/
classification2ssd.py
225 lines (174 loc) · 7.87 KB
/
classification2ssd.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
""" Credit to https://stackoverflow.com/questions/41161021/how-to-convert-a-dense-layer-to-an-equivalent-convolutional-layer-in-keras """
import h5py
from os.path import split, join, splitext
from argparse import ArgumentParser
from keras.models import Sequential
from keras.layers.convolutional import Convolution2D
from keras.engine import InputLayer
import keras
import numpy as np
from scipy.ndimage import zoom
from jpeg_deep.networks import vgga, vggd
from jpeg_deep.networks import vgga_dct, vggd_dct, vggd_dct_deconv
from jpeg_deep.networks import vgga_dct_conv, vggd_dct_conv, vggd_dct_deconv_conv
def to_fully_conv(model):
""" This methods converts a sequential model to a fully convolutional model."""
new_model = Sequential()
input_layer = InputLayer(input_shape=(None, None, 3), name="input_new")
new_model.add(input_layer)
for layer in model.layers:
if "Flatten" in str(layer):
flattened_ipt = True
f_dim = layer.input_shape
elif "Dense" in str(layer):
input_shape = layer.input_shape
output_dim = layer.get_weights()[1].shape[0]
W, b = layer.get_weights()
if flattened_ipt:
shape = (f_dim[1], f_dim[2], f_dim[3], output_dim)
new_W = W.reshape(shape)
new_layer = Convolution2D(output_dim,
(f_dim[1], f_dim[2]),
strides=(1, 1),
activation=layer.activation,
padding='valid',
weights=[new_W, b])
flattened_ipt = False
else:
shape = (1, 1, input_shape[1], output_dim)
new_W = W.reshape(shape)
new_layer = Convolution2D(output_dim,
(1, 1),
strides=(1, 1),
activation=layer.activation,
padding='valid',
weights=[new_W, b])
else:
new_layer = layer
new_model.add(new_layer)
return new_model
def to_fully_conv_dct(model, model_type):
if model_type == "vgg16":
new_model = vggd_dct_conv()
else:
new_model = vggd_dct_deconv_conv()
for layer in model.layers:
if "Dense" not in str(layer) and "Flatten" not in str(layer):
for layer_new in new_model.layers:
if layer_new.name == layer.name:
print("Setting layer : {}".format(layer_new.name))
layer_new.set_weights(layer.get_weights())
break
elif "Dense" in str(layer):
for layer_new in new_model.layers:
if layer_new.name == "conv2d_1" and layer.name == "fc1":
print("Setting layer : {}".format(layer_new.name))
W, b = layer.get_weights()
new_W = W.reshape((7, 7, 512, 4096))
layer_new.set_weights([new_W, b])
break
elif layer_new.name == "conv2d_2" and layer.name == "fc2":
print("Setting layer : {}".format(layer_new.name))
W, b = layer.get_weights()
new_W = W.reshape((1, 1, 4096, 4096))
layer_new.set_weights([new_W, b])
break
elif layer_new.name == "conv2d_3" and layer.name == "predictions":
print("Setting layer : {}".format(layer_new.name))
W, b = layer.get_weights()
new_W = W.reshape((1, 1, 4096, 1000))
layer_new.set_weights([new_W, b])
break
return new_model
parser = ArgumentParser(
description="Converts classification weights to fully convolutional network and to SSD backbone.")
