-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathfast_fgvr_semi_train.py
287 lines (227 loc) · 12.7 KB
/
fast_fgvr_semi_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
import os
import time
import json
import numpy as np
import tensorflow as tf
from pydoc import locate
from utils import user_io
import constants as const
from utils import os_utils
from utils import log_utils
from ranking import center_loss
from ranking import triplet_semi
from ranking import triplet_hard
import utils.tb_utils as tb_utils
import utils.tf_utils as tf_utils
from nets.conv_embed import ConvEmbed
from data_sampling.quick_tuple_loader import QuickTupleLoader
from data_sampling.triplet_tuple_loader import TripletTupleLoader
from config.base_config import BaseConfig
def touch_dir(path):
if(not os.path.exists(path)):
os.makedirs(path)
else:
print(path)
if not user_io.ask_yes_no_question('Model dir already exists, continue -- override?'):
quit()
def main(argv):
cfg = BaseConfig().parse(argv)
os.environ["CUDA_VISIBLE_DEVICES"] = cfg.gpu
save_model_dir = cfg.checkpoint_dir
model_basename = os.path.basename(save_model_dir)
touch_dir(save_model_dir)
args_file = os.path.join(cfg.checkpoint_dir,'args.json')
with open(args_file, 'w') as f:
json.dump(vars(cfg), f, ensure_ascii=False, indent=2, sort_keys=True)
# os_utils.touch_dir(save_model_dir)
log_file = os.path.join(cfg.checkpoint_dir, cfg.log_filename + '.txt')
os_utils.touch_dir(cfg.checkpoint_dir)
logger = log_utils.create_logger(log_file)
img_generator_class = locate(cfg.db_tuple_loader)
args = dict()
args['db_path'] = cfg.db_path
args['tuple_loader_queue_size'] = cfg.tuple_loader_queue_size
args['preprocess_func'] = cfg.preprocess_func
args['batch_size'] = cfg.batch_size
args['shuffle'] = False
args['csv_file'] = cfg.train_csv_file
args['img_size'] = const.max_frame_size
args['gen_hot_vector'] = True
train_iter = img_generator_class(args)
args['batch_size'] = cfg.batch_size
args['csv_file'] = cfg.test_csv_file
val_iter = img_generator_class(args)
trn_images, trn_lbls = train_iter.imgs_and_lbls()
val_imgs, val_lbls = val_iter.imgs_and_lbls()
with tf.Graph().as_default():
if cfg.train_mode == 'semi_hard' or cfg.train_mode == 'hard' or cfg.train_mode == 'cntr':
train_dataset = TripletTupleLoader(trn_images, trn_lbls,cfg).dataset
elif cfg.train_mode == 'vanilla':
train_dataset = QuickTupleLoader(trn_images, trn_lbls,cfg,is_training=True, shuffle=True,repeat=True).dataset
else:
raise NotImplementedError('{} is not a valid train mode'.format(cfg.train_mode))
val_dataset = QuickTupleLoader(val_imgs, val_lbls,cfg, is_training=False,repeat=False).dataset
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(
handle, train_dataset.output_types, train_dataset.output_shapes)
images_ph, lbls_ph = iterator.get_next()
network_class = locate(cfg.network_name)
model = network_class(cfg,images_ph=images_ph, lbls_ph=lbls_ph)
