forked from intel/caffe
-
Notifications
You must be signed in to change notification settings - Fork 0
/
convert_inception_v3.py
453 lines (402 loc) · 18.6 KB
/
convert_inception_v3.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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
from __future__ import print_function
import os.path
import re
import sys
import tarfile
import time
from datetime import datetime
# pylint: disable=unused-import,g-bad-import-order
import tensorflow.python.platform
from six.moves import urllib
import numpy as np
import tensorflow as tf
# pylint: enable=unused-import,g-bad-import-order
from tensorflow.python.platform import gfile
import h5py
import math
os.environ["GLOG_minloglevel"] ="3"
import caffe
from caffe.model_libs import *
from google.protobuf import text_format
paddings = {'VALID': [0, 0], 'SAME': [1, 1]}
FLAGS = tf.app.flags.FLAGS
# classify_image_graph_def.pb:
# Binary representation of the GraphDef protocol buffer.
# imagenet_synset_to_human_label_map.txt:
# Map from synset ID to a human readable string.
# imagenet_2012_challenge_label_map_proto.pbtxt:
# Text representation of a protocol buffer mapping a label to synset ID.
tf.app.flags.DEFINE_string(
'model_dir', '/tmp/imagenet',
"""Path to classify_image_graph_def.pb, """
"""imagenet_synset_to_human_label_map.txt, and """
"""imagenet_2012_challenge_label_map_proto.pbtxt.""")
tf.app.flags.DEFINE_string('image_file', '',
"""Absolute path to image file.""")
tf.app.flags.DEFINE_integer('num_top_predictions', 5,
"""Display this many predictions.""")
# pylint: disable=line-too-long
DATA_URL = 'http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz'
# pylint: enable=line-too-long
cur_dir = os.path.dirname(os.path.realpath(__file__))
caffe_root = '{}/../'.format(cur_dir)
labelmap_file = caffe_root + 'data/ILSVRC2016/labelmap_ilsvrc_clsloc.prototxt'
file = open(labelmap_file, 'r')
labelmap = caffe_pb2.LabelMap()
text_format.Merge(str(file.read()), labelmap)
def get_labelname(label):
num_labels = len(labelmap.item)
found = False
for i in xrange(0, num_labels):
if label == labelmap.item[i].label:
found = True
return labelmap.item[i].display_name
assert found == True
def create_graph():
"""Creates a graph from saved GraphDef file and returns a saver."""
# Creates graph from saved graph_def.pb.
with gfile.FastGFile(os.path.join(
FLAGS.model_dir, 'classify_image_graph_def.pb'), 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
with tf.device('/cpu:0'):
_ = tf.import_graph_def(graph_def, name='')
def make_padding(padding_name, conv_shape):
if padding_name == 'VALID':
return [0, 0]
elif padding_name == 'SAME':
return [int(math.ceil(conv_shape[0]/2)), int(math.ceil(conv_shape[1]/2))]
else:
sys.exit('Invalid padding name '+padding_name)
def dump_inputlayer(sess, net, operation='create'):
if operation == 'create':
resize = sess.graph.get_tensor_by_name('ResizeBilinear/size:0').eval()
[height, width] = resize
sub = sess.graph.get_tensor_by_name('Sub/y:0').eval()
mean = sub
if not type(mean) is list:
mean = [float(mean)]
else:
mean = [int(x) for x in mean]
mul = sess.graph.get_tensor_by_name('Mul/y:0').eval()
scale = float(mul)
net['data'] = L.Input(shape=dict(dim=[1, 3, int(height), int(width)]), transform_param=dict(mean_value=mean, scale=scale))
def dump_convbn(sess, net, from_layer, out_layer, operation='create'):
conv = sess.graph.get_operation_by_name(out_layer + '/Conv2D')
weights = sess.graph.get_tensor_by_name(out_layer + '/conv2d_params:0').eval()
padding = make_padding(conv.get_attr('padding'), weights.shape)
strides = conv.get_attr('strides')
beta = sess.graph.get_tensor_by_name(out_layer + '/batchnorm/beta:0').eval()
gamma = sess.graph.get_tensor_by_name(out_layer + '/batchnorm/gamma:0').eval()
mean = sess.graph.get_tensor_by_name(out_layer + '/batchnorm/moving_mean:0').eval()
std = sess.graph.get_tensor_by_name(out_layer + '/batchnorm/moving_variance:0').eval()
# TF weight matrix is of order: height x width x input_channels x output_channels
# make it to caffe format: output_channels x input_channels x height x width
weights = np.transpose(weights, (3, 2, 0, 1))
if operation == 'create':
assert from_layer in net.keys(), '{} not in net'.format(from_layer)
[num_output, channels, kernel_h, kernel_w] = weights.shape
[pad_h, pad_w] = padding
[stride_h, stride_w] = strides[1:3]
