forked from isht7/pytorch-deeplab-resnet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
convert_deeplab_resnet.py
285 lines (212 loc) · 8.52 KB
/
convert_deeplab_resnet.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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
import numpy as np
import os
os.environ["GLOG_minloglevel"] = "2"
import sys
sys.path.insert(0,'/data1/ravikiran/SketchObjPartSegmentation/src/caffe-switch/caffe/python')
#sys.path.insert(0,'/data1/ravikiran/deeplab-public-ver2/python')
import re
import caffe
import numpy as np
import skimage.io
import torch
from torch.autograd import Variable
import torchvision.models as models
import torch.nn.functional as F
import torch.nn as nn
import deeplab_resnet
from collections import OrderedDict
import matplotlib.pyplot as plt
import cv2
class CaffeParamProvider():
def __init__(self, caffe_net):
self.caffe_net = caffe_net
def conv_kernel(self, name):
k = self.caffe_net.params[name][0].data
return k
def conv_biases(self, name):
k = self.caffe_net.params[name][1].data
return k
def bn_gamma(self, name):
return self.caffe_net.params[name][0].data
def bn_beta(self, name):
return self.caffe_net.params[name][1].data
def bn_mean(self, name):
return (self.caffe_net.params[name][0].data/self.caffe_net.params[name][2].data)
def bn_variance(self, name):
return (self.caffe_net.params[name][1].data/self.caffe_net.params[name][2].data)
def fc_weights(self, name):
w = self.caffe_net.params[name][0].data
#w = w.transpose((1, 0))
return w
def fc_biases(self, name):
b = self.caffe_net.params[name][1].data
return b
def preprocess(out):
#"""Changes RGB [0,1] valued image to BGR [0,255] with mean subtracted."""
#out = np.copy(img) * 255.0
out = out[:, :, [2, 1, 0]] # swap channel from RGB to BGR
out[0] -= 104.008
out[1] -= 116.669
out[2] -= 122.675
return out
def assert_almost_equal(caffe_tensor, th_tensor):
t = th_tensor[0]
c = caffe_tensor[0]
if t.shape != c.shape:
print "t.shape", t.shape
print "c.shape", c.shape
d = np.linalg.norm(t - c)
print "d", d
assert d < 500
def dist_(caffe_tensor, th_tensor):
t = th_tensor[0]
c = caffe_tensor[0]
if t.shape != c.shape:
print "t.shape", t.shape
print "c.shape", c.shape
d = np.linalg.norm(t - c)
print "d", d
# returns image of shape [321, 321, 3]
# [height, width, depth]
def load_image(path, size=321):
img = cv2.imread(path)
resized_img = cv2.resize(img,(size,size)).astype(float)
return resized_img
def load_caffe(img_p):
caffe.set_mode_cpu()
#caffe.set_device(0)
prototxt = "data/test.prototxt"
caffemodel = "data/train_iter_20000.caffemodel"
net = caffe.Net(prototxt,caffemodel, caffe.TEST)
net.blobs['data'].data[0] = img_p.transpose((2, 0, 1))
assert net.blobs['data'].data[0].shape == (3, 321, 321)
net.forward()
return net
def parse_pth_varnames(p, pth_varname, num_layers):
# this function has been modified to fix #4
post = ''
EXP = 'voc12'
if ('weight' in pth_varname and 'conv2d_list' in pth_varname):
# #print ('res%d%s_branch%d%s'+post) % x
if len(post)!=0:
post = post[1:]+'_'
y = (EXP,int(pth_varname[25]))
return p.conv_kernel(('fc1_%s_'+post+ 'c%d') % y)
if ('bias' in pth_varname and 'conv2d_list' in pth_varname):
# #print ('res%d%s_branch%d%s'+post) % x
if len(post)!=0:
post = post[1:]+'_'
y = (EXP,int(pth_varname[25]))
return p.conv_biases(('fc1_%s_'+post+'c%d') % y)
if pth_varname == 'Scale.conv1.weight':
return p.conv_kernel('conv1'+post)
elif pth_varname == 'Scale.bn1.weight':
return p.bn_gamma('scale_conv1'+post)
elif pth_varname == 'Scale.bn1.bias':
return p.bn_beta('scale_conv1'+post)
elif pth_varname == 'Scale.bn1.running_mean':
return p.bn_mean('bn_conv1'+post)
elif pth_varname == 'Scale.bn1.running_var':
return p.bn_variance('bn_conv1'+post)
elif pth_varname == 'fc.weight':
return p.fc_weights('fc1000')
elif pth_varname == 'fc.bias':
return p.fc_biases('fc1000')
re1 = 'Scale.layer(\d+).(\d+).(downsample|conv1|bn1|conv2|bn2|conv3|bn3)'
#re1 = 'scale(\d+)/block(\d+)/(shortcut|a|b|c|A|B)'
