-
Notifications
You must be signed in to change notification settings - Fork 0
/
query_expansion_main.py
407 lines (339 loc) · 17.3 KB
/
query_expansion_main.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
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import cv2
import argparse
import time
import pickle
import sys
import math
import nsml
from nsml import DATASET_PATH
from tensorflow.python.tools.inspect_checkpoint import print_tensors_in_checkpoint_file
import numpy as np
import tensorflow as tf
from image_processing import preprocess, get_aug_config
from data_loader import get_assignment_map_from_checkpoint,\
get_balanced_dual_dataset, get_dual_dataset, image_load, train_data_loader, \
convert_to_query_db_data, convert_to_query_db_data_fixed_window, \
convert_to_query_db_data_for_generator
from measure import evaluate_mAP, evaluate_rank
from inference import get_feature
from train_utils import l2_normalize
from loss import batch_hard_triplet_loss
from model.delf_model import *
from imgaug import augmenters as iaa
import imgaug as ia
from delf import delf_config_pb2
from delf import feature_extractor
from delf import feature_io
local_infer = None
# bind training model with nsml
def bind_model(sess):
global local_infer
def save(dir_name):
os.makedirs(dir_name, exist_ok=True)
saver = tf.train.Saver()
saver.save(sess, os.path.join(dir_name, 'model'))
def load(file_path):
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(file_path)
if ckpt and ckpt.model_checkpoint_path:
checkpoint = os.path.basename(ckpt.model_checkpoint_path)
saver.restore(sess, os.path.join(file_path, checkpoint))
else:
raise NotImplementedError('No checkpoint!')
print('model loaded :' + file_path)
def infer(queries, references, _query_img=None, _reference_img=None, batch_size=128):
# load, and process images
if _query_img is None:
# not debug
test_path = DATASET_PATH + '/test/test_data'
db = [os.path.join(test_path, 'reference', path) for path in os.listdir(os.path.join(test_path, 'reference'))]
queries = [v.split('/')[-1].split('.')[0] for v in queries]
db = [v.split('/')[-1].split('.')[0] for v in db]
queries.sort()
db.sort()
queries_full_paths = list(map(lambda x: '/data/ir_ph2/test/test_data/query/' + x + '.jpg', queries))
db_full_path = list(map(lambda x: '/data/ir_ph2/test/test_data/reference/' + x + '.jpg', db))
_, query_vecs, _, reference_vecs = get_feature(queries_full_paths, db_full_path, sess, batch_size)
else:
# debug
_, query_vecs, _, reference_vecs = get_feature(_query_img, _reference_img, sess, batch_size)
db = references
expanded_queries= l2_normalize(query_vecs)
reference_vecs = l2_normalize(reference_vecs)
total_sim_matrix = np.empty(
(expanded_queries.shape[0], reference_vecs.shape[0]),
np.float32)
for expanded_query in expanded_queries:
sim_matrix = np.dot(expanded_query, reference_vecs.T)
sim_matrix = np.expand_dims(
np.sum(sim_matrix, axis=0),
axis=0)
np.append(total_sim_matrix, sim_matrix, axis=0)
indices = np.argsort(total_sim_matrix, axis=1)
indices = np.flip(indices, axis=1)