# Which loss fn to impose. For example, softmax only is applied in vanilla mode,
# while softmax + semi-hard triplet is applied in semi_hard mode.
if cfg.train_mode == 'semi_hard':
pre_logits = model.train_pre_logits
_, w, h, channels = pre_logits.shape
embed_dim = cfg.emb_dim
embedding_net = ConvEmbed(emb_dim=embed_dim, n_input=channels, n_h=h, n_w=w)
embedding = embedding_net.forward(pre_logits)
embedding = tf.nn.l2_normalize(embedding, axis=-1, epsilon=1e-10)
margin = cfg.margin
gt_lbls = tf.argmax(model.gt_lbls, 1);
metric_loss = triplet_semi.triplet_semihard_loss(gt_lbls, embedding, margin)
logger.info('Triplet loss lambda {}, with margin {}'.format(cfg.triplet_loss_lambda,margin))
total_loss = model.train_loss + cfg.triplet_loss_lambda * tf.reduce_mean(metric_loss)
elif cfg.train_mode == 'hard':
pre_logits = model.train_pre_logits
_, w, h, channels = pre_logits.shape
embed_dim = cfg.emb_dim
embedding_net = ConvEmbed(emb_dim=embed_dim, n_input=channels, n_h=h, n_w=w)
embedding = embedding_net.forward(pre_logits)
embedding = tf.nn.l2_normalize(embedding, axis=-1, epsilon=1e-10)
margin = cfg.margin
logger.info('Triplet loss lambda {}, with margin {}'.format(cfg.triplet_loss_lambda, margin))
gt_lbls = tf.argmax(model.gt_lbls, 1);
metric_loss = triplet_hard.batch_hard(gt_lbls, embedding, margin)
total_loss = model.train_loss + cfg.triplet_loss_lambda * tf.reduce_mean(metric_loss)
elif cfg.train_mode == 'cntr':
pre_logits = model.train_pre_logits
_, w, h, channels = pre_logits.shape
embed_dim = cfg.emb_dim
embedding_net = ConvEmbed(emb_dim=embed_dim, n_input=channels, n_h=h, n_w=w)
embedding = embedding_net.forward(pre_logits)
embedding = tf.nn.l2_normalize(embedding, axis=-1, epsilon=1e-10)
CENTER_LOSS_LAMBDA = 0.003
CENTER_LOSS_ALPHA = 0.5
num_fg_classes = cfg.num_classes
gt_lbls = tf.argmax(model.gt_lbls, 1);
center_loss_order, centroids, centers_update_op, appear_times, diff = center_loss.get_center_loss(embedding, gt_lbls,
CENTER_LOSS_ALPHA,
num_fg_classes)
# sample_centroid = tf.reshape(tf.gather(centroids, gt_lbls), [-1, config.emb_dim])
# center_loss_order = center_loss.center_loss(sample_centroid , embedding)
logger.info('Center loss lambda {}'.format(CENTER_LOSS_LAMBDA))
total_loss = model.train_loss + CENTER_LOSS_LAMBDA * tf.reduce_mean(center_loss_order)
elif cfg.train_mode == 'vanilla':
total_loss = model.train_loss
logger.info('Train Mode {}'.format(cfg.train_mode))
# variables_to_train = model.var_2_train();
# logger.info('variables_to_train ' + str(variables_to_train))
trainable_vars = tf.trainable_variables()
if cfg.caffe_iter_size > 1: ## Accumulated Gradient
## Creation of a list of variables with the same shape as the trainable ones
# initialized with 0s
accum_vars = [tf.Variable(tf.zeros_like(tv.initialized_value()), trainable=False) for tv in trainable_vars]
zero_ops = [tv.assign(tf.zeros_like(tv)) for tv in accum_vars]
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if cfg.train_mode == const.Train_Mode.CNTR:
update_ops.append(centers_update_op)
# print(update_ops)
with tf.control_dependencies(update_ops):
global_step = tf.Variable(0, name='global_step', trainable=False)
learning_rate = tf_utils.poly_lr(global_step,cfg)
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
if cfg.caffe_iter_size > 1: ## Accumulated Gradient
# grads = tf.Print(grads,[grads],'Grad Print');
grads = optimizer.compute_gradients(total_loss, trainable_vars)
# Adds to each element from the list you initialized earlier with zeros its gradient (works because accum_vars and gvs are in the same order)
accum_ops = [accum_vars[i].assign_add(gv[0]) for i, gv in enumerate(grads)]
iter_size = cfg.caffe_iter_size