parser.add_argument(
"mt", help="The type of the model to convert, for now one of vgg16/deconv.", type=str)
parser.add_argument(
"wp", help="The weights to be converted. The output will be saved in the same directory as the original weights.", type=str)
parser.add_argument(
"-dct", help="If the network is a dct version of the vgg16 or not.", action="store_true")
args = parser.parse_args()
if args.mt == "vgg16":
if args.dct:
model = vggd_dct(1000)
else:
model = vggd(1000)
elif args.mt == "deconv":
model = vggd_dct_deconv(1000)
else:
raise RuntimeError(
"Expected one of vgg16/deconv for mt argument but {} found".format(args.mt))
model.load_weights(args.wp)
if args.dct or args.mt == "deconv":
new_model = to_fully_conv_dct(model, args.mt)
else:
new_model = to_fully_conv(model)
# Extract saved name and resave a fully conv
path, file_name = split(args.wp)
file_name_conv = join(path, splitext(file_name)[0] + "_conv.h5")
file_name_ssd = join(path, splitext(file_name)[0] + "_ssd.h5")
new_model.save(file_name_conv)
# Also save as SSD ready
new_model.save(file_name_ssd)
with h5py.File(file_name_ssd, 'a') as f:
if not (args.dct or flag):
del f['model_weights']['dropout_1']
del f['model_weights']['dropout_2']
del f['model_weights']['conv2d_3']
# Removing old names
names = f['model_weights'].attrs["layer_names"]
if not args.dct:
idx = np.where(names == b'dropout_1')
names = np.delete(names, idx[0])
idx = np.where(names == b'dropout_2')
names = np.delete(names, idx[0])
idx = np.where(names == b'conv2d_3')
names = np.delete(names, idx[0])
idx = np.where(names == b'conv2d_1')
names[idx] = b'fc6'
idx = np.where(names == b'conv2d_2')
names[idx] = b'fc7'
f['model_weights'].attrs["layer_names"] = names
# Rename the groups
f['model_weights']['fc6'] = f['model_weights']['conv2d_1']
f['model_weights']['fc7'] = f['model_weights']['conv2d_2']
f['model_weights']['fc6']['fc6'] = f['model_weights']['conv2d_1']['conv2d_1']
f['model_weights']['fc7']['fc7'] = f['model_weights']['conv2d_2']['conv2d_2']
del f['model_weights']['conv2d_1']
del f['model_weights']['conv2d_2']
del f['model_weights']['fc6']['conv2d_1']
del f['model_weights']['fc7']['conv2d_2']
fc6_kernel = f['model_weights']['fc6']['fc6']['kernel:0']
fc6_bias = f['model_weights']['fc6']['fc6']['bias:0']
fc7_kernel = f['model_weights']['fc7']['fc7']['kernel:0']
fc7_bias = f['model_weights']['fc7']['fc7']['bias:0']
fc6_numpy = fc6_kernel.value
fc7_numpy = fc7_kernel.value
fc6_bias_numpy = fc6_bias.value
fc7_bias_numpy = fc7_bias.value
del f['model_weights']['fc6']['fc6']['kernel:0']
del f['model_weights']['fc7']['fc7']['kernel:0']
del f['model_weights']['fc6']['fc6']['bias:0']
del f['model_weights']['fc7']['fc7']['bias:0']
# Select 1024 randomly for fc6
fc6_numpy = fc6_numpy[:, :, :, ::4]
fc6_bias_numpy = fc6_bias_numpy[::4]
fc6_numpy = zoom(fc6_numpy, (0.42857, 0.42857, 1, 1))
# Select 1024 randomly for fc7
fc7_numpy = fc7_numpy[:, :, :, ::4]
fc7_bias_numpy = fc7_bias_numpy[::4]
fc7_numpy = fc7_numpy[:, :, ::4, :]
# Reset the new kernels
f.create_dataset('model_weights/fc6/fc6/kernel:0', data=fc6_numpy)
f.create_dataset('model_weights/fc7/fc7/kernel:0', data=fc7_numpy)
f.create_dataset('model_weights/fc6/fc6/bias:0', data=fc6_bias_numpy)
f.create_dataset('model_weights/fc7/fc7/bias:0', data=fc7_bias_numpy)
fc7_names = f['model_weights']['fc7'].attrs['weight_names']
fc7_names[0] = b'fc7/kernel:0'
fc7_names[1] = b'fc7/bias:0'
f['model_weights']['fc7'].attrs['weight_names'] = fc7_names
fc6_names = f['model_weights']['fc6'].attrs['weight_names']
fc6_names[0] = b'fc6/kernel:0'
fc6_names[1] = b'fc6/bias:0'
f['model_weights']['fc6'].attrs['weight_names'] = fc6_names