-
Notifications
You must be signed in to change notification settings - Fork 1
/
trainer.py
298 lines (264 loc) · 15.7 KB
/
trainer.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
import tensorflow as tf
from model import model_factory
from preprocessing import preprocessing_factory
import optimizer
import os, glob
from datetime import datetime
import dataset
import visualizer
import numpy as np
import json
import cv2
from tensorflow.contrib.tensorboard.plugins import projector
from grad_cam_plus_plus import GradCamPlusPlus
NUM_DATASET_MAP = {"mnist": [60000, 10000, 10, 1], "cifar10": [50000, 10000, 10, 3], "flowers": [3320, 350, 5, 3],
"block": [4579, 510, 3, 1],
"direction": [3036, 332, 4, 1]}
def write_summary(writer, name, imgs, sess):
img_tensor = tf.convert_to_tensor(np.array(imgs))
image_summaries = tf.summary.image(name, img_tensor, len(imgs))
merged_image_summary = tf.summary.merge([image_summaries])
writer.add_summary(sess.run(merged_image_summary))
def train(conf):
if conf.dataset_name not in NUM_DATASET_MAP:
num_dataset, num_classes = dataset.make_tfrecord(conf.dataset_name, conf.dataset_dir, conf.train_fraction,
conf.num_channel,
conf.num_dataset_parallel)
if num_dataset is None:
metadata = json.load(open(os.path.join(conf.dataset_dir, "metadata")))
NUM_DATASET_MAP[conf.dataset_name] = [metadata["num_train"],
metadata["num_validation"], metadata["num_classes"],
conf.num_channel]
else:
NUM_DATASET_MAP[conf.dataset_name] = [num_dataset * conf.train_fraction,
num_dataset * (1 - conf.train_fraction), num_classes,
conf.num_channel]
num_channel = NUM_DATASET_MAP[conf.dataset_name][3]
num_classes = NUM_DATASET_MAP[conf.dataset_name][2]
is_training = tf.placeholder(tf.bool, shape=(), name="is_training")
if conf.model_name[:6] == "nasnet":
model_f = model_factory.get_network_fn(conf.model_name, num_classes, weight_decay=conf.weight_decay,
is_training=True)
else:
model_f = model_factory.get_network_fn(conf.model_name, num_classes, weight_decay=conf.weight_decay,
is_training=is_training)
model_image_size = conf.model_image_size or model_f.default_image_size
def pre_process(example_proto, training):
features = {"image/encoded": tf.FixedLenFeature((), tf.string, default_value=""),
"image/class/label": tf.FixedLenFeature((), tf.int64, default_value=0),
'image/height': tf.FixedLenFeature((), tf.int64, default_value=0),
'image/width': tf.FixedLenFeature((), tf.int64, default_value=0)
}
parsed_features = tf.parse_single_example(example_proto, features)
if conf.preprocessing_name:
image_preprocessing_fn = preprocessing_factory.get_preprocessing(conf.preprocessing_name,
is_training=training)
image = tf.image.decode_image(parsed_features["image/encoded"], num_channel)
image = tf.clip_by_value(
image_preprocessing_fn(image, model_image_size, model_image_size), -1, 1.0)
else:
image = tf.clip_by_value(tf.image.per_image_standardization(
tf.image.resize_images(tf.image.decode_jpeg(parsed_features["image/encoded"], num_channel),
[model_image_size, model_image_size])), -1., 1.0)
if len(parsed_features["image/class/label"].get_shape()) == 0:
label = tf.one_hot(parsed_features["image/class/label"], num_classes)
else:
label = parsed_features["image/class/label"]
return image, label
def train_dataset_map(example_proto):
return pre_process(example_proto, True)
def test_dataset_map(example_proto):
return pre_process(example_proto, False)
def get_model():
model_name = conf.model_name
inputs = tf.placeholder(tf.float32, shape=[None, model_image_size, model_image_size, num_channel],
name="inputs")
labels = tf.placeholder(tf.float32, shape=[None, num_classes], name="labels")
global_step = tf.Variable(0, trainable=False)
learning_rate = optimizer.configure_learning_rate(NUM_DATASET_MAP[conf.dataset_name][0], global_step, conf)
# learning_rate = tf.placeholder(tf.float32, shape=(), name="learning_rate")
conf.num_channel = num_channel