# query = 1, ref = 10
# sim_matrix[0] = [0.1, 0.56, 0.2, 0.5, ....]
# Sort cosine similarity values to rank it
retrieval_results = {}
for (i, query) in enumerate(queries):
ranked_list = [db[k] for k in indices[i]]
ranked_list = ranked_list[:5000]
retrieval_results[query] = ranked_list
return list(zip(range(len(retrieval_results)), retrieval_results.items()))
# DONOTCHANGE: They are reserved for nsml
nsml.bind(save=save, load=load, infer=infer)
local_infer = infer
if __name__ == '__main__':
args = argparse.ArgumentParser()
# hyperparameters
args.add_argument('--epochs', type=int, default=200)
args.add_argument('--batch_size', type=int, default=64)
args.add_argument('--debug', action='store_true', help='debug mode')
args.add_argument('--debug_data', type=str, default="./debug_data", help='debug_data')
args.add_argument('--lr', type=float, default=0.0001, help='learning rate')
args.add_argument('--dev_querynum', type=int, default=300, help='dev split percentage')
args.add_argument('--dev_referencenum', type=int, default=20, help='dev split percentage')
# augmentation
args.add_argument('--augmentation', action='store_true', help='apply random crop in processing')
args.add_argument('--crop', action='store_true', help='set crop images')
args.add_argument('--fliplr', action='store_true', help='set fliplr')
args.add_argument('--flipud', action='store_true', help='set flipud')
args.add_argument('--gausian', action='store_true', help='set gausian')
args.add_argument('--dropout', action='store_true', help='set dropout')
args.add_argument('--noise', action='store_true', help='set noise')
args.add_argument('--rotate', action='store_true', help='rotate -45 degree to +45 degree')
# loss calculation
args.add_argument('--train_logits', action='store_true', help='train similarity and logit jointly')
args.add_argument('--train_sim', action='store_true', help='train similarity and logit jointly')
args.add_argument('--train_sim_dist', action='store_true', help='train similarity and logit jointly using squared loss')
args.add_argument('--train_max_neg', action='store_true', help='train max negative loss')
args.add_argument('--train_max_neg_topk', type=int, default=5, help='set top_k max negative')
args.add_argument('--train_triplet', action="store_true", help="train triplet loss")
# pre trained model
args.add_argument('--pretrained_model', type=str, default=None, help='restore pretrained model')
args.add_argument('--stop_gradient_sim', action='store_true', help='stop gradient similarity')
args.add_argument('--skipcon_attn', action='store_true', help='skip connection attention')
args.add_argument('--logit_concat_sim', action='store_true', help='skip connection attention')
# DONOTCHANGE: They are reserved for nsml
args.add_argument('--mode', type=str, default='train')
args.add_argument('--iteration', type=str, default='0')
args.add_argument('--pause', type=int, default=0)
config = args.parse_args()
print("Model configuration", config)
# training parameters
nb_epoch = config.epochs
batch_size = config.batch_size
"""-------- Model Part -------------------------------------------------"""
num_classes = 1384
input_shape = (224, 224, 3) # input image shape
# set input placeholders
X1 = tf.placeholder(
tf.float32,
[None, input_shape[0], input_shape[1], 3],
name="input_X1")
Y1 = tf.placeholder(tf.float32, [None, num_classes], name="input_Y1")
X2 = tf.placeholder(
tf.float32,
[None, input_shape[0], input_shape[1], 3],
name="input_X2"
)
Y2 = tf.placeholder(tf.float32, [None, num_classes], name="input_Y2")
# init model
global_step = tf.Variable(0, name="mandoo_global_step")
model = Delf_dual_model(X1, X2, num_classes,
skipcon_attn=config.skipcon_attn,
stop_gradient_sim=config.stop_gradient_sim,
logit_concat_sim=config.logit_concat_sim)
# define loss function to optimize
acc_logit = tf.zeros([])
loss_squared_logit = tf.zeros([])
