 'cls_train_steps': 94},
{'dataset': 'AWA1', 'num_shots': 0, 'generalized': True, 'cls_train_steps': 33},
{'dataset': 'AWA1', 'num_shots': 0, 'generalized': False, 'cls_train_steps': 25},
{'dataset': 'AWA1', 'num_shots': 1, 'generalized': True, 'cls_train_steps': 40},
{'dataset': 'AWA1', 'num_shots': 1, 'generalized': False, 'cls_train_steps': 81},
{'dataset': 'AWA1', 'num_shots': 5, 'generalized': True, 'cls_train_steps': 89},
{'dataset': 'AWA1', 'num_shots': 5, 'generalized': False, 'cls_train_steps': 62},
{'dataset': 'AWA1', 'num_shots': 2, 'generalized': True, 'cls_train_steps': 56},
{'dataset': 'AWA1', 'num_shots': 2, 'generalized': False, 'cls_train_steps': 59},
{'dataset': 'AWA1', 'num_shots': 10, 'generalized': True, 'cls_train_steps': 100},
{'dataset': 'AWA1', 'num_shots': 10, 'generalized': False, 'cls_train_steps': 50},
#changed for ablation {'dataset': 'CUB', 'num_shots': 0, 'generalized': True, 'cls_train_steps': 36},
{'dataset': 'CUB', 'num_shots': 0, 'generalized': True, 'cls_train_steps': 27},
{'dataset': 'CUB', 'num_shots': 0, 'generalized': False, 'cls_train_steps': 22},
{'dataset': 'CUB', 'num_shots': 1, 'generalized': True, 'cls_train_steps': 34},
{'dataset': 'CUB', 'num_shots': 1, 'generalized': False, 'cls_train_steps': 46},
{'dataset': 'CUB', 'num_shots': 5, 'generalized': True, 'cls_train_steps': 64},
{'dataset': 'CUB', 'num_shots': 5, 'generalized': False, 'cls_train_steps': 73},
{'dataset': 'CUB', 'num_shots': 2, 'generalized': True, 'cls_train_steps': 39},
{'dataset': 'CUB', 'num_shots': 2, 'generalized': False, 'cls_train_steps': 31},
{'dataset': 'CUB', 'num_shots': 10, 'generalized': True, 'cls_train_steps': 85},
{'dataset': 'CUB', 'num_shots': 10, 'generalized': False, 'cls_train_steps': 67},
{'dataset': 'AWA2', 'num_shots': 0, 'generalized': True, 'cls_train_steps': 31},
{'dataset': 'AWA2', 'num_shots': 0, 'generalized': False, 'cls_train_steps': 39},
{'dataset': 'AWA2', 'num_shots': 1, 'generalized': True, 'cls_train_steps': 44},
{'dataset': 'AWA2', 'num_shots': 1, 'generalized': False, 'cls_train_steps': 96},
{'dataset': 'AWA2', 'num_shots': 5, 'generalized': True, 'cls_train_steps': 99},
{'dataset': 'AWA2', 'num_shots': 5, 'generalized': False, 'cls_train_steps': 100},
{'dataset': 'AWA2', 'num_shots': 2, 'generalized': True, 'cls_train_steps': 69},
{'dataset': 'AWA2', 'num_shots': 2, 'generalized': False, 'cls_train_steps': 79},
{'dataset': 'AWA2', 'num_shots': 10, 'generalized': True, 'cls_train_steps': 86},
{'dataset': 'AWA2', 'num_shots': 10, 'generalized': False, 'cls_train_steps': 78}
]
##################################
# change some hyperparameters here
##################################
hyperparameters['dataset'] = args.dataset
#hyperparameters['num_shots']= args.num_shots
#hyperparameters['generalized']= args.generalized
hyperparameters['pretrain'] = args.pretrain
hyperparameters['mod_dataset'] = args.mod_dataset
hyperparameters['device']= args.device
hyperparameters['num_gen_iter']= args.num_gen_iter
hyperparameters['num_dis_iter']= args.num_dis_iter
hyperparameters['cls_train_steps'] = [x['cls_train_steps'] for x in cls_train_steps
if all([hyperparameters['dataset']==x['dataset'],
hyperparameters['num_shots']==x['num_shots'],
hyperparameters['generalized']==x['generalized'] ])][0]
print('***')
print(hyperparameters['cls_train_steps'] )
if hyperparameters['generalized']:
if hyperparameters['num_shots']==0:
hyperparameters['samples_per_class'] = {'CUB': (200, 0, 400, 0), 'SUN': (200, 0, 400, 0),
'APY': (200, 0, 400, 0), 'AWA1': (200, 0, 400, 0),
'AWA2': (200, 0, 400, 0), 'FLO': (200, 0, 400, 0)}
else:
hyperparameters['samples_per_class'] = {'CUB': (200, 0, 200, 200), 'SUN': (200, 0, 200, 200),
'APY': (200, 0, 200, 200), 'AWA1': (200, 0, 200, 200),
'AWA2': (200, 0, 200, 200), 'FLO': (200, 0, 200, 200)}
else:
if hyperparameters['num_shots']==0:
hyperparameters['samples_per_class'] = {'CUB': (0, 0, 200, 0), 'SUN': (0, 0, 200, 0),
'APY': (0, 0, 200, 0), 'AWA1': (0, 0, 200, 0),
'AWA2': (0, 0, 200, 0), 'FLO': (0, 0, 200, 0)}
else:
hyperparameters['samples_per_class'] = {'CUB': (0, 0, 200, 200), 'SUN': (0, 0, 200, 200),
'APY': (0, 0, 200, 200), 'AWA1': (0, 0, 200, 200),
'AWA2': (0, 0, 200, 200), 'FLO': (0, 0, 200, 200)}
if (hyperparameters['dataset'] == 'CUB'):
hyperparameters['latent_size'] = 128
# if (hyperparameters['dataset'] == 'SUN'):
# hyperparameters['latent_size'] = 128
model = Model( hyperparameters)
model.to(hyperparameters['device'])
losses_log = 0
if hyperparameters['pretrain']:
########################################
