-
Notifications
You must be signed in to change notification settings - Fork 83
/
train.py
226 lines (174 loc) · 8.61 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
#!/usr/bin/python3
import argparse
import os
import random
import logging
import tensorflow as tf
import numpy as np
from PIL import Image
from unet import UNet, Discriminator
from scripts.image_manips import resize
model_name = "matting"
logging.basicConfig(level=logging.INFO)
# Parse Arguments
def parse_args():
parser = argparse.ArgumentParser(description="Trains the unet")
parser.add_argument("data", type=str,
help="Path to a folder containing data to train")
parser.add_argument("--lr", type=float, default=1.0,
help="Learning rate used to optimize")
parser.add_argument("--d_coeff", type=float, default=1.0,
help="Discriminator loss coefficient")
parser.add_argument("--gen_epoch", type=int, default=4,
help="Number of training epochs")
parser.add_argument("--disc_epoch", type=int, default=1,
help="Number of training epochs")
parser.add_argument("--adv_epoch", type=int, default=5,
help="Number of training epochs")
parser.add_argument("--batch_size", dest="batch_size", type=int, default=4,
help="Size of the batches used in training")
parser.add_argument('--checkpoint', type=int, default=None,
help='Saved session checkpoint, -1 for latest.')
parser.add_argument('--logdir', default="log/" + model_name,
help='Directory where logs should be written.')
return parser.parse_args()
def apply_trimap(images, output, alpha):
masked_output = []
for channel in range(4):
masked_output.append(output[:,:,:,channel])
masked_output[channel] = tf.where(alpha < 0.25, images[:,:,:,channel], masked_output[channel])
masked_output[channel] = tf.where(alpha > 0.75, images[:,:,:,channel], masked_output[channel])
masked_output[channel] = masked_output[channel]
masked_output = tf.stack(masked_output, 3)
return masked_output
def main(args):
input_path = os.path.join(args.data, "input")
trimap_path = os.path.join(args.data, "trimap")
target_path = os.path.join(args.data, "target")
output_path = os.path.join(args.data, "output")
train_data_update_freq = args.batch_size
test_data_update_freq = 50*args.batch_size
sess_save_freq = 100*args.batch_size
if not os.path.isdir(output_path):
os.makedirs(output_path)
if not os.path.isdir(args.logdir):
os.makedirs(args.logdir)
ids = [[int(i) for i in os.path.splitext(filename)[0].split('_')] for filename in os.listdir(input_path)]
np.random.shuffle(ids)
split_point = int(round(0.85*len(ids))) #using 70% as training and 30% as Validation
train_ids = tf.get_variable('train_ids', initializer=ids[0:split_point], trainable=False)
valid_ids = tf.get_variable('valid_ids', initializer=ids[split_point:len(ids)], trainable=False)
global_step = tf.get_variable('global_step', initializer=0, trainable=False)
g_iter = int(args.gen_epoch * int(train_ids.shape[0]))
d_iter = int(args.disc_epoch * int(train_ids.shape[0]))
a_iter = int(args.adv_epoch * int(train_ids.shape[0]))
n_iter = g_iter+d_iter+a_iter
input_images = tf.placeholder(tf.float32, shape=[None, 480, 360, 4])
target_images = tf.placeholder(tf.float32, shape=[None, 480, 360, 4])
alpha = target_images[:,:,:,3][..., np.newaxis]
with tf.variable_scope("Gen"):
gen = UNet(4,4)
output = tf.sigmoid(gen(input_images))
g_loss = tf.losses.mean_squared_error(target_images, output)
with tf.variable_scope("Disc"):
disc = Discriminator(4)
d_real = disc(target_images)
d_fake = disc(output)
d_loss = tf.reduce_mean(tf.log(d_real) + tf.log(1-d_fake))
a_loss = g_loss + args.d_coeff * d_loss
g_loss_summary = tf.summary.scalar("g_loss", g_loss)
d_loss_summary = tf.summary.scalar("d_loss", d_loss)
a_loss_summary = tf.summary.scalar("a_loss", a_loss)
summary_op = tf.summary.merge(
[g_loss_summary, d_loss_summary, a_loss_summary])