loss_crossent_logit = tf.zeros([])
loss_sim = tf.zeros([])
loss_sim_dist = tf.zeros([])
loss_max_neg = tf.zeros([])
acc_sim = tf.zeros([])
loss_triplet = tf.zeros([])
Y_sim = tf.expand_dims(tf.reduce_sum(Y1 * Y2, axis=1), axis=-1)
if config.train_logits:
loss_crossent_1 = tf.nn.softmax_cross_entropy_with_logits_v2(logits=model.logits_1, labels=Y1)
loss_crossent_2 = tf.nn.softmax_cross_entropy_with_logits_v2(logits=model.logits_2, labels=Y2)
loss_crossent_logit = tf.reduce_sum(loss_crossent_1 + loss_crossent_2)
loss_squared_1 = tf.losses.mean_squared_error(labels=Y1, predictions=tf.nn.softmax(model.logits_1))
loss_squared_2 = tf.losses.mean_squared_error(labels=Y2, predictions=tf.nn.softmax(model.logits_2))
loss_squared_logit = tf.reduce_sum(loss_squared_1 + loss_squared_2)
pred_1 = tf.argmax(model.logits_1, 1, name="pred_1")
pred_2 = tf.argmax(model.logits_2, 1, name="pred_2")
acc_1 = tf.reduce_mean(tf.cast(tf.equal(pred_1, tf.argmax(Y1, 1)), "float"))
acc_2 = tf.reduce_mean(tf.cast(tf.equal(pred_2, tf.argmax(Y2, 1)), "float"))
acc_logit = (acc_1 + acc_2) / 2.0
if config.train_sim:
Y_sim = tf.expand_dims(tf.reduce_sum(Y1 * Y2, axis=1), axis=-1)
pred_sim = tf.cast(tf.greater(tf.nn.sigmoid(model.similarity), 0.5), tf.int64)
acc_sim = tf.reduce_mean(tf.cast(tf.equal(pred_sim, tf.cast(Y_sim, tf.int64)), "float"))
loss_sim = tf.nn.sigmoid_cross_entropy_with_logits(logits=model.similarity, labels=Y_sim)
loss_sim = tf.reduce_sum(loss_sim)
if config.train_sim_dist:
sig_f1 = tf.nn.sigmoid(model.feat_attn_1)
sig_f2 = tf.nn.sigmoid(model.feat_attn_2)
Y_sim_dist = (tf.expand_dims(tf.reduce_sum(Y1 * Y2, axis=1), axis=-1) - 0.5) * 2
loss_sim_dist = Y_sim_dist * tf.losses.mean_squared_error(labels=sig_f1, predictions=sig_f2)
loss_sim_dist = tf.reduce_sum(loss_sim_dist)
if config.train_max_neg:
logit_sum_1, _ = tf.reduce_sum(model.logits_1, axis=1)
pred_value_1, _ = tf.nn.top_k(model.logits_1, k=1)
logit_sum_2, _ = tf.reduce_sum(model.logits_2, axis=1)
pred_value_2, _ = tf.nn.top_k(model.logits_2, k=1)
loss_max_neg_1 = tf.reduce_sum(tf.maximum(tf.subtract(logit_sum_1, pred_value_1), 0.0))
loss_max_neg_2 = tf.reduce_sum(tf.maximum(tf.subtract(logit_sum_2, pred_value_2), 0.0))
loss_max_neg = (loss_max_neg_1 + loss_max_neg_2) / 2.0
if config.train_triplet:
loss_triplet = batch_hard_triplet_loss(Y1, model.feature_vector, 100, squared=True)
loss = loss_sim + loss_squared_logit + loss_crossent_logit \
+ loss_sim_dist + loss_max_neg + loss_triplet
# optimize the loss
optimizer = tf.train.AdamOptimizer(config.lr)
train_op = optimizer.minimize(loss, global_step=global_step)
"""------------------------------------------------------------------------------------"""
# start session
sess = tf.Session()
# load model from scratch or pretrained_model(resnet_v1_50)
if config.pretrained_model == None:
print("Initialize model from the scratch")
init = tf.global_variables_initializer()
sess.run(init)
else:
print("Initialize model from pretrained_model")
all_vars = tf.global_variables()
# load pretrained_model checkpoint
assignment_map, initialized_variable_names = get_assignment_map_from_checkpoint(
tvars=all_vars,
init_checkpoint=config.pretrained_model
)
for var in initialized_variable_names:
print(str(var) + " *INIT FROM CKPT* ")
print("Total {:g} variables are restored from ckpt : {}".format(
len(initialized_variable_names), str(config.pretrained_model)))
tf.train.init_from_checkpoint(
config.pretrained_model, assignment_map)
# find uninitialized variables and initialize it
is_initialized = sess.run([tf.is_variable_initialized(var)
for var in all_vars])
not_initialized_vars = [var
for (var, f) in zip(all_vars, is_initialized)
if not f]
if len(not_initialized_vars):
sess.run(tf.variables_initializer(not_initialized_vars))
saver = tf.train.Saver()
bind_model(sess)
if config.pause:
nsml.paused(scope=locals())
