-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain2.py
251 lines (204 loc) · 10.9 KB
/
main2.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
"""
Does the semi-supervised experiments.
"""
import os
import argparse
import pickle
import numpy as np
from skopt import gp_minimize
import keras
from keras.optimizers import SGD
from keras import backend as K
from resnet101 import resnet101
from multi_gpu import to_multi_gpu, to_single_gpu
import ml_loss
from datahandler import DataHandler
import metrics
import params
def update_mixed_labels(model):
"""
Propagates labels to unlabeled examples using the features extracted with the model
"""
if params.n_gpus > 1:
feat_model = to_single_gpu(model)
feat_model = keras.Model(feat_model.layers[0].input, feat_model.layers[-2].output) # pylint: disable=E1101
if params.n_gpus > 1:
feat_model = to_multi_gpu(feat_model, n_gpus=params.n_gpus)
no_batches = int(np.ceil(float(len(dh.inds_labeled)) / params.batch_size))
l_feats = feat_model.predict_generator(dh.generator('train_labeled_sorted', aug=False, shuffle_batches=False), no_batches)
l_feats = np.concatenate((l_feats[:(no_batches - 1) * params.batch_size], l_feats[-(len(dh.inds_labeled) - (no_batches - 1) * params.batch_size):]))
no_batches = int(np.ceil(float(len(dh.inds_unlabeled)) / params.batch_size))
ul_feats = feat_model.predict_generator(dh.generator('train_unlabeled', aug=False, shuffle_batches=False), no_batches)
ul_feats = np.concatenate((ul_feats[:(no_batches - 1) * params.batch_size], ul_feats[-(len(dh.inds_unlabeled) - (no_batches - 1) * params.batch_size):]))
min_dists = np.zeros(ul_feats.shape[0], dtype=np.float32)
min_dist_inds = np.zeros(ul_feats.shape[0], dtype=np.int)
for ind_unlabeled in range(ul_feats.shape[0]):
min_dists[ind_unlabeled] = np.Inf
for ind_labeled in range(l_feats.shape[0]):
dist = np.linalg.norm(l_feats[ind_labeled] - ul_feats[ind_unlabeled])
if dist < min_dists[ind_unlabeled]:
min_dists[ind_unlabeled] = dist
min_dist_inds[ind_unlabeled] = ind_labeled
no_labeled = l_feats.shape[0]
no_unlabeled = ul_feats.shape[0]
no_mixed = no_labeled + no_unlabeled
mean_dist = np.mean(min_dists) * (float(no_unlabeled) / no_mixed)
similarity_scores = np.exp(-min_dists / mean_dist)
mixed_labels = np.zeros((no_mixed, dh.no_classes[dh.dataset] + 1), dtype=np.float32)
mixed_labels[dh.inds_labeled_sorted, :-1] = dh.train_labels[dh.inds_labeled_sorted]
mixed_labels[dh.inds_labeled_sorted, -1] = 5
for ind_unlabeled in range(no_unlabeled):
mixed_labels[dh.inds_unlabeled[ind_unlabeled], :-1] = dh.train_labels[dh.inds_labeled_sorted[min_dist_inds[ind_unlabeled]]]
mixed_labels[dh.inds_unlabeled[ind_unlabeled], -1] = similarity_scores[ind_unlabeled]
dh.mixed_labels = mixed_labels
"""for ind in range(no_mixed):
if ind in dh.inds_labeled:
print('Labeled')
else:
print('Propagated', dh.mixed_labels[ind, -1])
print(np.where(dh.train_labels[ind] == 1)[0])
print(np.where(dh.mixed_labels[ind, :-1] == 1)[0])
if ind % 5 == 0:
print(ind)"""
def test_model(model):
"""
Calculates the metrics on the validation data. Slightly modified for semi-supervised training
"""
no_examples = len(dh.val_images)
no_batches = int(np.ceil(float(no_examples) / params.batch_size))
preds = np.empty((no_examples, dh.no_classes[args.dataset]), dtype=np.float32)
