-
Notifications
You must be signed in to change notification settings - Fork 17
/
Copy pathtrain.py
265 lines (226 loc) · 12.7 KB
/
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
# Author: Yuanwei Li (29 Sep 2018)
#
# Multiple landmark detection in 3D ultrasound images of fetal head
# Network training
#
# Reference
# Fast Multiple Landmark Localisation Using a Patch-based Iterative Network
# https://arxiv.org/abs/1806.06987
#
# ==============================================================================
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import numpy as np
import tensorflow as tf
from utils import input_data, shape_model_func, network, patch
class Config(object):
"""Training configurations."""
# File paths
data_dir = './data/Images'
label_dir = './data/Landmarks'
train_list_file = './data/list_train.txt'
test_list_file = './data/list_test.txt'
log_dir = './logs'
model_dir = './cnn_model'
# Shape model parameters
shape_model_file = './shape_model/shape_model/ShapeModel.mat'
eigvec_per = 0.995 # Percentage of eigenvectors to keep
sd = 3.0 # Standard deviation of shape parameters
landmark_count = 10 # Number of landmarks
landmark_unwant = [0, 8, 9, 13, 14, 15] # list of unwanted landmark indices
# Training parameters
resume = False # Whether to train from scratch or resume previous training
box_size = 101 # patch size (odd number)
alpha = 0.5 # Weighting given to the loss (0<=alpha<=1). loss = alpha*loss_c + (1-alpha)*loss_r
learning_rate = 0.001
max_steps = 100000 # Number of steps to train
save_interval = 25000 # Number of steps in between saving each model
batch_size = 64 # Training batch size
dropout = 0.5
def main():
config = Config()
# Load shape model
shape_model = shape_model_func.load_shape_model(config.shape_model_file, config.eigvec_per)
num_cnn_output_c = 2 * shape_model['Evectors'].shape[1]
num_cnn_output_r = shape_model['Evectors'].shape[1]
# Load images and landmarks
data = input_data.read_data_sets(config.data_dir,
config.label_dir,
config.train_list_file,
config.test_list_file,
config.landmark_count,
config.landmark_unwant,
shape_model)
# Build graph
print("Building graph...")
# Input placeholders
with tf.name_scope('input'):
x = tf.placeholder(tf.float32, [None, config.box_size, config.box_size, 3*config.landmark_count], name='x-input')
tf.add_to_collection('x', x)
yc_ = tf.placeholder(tf.float32, [None, num_cnn_output_c], name='yc-input') # one hot vector for classification labels (positive or negative for each shape parameter)
tf.add_to_collection('yc_', yc_)
yr_ = tf.placeholder(tf.float32, [None, num_cnn_output_r], name='yr-input') # regression output
tf.add_to_collection('yr_', yr_)
# Define CNN model
yc, yr, keep_prob = network.cnn(x, num_cnn_output_c, num_cnn_output_r)
tf.add_to_collection('yc', yc)
tf.add_to_collection('yr', yr)
tf.add_to_collection('keep_prob', keep_prob)
# Define prediction
with tf.name_scope('prediction'):
action_ind = tf.argmax(yc, 1)
tf.add_to_collection('action_ind', action_ind)
action_prob = tf.nn.softmax(yc)
tf.add_to_collection('action_prob', action_prob)
# Define loss
with tf.name_scope('loss'):
# Loss weight, alpha
alpha = tf.placeholder(tf.float32, name='alpha')
tf.add_to_collection('alpha', alpha)
# Classification loss (cross entropy)
loss_c = alpha * tf.nn.softmax_cross_entropy_with_logits(labels=yc_, logits=yc)
loss_c = tf.reduce_mean(loss_c)
tf.add_to_collection('loss_c', loss_c)
tf.summary.scalar('loss_c', loss_c)
# Regresssion loss (MSE)
loss_r = (1 - alpha) * tf.reduce_mean(tf.pow(yr - yr_, 2), axis=1)
loss_r = tf.reduce_mean(loss_r)
tf.add_to_collection('loss_r', loss_r)
tf.summary.scalar('loss_r', loss_r)
# Combined loss
loss = loss_c + loss_r
tf.add_to_collection('loss', loss)
tf.summary.scalar('loss', loss)
# Define optimizer
with tf.name_scope('train'):
# # learning rate decreases over time
# global_step = tf.Variable(0, trainable=False)
# starter_learning_rate = config.learning_rate
# learning_rate = tf.train.exponential_decay(starter_learning_rate, global_step, 250, 0.5, staircase=True)
# train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step)
# tf.summary.scalar('learning_rate', learning_rate)
# Constant learning rate
train_step = tf.train.AdamOptimizer(config.learning_rate).minimize(loss)
tf.add_to_collection('train_step', train_step)
with tf.name_scope('performance'):
# classification accuracy
correct_classification = tf.equal(tf.argmax(yc, 1), tf.argmax(yc_, 1))
correct_classification = tf.cast(correct_classification, tf.float32)
accuracy = tf.reduce_mean(correct_classification)
tf.add_to_collection('accuracy', accuracy)
tf.summary.scalar('accuracy', accuracy)
# Regression squared distance error is given by loss_r
# Run training
print("Start training...")