summary_image = tf.summary.image("result", output)
g_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Gen')
d_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Disc')
g_optimizer = tf.train.AdadeltaOptimizer(args.lr).minimize(g_loss, global_step=global_step, var_list=g_vars)
a_optimizer = tf.train.AdadeltaOptimizer(args.lr).minimize(a_loss, global_step=global_step, var_list=g_vars)
d_optimizer = tf.train.AdadeltaOptimizer(args.lr).minimize(-d_loss, global_step=global_step, var_list=d_vars)
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
train_writer = tf.summary.FileWriter(args.logdir + '/train')
test_writer = tf.summary.FileWriter(args.logdir + '/test')
saver = tf.train.Saver()
if args.checkpoint is not None and os.path.exists(os.path.join(args.logdir, 'checkpoint')):
if args.checkpoint == -1:#latest checkpoint
saver.restore(sess, tf.train.latest_checkpoint(args.logdir))
else:#Specified checkpoint
saver.restore(sess, os.path.join(args.logdir, model_name+".ckpt-"+str(args.checkpoint)))
logging.debug('Model restored to step ' + str(global_step.eval(sess)))
train_ids = list(train_ids.eval(sess))
valid_ids = list(valid_ids.eval(sess))
def load_batch(batch_ids):
images, targets = [], []
for i, j in batch_ids:
input_filename = os.path.join(input_path, str(i) + '_' + str(j) + '.jpg')
trimap_filename = os.path.join(trimap_path, str(i) + '_trimap.jpg')
target_filename = os.path.join(target_path, str(i) + '.png')
logging.debug(input_filename)
logging.debug(trimap_filename)
logging.debug(target_filename)
image = resize(Image.open(input_filename), 2)
trimap = resize(Image.open(trimap_filename), 2)
target = resize(Image.open(target_filename), 2)
image = np.array(image)
trimap = np.array(trimap)[..., np.newaxis]
image = np.concatenate((image, trimap), axis = 2).astype(np.float32) / 255
target = np.array(target).astype(np.float32) / 255
images.append(image)
targets.append(target)
return np.asarray(images), np.asarray(targets)
def test_step(batch_idx, summary_fct):
batch_range = random.sample(train_ids, args.batch_size)
images, targets = load_batch(batch_range)
loss, demo, summary = sess.run([g_loss, summary_image, summary_fct], feed_dict={
input_images: images,
target_images: targets,
})
test_writer.add_summary(summary, batch_idx)
test_writer.add_summary(demo, batch_idx)
logging.info('Validation Loss: {:.8f}'.format(loss))
try:
batch_idx = 0
while batch_idx < n_iter:
batch_idx = global_step.eval(sess) * args.batch_size
loss_fct = None
label = None
optimizers = []
if batch_idx < g_iter:
loss_fct = g_loss
summary_fct = g_loss_summary
label = 'Gen train'
optimizers = [g_optimizer]
elif batch_idx < g_iter+d_iter:
loss_fct = d_loss
summary_fct = d_loss_summary
label = 'Disc train'
optimizers = [d_optimizer]
else:
loss_fct = a_loss
summary_fct = summary_op
label = 'Adv train'
optimizers = [a_optimizer]
batch_range = random.sample(train_ids, args.batch_size)
images, targets = load_batch(batch_range)
loss, summary = sess.run([loss_fct, summary_fct] + optimizers, feed_dict={
input_images: np.array(images),
target_images: np.array(targets)})[0:2]
if batch_idx % train_data_update_freq == 0:
logging.info('{}: [{}/{} ({:.0f}%)]\tGen Loss: {:.8f}'.format(label, batch_idx, n_iter,
100. * (batch_idx+1) / n_iter, loss))
train_writer.add_summary(summary, batch_idx)
if batch_idx % test_data_update_freq == 0:
test_step(batch_idx, summary_fct)
if batch_idx % sess_save_freq == 0:
logging.debug('Saving model')
saver.save(sess, os.path.join(args.logdir, model_name+".ckpt"), global_step=batch_idx)
except Exception:
saver.save(sess, os.path.join(args.logdir, 'crash_save_'+model_name+".ckpt"), global_step=batch_idx)
saver.save(sess, os.path.join(args.logdir, model_name+".ckpt"), global_step=batch_idx)
if __name__ == '__main__':
args = parse_args()
main(args)