labels = np.empty((no_examples, dh.no_classes[args.dataset]), dtype=np.float32)
gen = dh.generator('val', aug=False, shuffle_batches=False)
for ind_batch in range(no_batches):
images_batch, labels_batch = gen.next()
preds_batch = model.predict(images_batch, batch_size=params.batch_size)
if ind_batch == no_batches - 1:
preds[no_examples - params.batch_size:no_examples] = preds_batch[:, :-1]
labels[no_examples - params.batch_size:no_examples] = labels_batch
else:
preds[ind_batch * params.batch_size:(ind_batch + 1) * params.batch_size] = preds_batch[:, :-1]
labels[ind_batch * params.batch_size:(ind_batch + 1) * params.batch_size] = labels_batch
return metrics.calculate_metrics(labels, preds)
def run_experiment(x):
"""
Runs a single experiment with the given learning rate and weight decay parameters
"""
learning_rate = x[0]
weight_decay = x[1]
global log_path
if args.optimize:
global step
step += 1 # pylint: disable=E0602
log_path = os.path.join('log2', args.dataset, args.ml_method, args.init, str(args.labeled_ratio), str(args.corruption_ratio), str(step))
if not os.path.exists(log_path):
os.makedirs(log_path)
with open(os.path.join(log_path, 'his.txt'), 'w') as f:
f.write('Learning rate: ' + str(x[0]) + '\n')
f.write('Weight decay: ' + str(x[1]) + '\n')
K.clear_session()
# load pretrained model for label propagation
if args.ml_method == 'robust_warp':
model_path = os.path.join('log', args.dataset, 'robust_warp_sup', args.init, str(args.labeled_ratio), str(args.corruption_ratio), 'best_cp.h5')
else:
model_path = os.path.join('log', args.dataset, args.ml_method, args.init, str(args.labeled_ratio), str(args.corruption_ratio), 'best_cp.h5')
model_orig = resnet101(dh.no_classes[args.dataset], initialization='random', weight_decay=weight_decay)
if params.n_gpus > 1:
model_orig = to_multi_gpu(model_orig, n_gpus=params.n_gpus)
model_orig.load_weights(model_path)
update_mixed_labels(model_orig)
model = resnet101(dh.no_classes[args.dataset] + 1, initialization='random', weight_decay=weight_decay)
if params.n_gpus > 1:
model_orig = to_single_gpu(model_orig)
for ind_layer in range(len(model.layers)):
if model.layers[ind_layer].name == model_orig.layers[ind_layer].name:
model.layers[ind_layer].set_weights(model_orig.layers[ind_layer].get_weights())
if params.n_gpus > 1:
model = to_multi_gpu(model, n_gpus=params.n_gpus)
sgd = SGD(lr=learning_rate, momentum=0.9, decay=0.0, nesterov=True)
model.compile(loss=loss_function, optimizer=sgd, metrics=[loss_function])
ind_lr_step = 0
train_losses = []
val_losses = []
patience_losses = []
for ind_epoch in range(params.max_epoch):
if ind_epoch % 20 == 0 and not ind_epoch == 0:
update_mixed_labels(model)
his = model.fit_generator(generator=dh.generator('train_mixed', aug=True),
steps_per_epoch=dh.mixed_labels.shape[0] / params.batch_size,
validation_data=dh.generator('val', aug=False),
validation_steps=len(dh.val_images) / params.batch_size / 10,
verbose=2)
train_losses.append(his.history['loss'][0])
val_losses.append(his.history['val_loss'][0])
patience_losses.append(his.history['val_loss'][0])
if min(patience_losses) == patience_losses[-1]:
model.save_weights(os.path.join(log_path, str(ind_lr_step) + '_cp.h5'))
elif np.argmin(np.array(patience_losses)) < len(patience_losses) - 1 - params.lr_patience:
