forked from tensorpack/tensorpack
-
Notifications
You must be signed in to change notification settings - Fork 0
/
hed.py
executable file
·230 lines (197 loc) · 7.97 KB
/
hed.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
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
# File: hed.py
# Author: Yuxin Wu <[email protected]>
import cv2
import tensorflow as tf
import argparse
import numpy as np
from six.moves import zip
import os
import sys
from tensorpack import *
from tensorpack.tfutils.symbolic_functions import *
from tensorpack.tfutils.summary import *
class Model(ModelDesc):
def _get_inputs(self):
return [InputDesc(tf.float32, [None, None, None, 3], 'image'),
InputDesc(tf.int32, [None, None, None], 'edgemap')]
def _build_graph(self, inputs):
image, edgemap = inputs
image = image - tf.constant([104, 116, 122], dtype='float32')
edgemap = tf.expand_dims(edgemap, 3, name='edgemap4d')
def branch(name, l, up):
with tf.variable_scope(name) as scope:
l = Conv2D('convfc', l, 1, kernel_shape=1, nl=tf.identity,
use_bias=True,
W_init=tf.constant_initializer(),
b_init=tf.constant_initializer())
while up != 1:
l = BilinearUpSample('upsample{}'.format(up), l, 2)
up = up / 2
return l
with argscope(Conv2D, kernel_shape=3, nl=tf.nn.relu):
l = Conv2D('conv1_1', image, 64)
l = Conv2D('conv1_2', l, 64)
b1 = branch('branch1', l, 1)
l = MaxPooling('pool1', l, 2)
l = Conv2D('conv2_1', l, 128)
l = Conv2D('conv2_2', l, 128)
b2 = branch('branch2', l, 2)
l = MaxPooling('pool2', l, 2)
l = Conv2D('conv3_1', l, 256)
l = Conv2D('conv3_2', l, 256)
l = Conv2D('conv3_3', l, 256)
b3 = branch('branch3', l, 4)
l = MaxPooling('pool3', l, 2)
l = Conv2D('conv4_1', l, 512)
l = Conv2D('conv4_2', l, 512)
l = Conv2D('conv4_3', l, 512)
b4 = branch('branch4', l, 8)
l = MaxPooling('pool4', l, 2)
l = Conv2D('conv5_1', l, 512)
l = Conv2D('conv5_2', l, 512)
l = Conv2D('conv5_3', l, 512)
b5 = branch('branch5', l, 16)
final_map = Conv2D('convfcweight',
tf.concat([b1, b2, b3, b4, b5], 3), 1, 1,
W_init=tf.constant_initializer(0.2),
use_bias=False, nl=tf.identity)
costs = []
for idx, b in enumerate([b1, b2, b3, b4, b5, final_map]):
output = tf.nn.sigmoid(b, name='output{}'.format(idx + 1))
xentropy = class_balanced_sigmoid_cross_entropy(
b, edgemap,
name='xentropy{}'.format(idx + 1))
costs.append(xentropy)
# some magic threshold
pred = tf.cast(tf.greater(output, 0.5), tf.int32, name='prediction')
wrong = tf.cast(tf.not_equal(pred, edgemap), tf.float32)
wrong = tf.reduce_mean(wrong, name='train_error')
if get_current_tower_context().is_training:
wd_w = tf.train.exponential_decay(2e-4, get_global_step_var(),
80000, 0.7, True)
wd_cost = tf.multiply(wd_w, regularize_cost('.*/W', tf.nn.l2_loss), name='wd_cost')
costs.append(wd_cost)
add_param_summary(('.*/W', ['histogram'])) # monitor W
self.cost = tf.add_n(costs, name='cost')
add_moving_summary(costs + [wrong, self.cost])
def _get_optimizer(self):
lr = get_scalar_var('learning_rate', 3e-5, summary=True)
opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
return optimizer.apply_grad_processors(
opt, [gradproc.ScaleGradient(
[('convfcweight.*', 0.1), ('conv5_.*', 5)])])
def get_data(name):
isTrain = name == 'train'
ds = dataset.BSDS500(name, shuffle=True)
class CropMultiple16(imgaug.ImageAugmentor):