# Define the training step (part with variable value update)
train_op = optimizer.apply_gradients([(accum_vars[i] / iter_size, gv[1]) for i, gv in enumerate(grads)],
global_step=global_step)
else:
grads = optimizer.compute_gradients(total_loss)
train_op = optimizer.apply_gradients(grads, global_step=global_step)
sess = tf.InteractiveSession()
training_iterator = train_dataset.make_one_shot_iterator()
validation_iterator = val_dataset.make_initializable_iterator()
training_handle = sess.run(training_iterator.string_handle())
validation_handle = sess.run(validation_iterator.string_handle())
tb_path = save_model_dir
logger.info(tb_path)
start_iter = tb_utils.get_latest_iteration(tb_path)
train_writer = tf.summary.FileWriter(tb_path, sess.graph)
tf.global_variables_initializer().run()
saver = tf.train.Saver() # saves variables learned during training
ckpt_file = tf.train.latest_checkpoint(save_model_dir)
logger.info('Model Path {}'.format(ckpt_file))
load_model_msg = model.load_model(save_model_dir, ckpt_file, sess, saver, load_logits=False)
logger.info(load_model_msg)
ckpt_file = os.path.join(save_model_dir, cfg.checkpoint_filename)
val_loss = tf.summary.scalar('Val_Loss', model.val_loss)
val_acc_op = tf.summary.scalar('Batch_Val_Acc', model.val_accuracy)
model_acc_op = tf.summary.scalar('Split_Val_Accuracy', model.val_accumulated_accuracy)
best_model_step = 0
best_acc = 0
logger.info('Start Training from {}, till {}'.format(start_iter,cfg.train_iters))
# Start Training
for step in range(start_iter + 1, cfg.train_iters + 1):
start_time_train = time.time()
# Update network weights while supporting caffe_iter_size
for mini_batch in range(cfg.caffe_iter_size - 1):
feed_dict = {handle: training_handle}
model_loss_value, accuracy_value, _ = sess.run(
[model.train_loss, model.train_accuracy, accum_ops], feed_dict)
feed_dict = {handle: training_handle}
model_loss_value, accuracy_value, _= sess.run([model.train_loss, model.train_accuracy, train_op],
feed_dict)
if cfg.caffe_iter_size > 1: ## Accumulated Gradient
sess.run(zero_ops)
train_time = time.time() - start_time_train
if (step == 1 or step % cfg.logging_threshold == 0):
logger.info(
'i {0:04d} loss {1:4f} Acc {2:2f} Batch Time {3:3f}'.format(step,model_loss_value,
accuracy_value,
train_time))
if (step % cfg.test_interval == 0):
run_metadata = tf.RunMetadata()
tf.local_variables_initializer().run()
sess.run(validation_iterator.initializer)
_val_acc_op = 0
while True:
try:
# Eval network on validation/testing split
feed_dict = {handle: validation_handle}
val_loss_op, batch_accuracy, accuracy_op, _val_acc_op, _val_acc, c_cnf_mat,macro_acc = sess.run(
[val_loss, model.val_accuracy, model_acc_op, val_acc_op, model.val_accumulated_accuracy,
model.val_confusion_mat,model.val_per_class_acc_acc], feed_dict)
except tf.errors.OutOfRangeError:
logger.info('Val Acc {0}, Macro Acc: {1}'.format(_val_acc,macro_acc))
break
train_writer.add_run_metadata(run_metadata, 'step%03d' % step)
train_writer.add_summary(val_loss_op, step)
train_writer.add_summary(_val_acc_op, step)
train_writer.add_summary(accuracy_op, step)
train_writer.flush()
if (step % 100 == 0):
saver.save(sess, ckpt_file)
if best_acc < _val_acc:
saver.save(sess, ckpt_file + 'best')
best_acc = _val_acc
best_model_step = step
logger.info('Best Acc {0} at {1} == {2}'.format(best_acc, best_model_step, model_basename))
logger.info('Triplet loss lambda {}'.format(cfg.triplet_loss_lambda))
logger.info('Mode {}'.format(cfg.train_mode))
logger.info('Loop complete')
sess.close()
if __name__ == '__main__':
arg_db_name = 'cars'
arg_net = 'inc4'
arg_ckpt = 'test_{}_{}'.format(arg_db_name,arg_net)
args = [
'--gpu', '0',
'--checkpoint_dir',arg_ckpt,
'--db_name',arg_db_name ,
'--net',arg_net ,
]
main(args)