forked from dhlab-epfl/dhSegment
-
Notifications
You must be signed in to change notification settings - Fork 0
/
edit_checkpoint.py
153 lines (129 loc) · 6.36 KB
/
edit_checkpoint.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
from tensorflow.python import pywrap_tensorflow
from tensorflow.python.tools.inspect_checkpoint import print_tensors_in_checkpoint_file
import numpy as np
import tensorflow as tf
import os
flags = tf.app.flags
flags.DEFINE_string('input_path',
'pretrained_models/resnet_v1_50.ckpt',
'path of pretrained_checkpoint')
flags.DEFINE_string('output_path',
'pretrained_models/msbin_resnet_zero_init.ckpt',
'output checkpoint')
flags.DEFINE_string('feature_extractor', 'resnet_v1_50',
'name of first checkpoint')
flags.DEFINE_string(
'num_input_channels', "12",
'number of input channel. Each image, background, diff image require 3 channels'
)
flags.DEFINE_string(
'edit_method', 'spread',
'divide the checkpoint convolution variable by the number of channels'
' divided by 3 and clone it to every set of 3 channels. random: initialize'
' extra channels feature map to random truncated_normal with sttdev=0.2'
'. clone: clone the value to new channels')
# flags = tf.app.flags
# flags.DEFINE_string('input_path', 'pretrained_models/model_try/resnet_v1_50.ckpt',
# 'path of pretrained_checkpoint')
# flags.DEFINE_string('output_path', 'pretrained_models/model_try_output/resnet_mod.ckpt', 'output checkpoint')
# flags.DEFINE_string('feature_extractor', 'resnet_v1_50', 'name of first checkpoint')
# flags.DEFINE_string('num_input_channels', '6', 'number of input channel. Each image, background, diff image require 3 channels')
# flags.DEFINE_string('edit_method', 'spread', 'divide the checkpoint convolution variable by the number of channels'
# ' divided by 3 and clone it to every set of 3 channels. random: initialize'
# ' extra channels feature map to random truncated_normal with sttdev=0.2'
# '. clone: clone the value to new channels')
FLAGS = flags.FLAGS
if __name__ == '__main__':
reader = pywrap_tensorflow.NewCheckpointReader(FLAGS.input_path)
# print_tensors_in_checkpoint_file('pretrained_models/model_try/resnet_v1_101.ckpt', 'all_tensors')
# print_tensors_in_checkpoint_file('pretrained_models/model_try/resnet_v1_101.ckpt', all_tensors=True, tensor_name='')
os.environ['CUDA_VISIBLE_DEVICES'] = "3"
var_to_shape_map = reader.get_variable_to_shape_map()
var_to_edit_names = [
'resnet_v1_50/conv1/weights'.format(FLAGS.feature_extractor),
'resnet_v1_50/mean_rgb'.format(FLAGS.feature_extractor),
]
print('Loading checkpoint...')
# for key in sorted(var_to_shape_map):
# print("Found variable: {}".format(key))
for key in sorted(var_to_shape_map):
if key not in var_to_edit_names:
var = tf.Variable(
reader.get_tensor(key), name=key, dtype=tf.float32)
else:
print("Found variable: {}".format(key))
vars_to_edit = []
for name in var_to_edit_names:
if reader.has_tensor(name):
vars_to_edit.append(reader.get_tensor(name))
else:
raise Exception(
"{} not found in checkpoint. Check feature extractor name. Exiting."
.format(name))
new_vars = []
sess = tf.Session()
i = 0
for name, var_to_edit in zip(var_to_edit_names, vars_to_edit):
if FLAGS.edit_method in ['spread', 'clone']:
print('shape: ' + str(var_to_edit.shape))
if i == 0:
checkpoint_num_input_channels = var_to_edit.shape[2]
i = i+1
# if FLAGS.num_input_channels % checkpoint_num_input_channels != 0:
# raise Exception('For spread edit method, num_input_channels must be divisible by num input channels of checkpoint!')
num_clones = int(
int(FLAGS.num_input_channels) / checkpoint_num_input_channels)
# print('var_to_edit: ' + var_to_edit)
# print(var_to_edit)
if FLAGS.edit_method == 'spread':
spreaded_var = var_to_edit / num_clones
else:
spreaded_var = var_to_edit
print('spreaded_var: ' + str(spreaded_var.shape))
print('num_clones: ' + str(num_clones))
new_var = np.tile(spreaded_var, [1, 1, num_clones, 1])
new_var = np.zeros(new_var.shape)
# new_var = np.tile(spreaded_var, [1, 1, num_clones, 1])
# print(new_var.shape)
# red_channel = new_var[:,:,0:1,:]
# green_channel = new_var[:,:,1:2,:]
# blue_channel = new_var[:,:,2:3,:]
# # vis light:
# new_var[:,:,0:1,:] = green_channel
# # 365 nm:
# new_var[:,:,1:2,:] = blue_channel
# # 450 nm:
# new_var[:,:,2:3,:] = blue_channel
# # 465 nm:
# new_var[:,:,3:4,:] = blue_channel
# # 505 nm:
# new_var[:,:,4:5,:] = green_channel
# # 535 nm:
# new_var[:,:,5:6,:] = green_channel
# # 570 nm:
# new_var[:,:,6:7,:] = red_channel
# # 625 nm:
# new_var[:,:,7:8,:] = red_channel
# # 700 nm:
# new_var[:,:,8:9,:] = red_channel
# # 780 nm:
# new_var[:,:,9:10,:] = red_channel
# # 870 nm:
# new_var[:,:,10:11,:] = red_channel
# # 940 nm:
# new_var[:,:,11:12,:] = red_channel
print('msi-shape: ' + str(new_var.shape))
new_vars.append(tf.Variable(new_var, name=name, dtype=tf.float32))
elif FLAGS.edit_method == 'random':
random_shape = list(var_to_edit.shape)
random_shape[2] = FLAGS.num_input_channels - 3
random_var = tf.truncated_normal(
shape=random_shape, stddev=0.01).eval(session=sess)
new_var = np.concatenate([var_to_edit, random_var], axis=2)
new_vars.append(tf.Variable(new_var, name=name, dtype=tf.float32))
else:
raise Exception("Edit method must be spread or zeros or clone!")
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.save(sess, FLAGS.output_path)
#Only need .0000-of-0001 and .index file. Good to go!