with open(os.path.join(log_path, 'his.txt'), 'a') as f:
f.write('loss: ' + str(train_losses) + '\n')
f.write('val_loss: ' + str(val_losses) + '\n')
if ind_lr_step == params.no_lr_steps - 1:
model.load_weights(os.path.join(log_path, str(ind_lr_step) + '_cp.h5'), by_name=True)
break
else:
model.load_weights(os.path.join(log_path, str(ind_lr_step) + '_cp.h5'), by_name=True)
learning_rate /= 10
K.set_value(model.optimizer.lr, learning_rate)
ind_lr_step += 1
model.save_weights(os.path.join(log_path, str(ind_lr_step) + '_cp.h5'))
train_losses = []
val_losses = []
model.save_weights(os.path.join(log_path, 'best_cp.h5'))
res_metrics = test_model(model)
with open(os.path.join(log_path, 'metrics.p'), 'wb') as f:
pickle.dump(res_metrics, f, pickle.HIGHEST_PROTOCOL)
with open(os.path.join(log_path, 'metrics.txt'), 'w') as f:
f.write(str(res_metrics) + '\n')
return -res_metrics['f1c_top3'] # negative because gp_minimize tries to minimize the result
def main():
global log_path
global step
if args.optimize:
if args.cont:
with open(os.path.join(log_path, 'opt_res.p'), 'rb') as f:
past_res = pickle.load(f)
step = len(past_res['x_iters'])
res = gp_minimize(run_experiment, params.opt_interval[args.init + '2'], n_random_starts=params.no_random_starts, n_calls=params.no_opt_iters,
x0=past_res['x_iters'], y0=past_res['func_vals'])
else:
res = gp_minimize(run_experiment, params.opt_interval[args.init], n_random_starts=params.no_random_starts, n_calls=params.no_opt_iters)
log_path = os.path.join('log2', args.dataset, args.ml_method, args.init, str(args.labeled_ratio), str(args.corruption_ratio))
with open(os.path.join(log_path, 'opt_res.p'), 'wb') as f:
pickle.dump(res, f, pickle.HIGHEST_PROTOCOL)
with open(os.path.join(log_path, 'opt_res.txt'), 'w') as f:
f.write('Optimization results:\n')
f.write(str(res) + '\n')
else:
run_experiment([params.learning_rate[args.init + '2'], params.weight_decay[args.init + '2']])
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('dataset', choices=['nus_wide', 'ms_coco'])
parser.add_argument('init', choices=['imagenet', 'random'])
parser.add_argument('ml_method', choices=['br', 'sm', 'pwr', 'warp', 'robust_warp', 'robust_warp_sup'])
parser.add_argument('labeled_ratio', type=int, choices=[10, 20, 100]) # percentage
parser.add_argument('corruption_ratio', type=int, choices=range(0, 60, 10)) # percentage
parser.add_argument('--optimize', action='store_true', help='does hyperparameter optimization')
parser.add_argument('--cont', action='store_true', help='continues optimization')
args = parser.parse_args()
log_path = os.path.join('log2', args.dataset, args.ml_method, args.init, str(args.labeled_ratio), str(args.corruption_ratio))
if not os.path.exists(log_path):
os.makedirs(log_path)
dh = DataHandler(args.dataset, args.labeled_ratio, args.corruption_ratio)
if args.ml_method == 'br':
loss_function = ml_loss.binary_relevance.binary_relevance
elif args.ml_method == 'sm':
loss_function = ml_loss.ml_softmax.ml_softmax
elif args.ml_method == 'pwr':
loss_function = ml_loss.pairwise_ranking.pairwise_ranking
elif args.ml_method == 'warp':
loss_function = ml_loss.warp_py.warp
elif args.ml_method == 'robust_warp':
loss_function = ml_loss.robust_warp_py.robust_warp
if args.optimize:
step = 0
main()