-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
150 lines (130 loc) · 4.58 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
import model as model
import json
import numpy
import os, os.path
import random
import skimage.io, skimage.transform
import sys
import tensorflow as tf
import time
Size = 256
MaskSize = 32
all_tiles = {}
for fname in os.listdir('images/'):
if not fname.endswith('.jpg'):
continue
label = fname.split('.jpg')[0]
parts = label.split('_')
x = int(parts[0])
y = int(parts[1])
im = skimage.io.imread('images/'+label+'.jpg')
mask = skimage.io.imread('images/'+label+'.png')
mask = skimage.transform.resize(mask, [MaskSize, MaskSize], order=0, preserve_range=True).astype('uint8')
all_tiles[(x, y)] = (im, mask)
examples = []
for ((x, y), (im, mask)) in all_tiles.items():
if numpy.count_nonzero(mask) < 32:
continue
ok = True
big_im = numpy.zeros((3*Size, 3*Size, 3), dtype='uint8')
big_mask = numpy.zeros((3*MaskSize, 3*MaskSize), dtype='uint8')
for ox in [-1, 0, 1]:
for oy in [-1, 0, 1]:
if (x+ox, y+oy) not in all_tiles:
ok = False
continue
big_im[(oy+1)*Size:(oy+2)*Size, (ox+1)*Size:(ox+2)*Size, :] = all_tiles[(x+ox, y+oy)][0]
big_mask[(oy+1)*MaskSize:(oy+2)*MaskSize, (ox+1)*MaskSize:(ox+2)*MaskSize] = all_tiles[(x+ox, y+oy)][1]
if not ok:
continue
examples.append((big_im, big_mask, (x, y)))
model_path = 'model/model'
random.shuffle(examples)
val_examples = [example for example in examples if example[2][0] < 29898]
train_examples = [example for example in examples if example[2][0] > 29898]
zero_mask = numpy.zeros((MaskSize, MaskSize), dtype='uint8')
for fname in os.listdir('images_negative/'):
im = skimage.io.imread('images_negative/'+fname)
train_examples.append((im, zero_mask))
def prepare(example):
im = example[0]
mask = example[1]
if im.shape[0] > Size:
x = random.randint(MaskSize-MaskSize//4, MaskSize+MaskSize//4)
y = random.randint(MaskSize-MaskSize//4, MaskSize+MaskSize//4)
factor = Size//MaskSize
im = im[y*factor:y*factor+Size, x*factor:x*factor+Size, :]
mask = mask[y:y+MaskSize, x:x+MaskSize]
# random rotation
rotations = random.randint(0, 3)
if rotations > 0:
im = numpy.rot90(im, k=rotations)
mask = numpy.rot90(mask, k=rotations)
return im, mask
val_prepared = [prepare(example) for example in val_examples]
# train
m = model.Model(256, 256)
session = tf.Session()
session.run(m.init_op)
batch_size = 32
best_loss = None
for epoch in range(9999):
start_time = time.time()
train_losses = []
for i in range(256):
batch = [prepare(example) for example in random.sample(train_examples, batch_size)]
_, loss = session.run([m.optimizer, m.loss], feed_dict={
m.inputs: [example[0] for example in batch],
m.targets: [example[1] for example in batch],
m.learning_rate: 1e-3,
m.is_training: True,
})
train_losses.append(loss)
train_loss = numpy.mean(train_losses)
train_time = time.time()
val_losses = []
for i in range(0, len(val_prepared), batch_size):
batch = val_prepared[i:i+batch_size]
loss = session.run(m.loss, feed_dict={
m.inputs: [example[0] for example in batch],
m.targets: [example[1] for example in batch],
m.is_training: False,
})
val_losses.append(loss)
val_loss = numpy.mean(val_losses)
val_time = time.time()
print(
'iteration {}: train_time={}, val_time={}, train_loss={}, val_loss={}/{}'.format(
epoch, int(train_time - start_time), int(val_time - train_time), train_loss, val_loss, best_loss
))
if best_loss is None or val_loss < best_loss:
best_loss = val_loss
m.saver.save(session, model_path)
def test():
for i, example in enumerate(val_prepared):
im = example[0]
mask = example[1]
output = session.run(m.outputs, feed_dict={
m.inputs: [im],
m.is_training: False,
})[0, :, :]
skimage.io.imsave('/home/ubuntu/vis/{}_im.jpg'.format(i), im)
skimage.io.imsave('/home/ubuntu/vis/{}_mask.png'.format(i), skimage.transform.resize(mask, [Size, Size], order=0, preserve_range=True).astype('uint8'))
skimage.io.imsave('/home/ubuntu/vis/{}_out.png'.format(i), (skimage.transform.resize(output, [Size, Size], order=0)*255).astype('uint8'))
def apply():
m = model.Model(2048, 2048)
session = tf.Session()
m.saver.restore(session, model_path)
im = skimage.io.imread('tile.jpg')
output = numpy.zeros((im.shape[1], im.shape[0]), dtype='uint8')
for x in range(0, im.shape[1], 2048):
for y in range(0, im.shape[0], 2048):
print(x, y)
cur_output = session.run(m.outputs, feed_dict={
m.inputs: [im[y:y+2048, x:x+2048, :]],
m.is_training: False,
})[0, :, :]
cur_output *= 255
cur_output = skimage.transform.resize(cur_output, [2048, 2048], order=0, preserve_range=True)
output[y:y+2048, x:x+2048] = cur_output.astype('uint8')
skimage.io.imsave('out.png', output)