sess = tf.InteractiveSession()
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(config.log_dir + '/train', sess.graph)
test_writer = tf.summary.FileWriter(config.log_dir + '/test')
if config.resume:
# Resume previous training
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(config.model_dir))
saver = tf.train.Saver(max_to_keep=20)
ite_start = int(tf.train.latest_checkpoint(config.model_dir).split('-')[-1])
ite_end = ite_start + config.max_steps
else:
# Start new training
tf.global_variables_initializer().run()
saver = tf.train.Saver(max_to_keep=20)
ite_start = 0
ite_end = config.max_steps
for i in xrange(ite_start, ite_end):
patches_train, actions_train, dbs_train, _ = get_train_pairs(config.batch_size,
data.train.images,
data.train.shape_params,
config.box_size,
num_cnn_output_c,
num_cnn_output_r,
shape_model,
config.sd)
if i % 1000 == 0:
# Record summaries and test-set loss and accuracy
patches_test, actions_test, dbs_test, _ = get_train_pairs(config.batch_size,
data.test.images,
data.test.shape_params,
config.box_size,
num_cnn_output_c,
num_cnn_output_r,
shape_model,
config.sd)
summary_test, l_test, lc_test, lr_test, acc_test = sess.run([merged, loss, loss_c, loss_r, accuracy], feed_dict={x: patches_test,
yc_: actions_test,
yr_: dbs_test,
keep_prob: 1.0,
alpha: config.alpha})
test_writer.add_summary(summary_test, i)
# Record summaries and train-set loss and accuracy
summary_train, l_train, lc_train, lr_train, acc_train = sess.run([merged, loss, loss_c, loss_r, accuracy], feed_dict={x: patches_train,
yc_: actions_train,
yr_: dbs_train,
keep_prob: 1.0,
alpha: config.alpha})
train_writer.add_summary(summary_train, i)
print('Step {}: \ttrain: loss={:11.6f}, loss_c={:9.6f}, loss_r={:11.6f}, acc={:8.6f}. \ttest: loss={:11.6f}, loss_c={:9.6f}, loss_r={:11.6f}, acc={:8.6f}.'.format
(i, l_train, lc_train, lr_train, acc_train, l_test, lc_test, lr_test, acc_test))
# Train one step
_ = sess.run(train_step, feed_dict={x: patches_train,
yc_: actions_train,
yr_: dbs_train,
keep_prob: config.dropout,
alpha: config.alpha})
# Save trained model
if ((i+1) % config.save_interval) == 0:
saver.save(sess, os.path.join(config.model_dir, 'model'), global_step=i+1)
print("Trained model save successfully in {} at step {}".format(os.path.join(config.model_dir, 'model'), i+1))
train_writer.close()
test_writer.close()
sess.close()
def get_train_pairs(batch_size, images, bs_gt, box_size, num_actions, num_regression_output, shape_model, sd):
"""Randomly sample image patches and corresponding ground truth classification and regression outputs.
Args:
batch_size: mini batch size
images: list of img_count images. Each image is [width, height, depth, channel], [x,y,z,channel]
bs_gt: Ground truth shape parameters. [img_count, num_shape_params]
box_size: size of image patch. Scalar.
num_actions: number of classification outputs
num_regression_output: number of regression outputs
shape_model: structure containing shape models
sd: standard deviation of shape model. Bounds from which to sample bs.
Returns:
patches: 2D image patches, [batch_size, box_size, box_size, 3*num_landmarks]
actions: Ground truth classification output. [batch_size, num_actions], each row is a one hot vector [positive or negative for each shape parameter]
dbs: Ground truth regression output. [batch_size, num_regression_output]. dbs = bs - bs_gt.
bs: sampled shape parameters [batch_size, num_regression_output]
"""
img_count = len(images)
num_landmarks = shape_model['Evectors'].shape[0] / 3
box_r = int((box_size - 1) / 2)
patches = np.zeros((batch_size, box_size, box_size, int(3*num_landmarks)), np.float32)
actions_ind = np.zeros(batch_size, dtype=np.uint16)
actions = np.zeros((batch_size, num_actions), np.float32)
# get image indices randomly for a mini-batch
ind = np.random.randint(img_count, size=batch_size)
# Randomly sample shape parameters, bs
bounds = sd * np.sqrt(shape_model['Evalues'])
bs = np.random.rand(batch_size, num_regression_output) * 2 * bounds - bounds
# Convert shape parameters to landmark
landmarks = shape_model_func.b2landmarks(bs, shape_model)
# Extract image patch
for i in xrange(batch_size):
image = images[ind[i]]
patches[i] = patch.extract_patch_all_landmarks(image, landmarks[i], box_r)
# Regression values (distances between predicted and GT shape parameters)
dbs = bs - bs_gt[ind]
# Extract classification labels as a one-hot vector
max_db_ind = np.argmax(np.abs(dbs), axis=1) # [batch_size]
max_db = dbs[np.arange(dbs.shape[0]), max_db_ind] # [batch_size]
is_positive = (max_db > 0)
actions_ind[is_positive] = max_db_ind[is_positive] * 2
actions_ind[np.logical_not(is_positive)] = max_db_ind[np.logical_not(is_positive)] * 2 + 1
actions[np.arange(batch_size), actions_ind] = 1
return patches, actions, dbs, bs
if __name__ == '__main__':
main()