bTrainmode = False
if config.mode == 'train':
bTrainmode = True
""" Load data """
print(DATASET_PATH)
output_path = ['./img_list.pkl', './label_list.pkl']
train_dataset_path = DATASET_PATH + '/train/train_data'
if nsml.IS_ON_NSML:
# Caching file
nsml.cache(train_data_loader, data_path=train_dataset_path,
output_path=output_path)
else:
train_dataset_path = config.debug_data
train_data_loader(train_dataset_path, output_path=output_path)
with open(output_path[0], 'rb') as img_f:
img_list = pickle.load(img_f)
with open(output_path[1], 'rb') as label_f:
label_list = pickle.load(label_f)
queries, references, queries_img, reference_img \
= convert_to_query_db_data_for_generator(img_list, label_list, input_shape, config.dev_querynum, config.dev_referencenum)
print("mAP devset : query(%d), reference(%d) " % (len(queries), len(references)))
dataset = get_balanced_dual_dataset(
train_dataset_path, batch_size, nb_epoch, num_classes=num_classes)
size_of_epoch = len(label_list) * 2
iterator = dataset.make_initializable_iterator()
img_batch_1, img_batch_2, label_batch_1, label_batch_2 = iterator.get_next()
sess.run([iterator.initializer])
# set data augmentation
seq = iaa.Sequential(get_aug_config(config))
# train batches
def train_step(img_batch_1, label_batch_1, img_batch_2, label_batch_2, config):
images_1, labels_1, images_2, labels_2 = sess.run(
[img_batch_1, label_batch_1, img_batch_2, label_batch_2])
if config.augmentation:
images_1 = seq.augment_images(images_1)
images_2 = seq.augment_images(images_2)
feed_dict = {
X1: images_1, Y1: labels_1,
X2: images_2, Y2: labels_2,
}
outputs = sess.run(
[train_op,
loss,
loss_crossent_logit,
loss_squared_logit,
loss_sim,
loss_sim_dist,
loss_max_neg,
global_step,
acc_logit,
acc_sim,
loss_triplet ], feed_dict)
return outputs
""" Training loop """
print("Mandoo model train start!..")
best_mAP = 0.
mAP = 0
best_mAP_step = 0
best_loss = 99999.
best_loss_step = 0
best_min_first_K = 99999.
best_min_first_K_step = 0.
start_time = time.time()
epoch = 0
while True:
try:
_, train_loss, train_loss_logit_cross, train_loss_logit_sqrt, train_loss_sim, train_loss_sim_dist, \
train_loss_max_neg, step, train_acc_logit, train_acc_sim, train_loss_triplet = train_step(
img_batch_1, label_batch_1, img_batch_2, label_batch_2, config)
prev_epoch = epoch
epoch = math.floor(step * batch_size / size_of_epoch)
# print process
if step % 30 == 0 or config.debug:
print_second = int(time.time() - start_time)
start_time = time.time()
print("[{:g} sec] epoch {:g}, step {:g}, acc_logit {:g}, acc_sim {:g}".format(
print_second, epoch, step, train_acc_logit, train_acc_sim))
print("..... total_loss {:g}, loss_logit_cross {:g}, loss_logit_sqrt {:g}, loss_sim {:g}, "
"loss_sim_dist {:g}, loss_max_neg {:g}, loss_triplet {:g}".format(
train_loss, train_loss_logit_cross, train_loss_logit_sqrt, train_loss_sim,
train_loss_sim_dist, train_loss_max_neg, train_loss_triplet))
do_save = False
if step % 150 == 0 or (config.debug and step % 1 == 0):
infer_result = local_infer(queries, references, queries_img, reference_img, batch_size)
mAP, mean_recall_at_K, min_first_1_at_K = evaluate_rank(infer_result)
if best_min_first_K >= min_first_1_at_K:
best_min_first_K = min_first_1_at_K
best_min_first_K_step = step
print("----> First_K @ 1 recall : %d / %d" % (min_first_1_at_K, len(mean_recall_at_K)))
do_save = True
if best_mAP <= mAP :
best_mAP = mAP
print("----> Best mAP : best-mAP {:g}".format(best_mAP))
do_save = True
if epoch - prev_epoch == 1:
print("----> Epoch changed saving")
do_save = True
if do_save:
# save model
nsml.report(summary=True, epoch=str(step), epoch_total=nb_epoch)
nsml.save(step)
print("Model saved : %d step" % step)
print("=============================================================================================================")
except tf.errors.OutOfRangeError:
print("finish train!")
break