conf.num_classes = num_classes
if model_name in ["deconv", "ed", "deconv_conv"]:
logits, gen_x, gen_x_ = model_f(inputs, model_conf=conf)
class_loss_op = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
gen_loss_op = tf.log(
tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=gen_x_, logits=gen_x)))
loss_op = tf.add(class_loss_op, gen_loss_op)
ops = [class_loss_op, loss_op, gen_loss_op]
ops_key = ["class_loss_op", "loss_op", "gen_loss_op"]
else:
if model_name == "conv":
logits = model_f(inputs, model_conf=conf)
else:
logits, end_points = model_f(inputs)
if model_name == "resnet":
logits = tf.reshape(logits, [-1, num_classes])
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
ops = [loss_op]
ops_key = ["loss_op"]
if conf.use_regularizer:
weights = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
regularizer = 0
for weight in weights:
regularizer += tf.nn.l2_loss(weight)
regularizer *= conf.weight_decay
loss_op += regularizer
tf.summary.scalar('loss', loss_op)
opt = optimizer.configure_optimizer(learning_rate, conf)
train_op = opt.minimize(loss_op, global_step=global_step)
accuracy_op = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1)), tf.float32))
ops.append(tf.argmax(logits, 1))
ops_key.append("predict_idx")
tf.summary.scalar('accuracy', accuracy_op)
merged = tf.summary.merge_all()
return inputs, labels, train_op, accuracy_op, merged, ops, ops_key, logits, end_points
if not os.path.exists(conf.dataset_dir):
conf.dataset_dir = os.path.join("/home/data", conf.dataset_name)
train_filenames = glob.glob(os.path.join(conf.dataset_dir, conf.dataset_name + ("_%s*tfrecord" % conf.train_name)))
test_filenames = glob.glob(os.path.join(conf.dataset_dir, conf.dataset_name + ("_%s*tfrecord" % conf.test_name)))
inputs, labels, train_op, accuracy_op, merged, ops, ops_key, logits, end_points = get_model()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
summary_dir = os.path.join(conf.log_dir, "summary")
train_writer = tf.summary.FileWriter(summary_dir + '/train', sess.graph)
test_writer = tf.summary.FileWriter(summary_dir + '/test')
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if conf.restore_model_path and len(glob.glob(conf.restore_model_path + ".data-00000-of-00001")) > 0:
print("restore!!")
saver.restore(sess, conf.restore_model_path)
train_iterator = tf.data.TFRecordDataset(train_filenames).map(train_dataset_map,
conf.num_dataset_parallel).shuffle(
buffer_size=conf.shuffle_buffer).batch(conf.batch_size).make_initializable_iterator()
train_next = train_iterator.get_next()
test_iterator = tf.data.TFRecordDataset(test_filenames).map(test_dataset_map, conf.num_dataset_parallel).batch(
conf.batch_size).make_initializable_iterator()
test_next = test_iterator.get_next()
num_train = NUM_DATASET_MAP[conf.dataset_name][0] // conf.batch_size
num_test = NUM_DATASET_MAP[conf.dataset_name][1] // conf.batch_size
if conf.vis_epoch is not None:
config = projector.ProjectorConfig()
vis_dir = os.path.join(conf.log_dir, "embedding")
total_dataset = None
total_labels = None
total_activations = None
heatmap_imgs = {}
bb_imgs = {}
for epoch in range(conf.epoch):
train_step = 0
if conf.vis_epoch is not None and total_dataset is not None:
total_dataset = None
total_labels = None
total_activations = None
if conf.train:
sess.run(train_iterator.initializer)
total_train_accuracy = .0
inner_train_step = 0
while True:
try:
batch_xs, batch_ys = sess.run(train_next)
results = sess.run([train_op, merged, accuracy_op, ] + ops,
feed_dict={inputs: batch_xs, labels: batch_ys, is_training: True})
total_train_accuracy += results[2]
now = datetime.now().strftime('%Y/%m/%d %H:%M:%S')
if train_step % conf.summary_interval == 0:
ops_results = " ".join(list(map(lambda x: str(x), list(zip(ops_key, results[3:])))))
print(
("[%s TRAIN %d epoch, %d / %d step] accuracy: %f" % (
now, epoch, train_step, num_train, results[2])) + ops_results)