### load model where u left
########################################
# load VAEGAN parameters from modified model
if hyperparameters['mod_dataset']:
saved_best_state = torch.load('./param/CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_BestEp.pth.tar')
best_epoch = saved_best_state['best_epoch']
saved_state = torch.load('./param/CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_ep' + str(best_epoch) + '.pth.tar')
print('\n USING SAVED PRETRAIN PARAMETER \n CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_ep' + str(best_epoch) + '.pth.tar \n')
for d in model.all_data_sources:
model.encoder[d].load_state_dict(saved_state['encoder'][d])
model.decoder[d].load_state_dict(saved_state['decoder'][d])
# load VAEGAN parameters from original model
else:
saved_best_state = torch.load('./param/CADA_trained_' + hyperparameters['dataset'] + '_BestEp.pth.tar')
best_epoch = saved_best_state['best_epoch']
saved_state = torch.load('./param/CADA_trained_' + hyperparameters['dataset'] + '_ep' + str(best_epoch) + '.pth.tar')
print('\n USING SAVED PRETRAIN PARAMETER \n CADA_trained_' + hyperparameters['dataset'] + '_ep' + str(best_epoch) + '.pth.tar \n')
for d in model.all_data_sources:
model.encoder[d].load_state_dict(saved_state['encoder'][d])
model.decoder[d].load_state_dict(saved_state['decoder'][d])
########################################
# train the classifier using the loaded VAEGAN parameters
model.clf.eval()
u,s,h,history = model.train_classifier()
# train VAEGAN model without pretrained parameters
else:
best_u = 0.0
best_s = 0.0
best_h = -1.0
history = [0.0, 0.0, 0.0, 0.0]
best_epoch = 0
for epoch in range(0, hyperparameters['epochs']):
losses_G, losses_D_att, losses_D_img, losses_log = model.train_vae(epoch)
# train the classifier
if (epoch%5==0 and epoch >= 20) or (epoch==(hyperparameters['epochs']-1)):
u,s,h,history = model.train_classifier()
if h > best_h:
best_u = u
best_s = s
best_h = h
best_epoch = epoch
print('\nbest epoch :' + str(best_epoch) + '\n')
print(model.encoder['resnet_features'].state_dict()['feature_encoder.0.weight'])
state = {
'model': model.state_dict(),
'hyperparameters':hyperparameters,
'epoch': epoch,
'encoder':{'resnet_features': model.encoder['resnet_features'].state_dict(),
'attributes': model.encoder['attributes'].state_dict()
},
'decoder':{'resnet_features': model.decoder['resnet_features'].state_dict(),
'attributes': model.decoder['attributes'].state_dict()
},
'discriminator': {'net_D_Att': model.net_D_Att.state_dict(),
'net_D_Img': model.net_D_Img.state_dict()
},
'optimizer_G': model.optimizer_G.state_dict(),
'optimized_D': model.optimizer_D.state_dict(),
'loss_log': losses_log
}
# saving VAEGAN parameters for each epoch
if hyperparameters['mod_dataset']:
torch.save(state, 'param/CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_ep' + str(epoch) + '.pth.tar')
print('>> saved CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_ep' + str(epoch) + '.pth.tar')
else:
torch.save(state, 'param/CADA_trained_' + hyperparameters['dataset'] + '_ep' + str(epoch) + '.pth.tar')
print('>> saved CADA_trained_' + hyperparameters['dataset'] + '_ep' + str(epoch) + '.pth.tar')
print('\nBest VAE Epoch %.1f | Novel %.4f | Seen %.4f | H %.4f \n' % (
best_epoch, u, s, best_h))
# checking the saved parameters
best_state = {'best_epoch': best_epoch}
if hyperparameters['mod_dataset']:
torch.save(best_state, 'param/CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_BestEp.pth.tar')
print('>> saved CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_BestEp.pth.tar')
else:
torch.save(best_state, 'param/CADA_trained_' + hyperparameters['dataset'] + '_BestEp.pth.tar')
print('>> saved CADA_trained_' + hyperparameters['dataset'] + '_BestEp.pth.tar')
hyperparameters['pretrain']=True
model.pretrain = True
print('\n\nCHECKING BEST CLASSIFIER MODEL PARAMETER')
if hyperparameters['mod_dataset']:
saved_best_state = torch.load('./param/CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_BestEp.pth.tar')
best_epoch = saved_best_state['best_epoch']
saved_state = torch.load('./param/CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_ep' + str(best_epoch) + '.pth.tar')
print('\nUSING SAVED PRETRAIN PARAMETER \n CADA_trained_' + hyperparameters['dataset'] + '_' + hyperparameters['mod_dataset'] + '_ep' + str(best_epoch) + '.pth.tar \n')
for d in model.all_data_sources:
model.encoder[d].load_state_dict(saved_state['encoder'][d])
model.decoder[d].load_state_dict(saved_state['decoder'][d])
print(model.encoder['resnet_features'].state_dict()['feature_encoder.0.weight'])