m = re.search(re1, pth_varname)
def letter(i):
return chr(ord('a') + i - 1)
scale_num = int(m.group(1)) + 1
block_num = int(m.group(2)) + 1
if scale_num == 2:
# scale 2 uses block letters
block_str = letter(block_num)
elif scale_num == 3 or scale_num == 4:
# scale 3 uses numbered blocks
# scale 4 uses numbered blocks
if num_layers == 50:
block_str = letter(block_num)
else:
if block_num == 1:
block_str = 'a'
else:
block_str = 'b%d' % (block_num - 1)
elif scale_num == 5:
# scale 5 uses block letters
block_str = letter(block_num)
else:
raise ValueError("unexpected scale_num %d" % scale_num)
branch = m.group(3)
if branch == "downsample":
branch_num = 1
conv_letter = ''
else:
branch_num = 2
conv_letter = letter(int(branch[-1]))
x = (scale_num, block_str, branch_num, conv_letter)
if ('weight' in pth_varname and 'conv' in pth_varname) or 'downsample.0.weight' in pth_varname:
return p.conv_kernel(('res%d%s_branch%d%s'+post) % x)
if ('weight' in pth_varname and 'bn' in pth_varname) or 'downsample.1.weight' in pth_varname:
return p.bn_gamma(('scale%d%s_branch%d%s'+post) % x)
if ('bias' in pth_varname and 'bn' in pth_varname) or 'downsample.1.bias' in pth_varname:
return p.bn_beta(('scale%d%s_branch%d%s'+post) % x)
if ('running_mean' in pth_varname and 'bn' in pth_varname) or 'downsample.1.running_mean' in pth_varname:
return p.bn_mean(('bn%d%s_branch%d%s'+post) % x)
if ('running_var' in pth_varname and 'bn' in pth_varname) or 'downsample.1.running_var' in pth_varname:
return p.bn_variance(('bn%d%s_branch%d%s'+post) % x)
raise ValueError('unhandled var ' + pth_varname)
def checkpoint_fn(layers):
return 'resnet%d.pth' % layers
def convert(img_p, layers):
caffe_model = load_caffe(img_p)
param_provider = CaffeParamProvider(caffe_model)
model = deeplab_resnet.Res_Deeplab(21)
old_dict = model.state_dict()
new_state_dict = OrderedDict()
keys = model.state_dict().keys()
for var_name in keys[:]:
data = parse_pth_varnames(param_provider, var_name, layers)
new_state_dict[var_name] = torch.from_numpy(data).float()
model.load_state_dict(new_state_dict)
o = []
def hook(module, input, output):
#print module
o.append(input[0].data.numpy())
model.Scale.conv1.register_forward_hook(hook) #0, data
model.Scale.bn1.register_forward_hook(hook) #1 conv1 out
model.Scale.relu.register_forward_hook(hook) #2 batch norm out
model.Scale.maxpool.register_forward_hook(hook) #3 bn1, relu out
model.Scale.layer1._modules['0'].conv1.register_forward_hook(hook) #4, pool1 out
model.Scale.layer1._modules['1'].conv1.register_forward_hook(hook) #5, res2a out
model.Scale.layer5.conv2d_list._modules['0'].register_forward_hook(hook) #6, res5c out
model.eval()
output = model(Variable(torch.from_numpy(img_p[np.newaxis, :].transpose(0,3,1,2)).float(),volatile=True))
interp = nn.UpsamplingBilinear2d(size=(321, 321))
output_temp = interp(output[3]).cpu().data[0].numpy()
output_temp = output_temp.transpose(1,2,0)
output_temp = np.argmax(output_temp,axis = 2)
#plt.imshow(output_temp)
#plt.show()
dist_(caffe_model.blobs['data'].data,o[0])
dist_(caffe_model.blobs['conv1'].data,o[3])
dist_(caffe_model.blobs['pool1'].data,o[4])
dist_(caffe_model.blobs['res2a'].data,o[5])
dist_(caffe_model.blobs['res5c'].data,o[6])
dist_(caffe_model.blobs['fc1_voc12'].data,output[0].data.numpy())
dist_(caffe_model.blobs['fc1_voc12_res075_interp'].data,output[1].data.numpy())
dist_(caffe_model.blobs['fc1_voc12_res05'].data,output[2].data.numpy())
dist_(caffe_model.blobs['fc_fusion'].data,output[3].data.numpy())
print 'input image shape',img_p[np.newaxis, :].transpose(0,3,1,2).shape
print 'output shapes -'
for a in output:
print a.data.numpy().shape
torch.save(model.state_dict(),'data/MS_DeepLab_resnet_trained_VOC.pth')
def main():
img = load_image("data/cat.jpg")
#img = load_image("data/2007_000033.jpg")
img_p = preprocess(img)
print "CONVERTING Multi-scale DeepLab_resnet"
convert(img_p, layers = 101)
if __name__ == '__main__':
main()