train_writer.add_summary(results[1], train_step + epoch * num_train)
train_step += 1
inner_train_step += 1
except tf.errors.OutOfRangeError:
break
if inner_train_step > 0:
print("Avg Train Accuracy : %f" % (float(total_train_accuracy) / inner_train_step))
if epoch % conf.num_save_interval == 0:
saver.save(sess, conf.log_dir + "/model_epoch_%d.ckpt" % epoch)
if conf.eval:
total_accuracy = 0
test_step = 0
sess.run(test_iterator.initializer)
while True:
try:
test_xs, test_ys = sess.run(test_next)
results = sess.run(
[merged, accuracy_op, logits] + ops,
feed_dict={inputs: test_xs, labels: test_ys, is_training: False})
total_accuracy += results[1]
now = datetime.now().strftime('%Y/%m/%d %H:%M:%S')
ops_results = " ".join(list(map(lambda x: str(x), list(zip(ops_key, results[3:])))))
print(("[%s TEST %d epoch, %d /%d step] accuracy: %f" % (
now, epoch, test_step, num_test, results[1])) + ops_results + "labels", test_ys.argmax(1))
test_writer.add_summary(results[0], test_step + (train_step + epoch * num_train))
test_step += 1
if conf.vis_epoch is not None and epoch % conf.vis_epoch == 0:
if conf.num_vis_steps >= test_step:
if conf.use_predict_of_test_for_embed_vis:
predict_y = np.zeros((conf.batch_size, num_classes))
predict_y[np.arange(conf.batch_size), results[-1]] = 1
tmp_labels = predict_y
else:
tmp_labels = test_ys
if total_dataset is None:
total_dataset = test_xs
total_labels = tmp_labels
total_activations = results[2]
else:
total_dataset = np.append(test_xs, total_dataset, axis=0)
total_labels = np.append(tmp_labels, total_labels, axis=0)
total_activations = np.append(results[2], total_activations, axis=0)
### Create CAM image
if end_points and conf.num_cam:
grad_cam_plus_plus = GradCamPlusPlus(end_points[model_f.default_logit_layer_name],
end_points[model_f.default_last_conv_layer_name],
inputs, is_training)
cam_imgs, class_indices = grad_cam_plus_plus.create_cam_imgs(sess, test_xs, results[2])
for i in range(conf.num_cam):
box_img = np.copy(test_xs[i])
# for j in range(GradCamPlusPlus.TOP3):
### Overlay heatmap
heapmap = grad_cam_plus_plus.convert_cam_2_heatmap(cam_imgs[i][0])
overlay_img = grad_cam_plus_plus.overlay_heatmap(test_xs[i], heapmap)
if test_ys[i].argmax() == results[2][i].argmax():
key = "label_%dd" % test_ys[i].argmax()
else:
key = "fail_label_%d_pred_%d" % (test_ys[i].argmax(), results[2][i].argmax())
if key not in heatmap_imgs:
heatmap_imgs[key] = []
bb_imgs[key] = []
if len(test_xs[i].shape) != 3 or test_xs[i].shape[2] != 3:
test = cv2.cvtColor(test_xs[i], cv2.COLOR_GRAY2BGR)[..., ::-1]
else:
test = test_xs[i]
heatmap_imgs[key].append(overlay_img[..., ::-1])
heatmap_imgs[key].append(test)
### Boxing
box_img = grad_cam_plus_plus.draw_rectangle(box_img, cam_imgs[i][0], [255, 0, 0])
bb_imgs[key].append(box_img)
except tf.errors.OutOfRangeError:
break
if conf.vis_epoch is not None and epoch % conf.vis_epoch == 0:
for key in heatmap_imgs:
write_summary(test_writer, "heatmap_epoch_%d_%s" % (epoch, key), heatmap_imgs[key], sess)
write_summary(test_writer, "bb_epoch_%d_%s" % (epoch, key), bb_imgs[key], sess)
heatmap_imgs = {}
bb_imgs = {}
if test_step > 0:
print("Avg Accuracy : %f" % (float(total_accuracy) / test_step))
if conf.vis_epoch is not None and epoch % conf.vis_epoch == 0:
# vis_dir = os.path.join(conf.log_dir, "embed_vis_%d" % epoch)
visualizer.add_embedding(config, sess=sess, embedding_list=[total_activations],
embedding_path=vis_dir, image_size=model_image_size,
channel=num_channel, labels=total_labels, prefix="epoch" + str(epoch))
if not conf.train:
break
### Write summary
# write_summary(test_writer, summary_names, result_imgs, sess)
if conf.vis_epoch is not None and conf.eval and total_dataset is not None:
visualizer.write_embedding(config, sess, total_dataset, embedding_path=vis_dir, image_size=model_image_size,
channel=num_channel, labels=total_labels)
sess.close()