def _get_augment_params(self, img):
newh = img.shape[0] // 16 * 16
neww = img.shape[1] // 16 * 16
assert newh > 0 and neww > 0
diffh = img.shape[0] - newh
h0 = 0 if diffh == 0 else self.rng.randint(diffh)
diffw = img.shape[1] - neww
w0 = 0 if diffw == 0 else self.rng.randint(diffw)
return (h0, w0, newh, neww)
def _augment(self, img, param):
h0, w0, newh, neww = param
return img[h0:h0 + newh, w0:w0 + neww]
if isTrain:
shape_aug = [
imgaug.RandomResize(xrange=(0.7, 1.5), yrange=(0.7, 1.5),
aspect_ratio_thres=0.15),
imgaug.RotationAndCropValid(90),
CropMultiple16(),
imgaug.Flip(horiz=True),
imgaug.Flip(vert=True)
]
else:
# the original image shape (321x481) in BSDS is not a multiple of 16
IMAGE_SHAPE = (320, 480)
shape_aug = [imgaug.CenterCrop(IMAGE_SHAPE)]
ds = AugmentImageComponents(ds, shape_aug, (0, 1), copy=False)
def f(m): # thresholding
m[m >= 0.50] = 1
m[m < 0.50] = 0
return m
ds = MapDataComponent(ds, f, 1)
if isTrain:
augmentors = [
imgaug.Brightness(63, clip=False),
imgaug.Contrast((0.4, 1.5)),
]
ds = AugmentImageComponent(ds, augmentors, copy=False)
ds = BatchDataByShape(ds, 8, idx=0)
ds = PrefetchDataZMQ(ds, 1)
else:
ds = BatchData(ds, 1)
return ds
def view_data():
ds = RepeatedData(get_data('train'), -1)
ds.reset_state()
for ims, edgemaps in ds.get_data():
for im, edgemap in zip(ims, edgemaps):
assert im.shape[0] % 16 == 0 and im.shape[1] % 16 == 0, im.shape
cv2.imshow("im", im / 255.0)
cv2.waitKey(1000)
cv2.imshow("edge", edgemap)
cv2.waitKey(1000)
def get_config():
logger.auto_set_dir()
dataset_train = get_data('train')
steps_per_epoch = dataset_train.size() * 40
dataset_val = get_data('val')
return TrainConfig(
dataflow=dataset_train,
callbacks=[
ModelSaver(),
ScheduledHyperParamSetter('learning_rate', [(30, 6e-6), (45, 1e-6), (60, 8e-7)]),
HumanHyperParamSetter('learning_rate'),
InferenceRunner(dataset_val,
BinaryClassificationStats('prediction', 'edgemap4d'))
],
model=Model(),
steps_per_epoch=steps_per_epoch,
max_epoch=100,
)
def run(model_path, image_path, output):
pred_config = PredictConfig(
model=Model(),
session_init=get_model_loader(model_path),
input_names=['image'],
output_names=['output' + str(k) for k in range(1, 7)])
predictor = OfflinePredictor(pred_config)
im = cv2.imread(image_path)
assert im is not None
im = cv2.resize(im, (im.shape[1] // 16 * 16, im.shape[0] // 16 * 16))
outputs = predictor([[im.astype('float32')]])
if output is None:
for k in range(6):
pred = outputs[k][0]
cv2.imwrite("out{}.png".format(
'-fused' if k == 5 else str(k + 1)), pred * 255)
else:
pred = outputs[5][0]
cv2.imwrite(output, pred * 255)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
parser.add_argument('--load', help='load model')
parser.add_argument('--view', help='view dataset', action='store_true')
parser.add_argument('--run', help='run model on images')
parser.add_argument('--output', help='fused output filename. default to out-fused.png')
args = parser.parse_args()
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
if args.view:
view_data()
elif args.run:
run(args.load, args.run, args.output)
else:
config = get_config()
if args.load:
config.session_init = get_model_loader(args.load)
config.nr_tower = max(get_nr_gpu(), 1)
SyncMultiGPUTrainer(config).train()