std_eps = 0.001
# parameters for convolution layer with batchnorm.
conv_prefix = ''
conv_postfix = ''
kwargs = {
'param': [dict(lr_mult=1, decay_mult=1)],
'weight_filler': dict(type='gaussian', std=0.01),
'bias_term': False,
}
conv_name = '{}{}{}'.format(conv_prefix, out_layer, conv_postfix)
if kernel_h != kernel_w:
net[conv_name] = L.Convolution(net[from_layer], num_output=num_output,
kernel_h=kernel_h, kernel_w=kernel_w, pad_h=pad_h, pad_w=pad_w,
stride_h=stride_h, stride_w=stride_w, **kwargs)
else:
net[conv_name] = L.Convolution(net[from_layer], num_output=num_output,
kernel_size=kernel_h, pad=pad_h, stride=stride_h, **kwargs)
# parameters for batchnorm layer.
bn_prefix = ''
bn_postfix = '_bn'
bn_kwargs = {
'param': [dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0), dict(lr_mult=0, decay_mult=0)],
}
bn_name = '{}{}{}'.format(bn_prefix, conv_name, bn_postfix)
net[bn_name] = L.BatchNorm(net[conv_name], in_place=True,
batch_norm_param=dict(eps=std_eps), **bn_kwargs)
# parameters for scale bias layer after batchnorm.
bias_prefix = ''
bias_postfix = '_bias'
bias_kwargs = {
'param': [dict(lr_mult=1, decay_mult=0)],
'filler': dict(type='constant', value=0.0),
}
bias_name = '{}{}{}'.format(bias_prefix, conv_name, bias_postfix)
net[bias_name] = L.Bias(net[bn_name], in_place=True, **bias_kwargs)
# relu layer.
relu_name = '{}_relu'.format(conv_name)
net[relu_name] = L.ReLU(net[conv_name], in_place=True)
elif operation == 'save':
conv_prefix = ''
conv_postfix = ''
conv_name = '{}{}{}'.format(conv_prefix, out_layer, conv_postfix)
net.params[conv_name][0].data.flat = weights.flat
# Copy bn parameters.
bn_prefix = ''
bn_postfix = '_bn'
bn_name = '{}{}{}'.format(bn_prefix, conv_name, bn_postfix)
net.params[bn_name][0].data.flat = mean
net.params[bn_name][1].data.flat = std
net.params[bn_name][2].data.flat = 1.
# Copy scale parameters.
bias_prefix = ''
bias_postfix = '_bias'
bias_name = '{}{}{}'.format(bias_prefix, conv_name, bias_postfix)
net.params[bias_name][0].data.flat = beta
def dump_pool(sess, net, from_layer, out_layer, operation='create'):
pooling = sess.graph.get_operation_by_name(out_layer)
ismax = pooling.type=='MaxPool' and 1 or 0
ksize = pooling.get_attr('ksize')
padding = make_padding(pooling.get_attr('padding'), ksize[1:3])
strides = pooling.get_attr('strides')
if operation == 'create':
if ismax:
pool = P.Pooling.MAX
else:
pool = P.Pooling.AVE
assert from_layer in net.keys()
[kernel_h, kernel_w] = ksize[1:3]
[pad_h, pad_w] = padding
[stride_h, stride_w] = strides[1:3]
if kernel_h != kernel_w:
net[out_layer] = L.Pooling(net[from_layer], pool=pool,
kernel_h=kernel_h, kernel_w=kernel_w, pad_h=pad_h, pad_w=pad_w,
stride_h=stride_h, stride_w=stride_w)
else:
net[out_layer] = L.Pooling(net[from_layer], pool=pool,
kernel_size=kernel_h, pad=pad_h, stride=stride_h)
def dump_softmax(sess, net, from_layer, out_layer, operation='create'):
softmax_w = sess.graph.get_tensor_by_name('softmax/weights:0').eval()
softmax_b = sess.graph.get_tensor_by_name('softmax/biases:0').eval()
softmax_w = np.transpose(softmax_w, (1, 0))
if operation == 'create':
assert from_layer in net.keys()
kwargs = {
'param': [dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)],
'weight_filler': dict(type='xavier'),
'bias_filler': dict(type='constant', value=0)
}
[num_output, channels] = softmax_w.shape
net[out_layer] = L.InnerProduct(net[from_layer], num_output=num_output, **kwargs)
prob_layer = '{}_prob'.format(out_layer)
net[prob_layer] = L.Softmax(net[out_layer])
elif operation == 'save':
net.params[out_layer][0].data.flat = softmax_w.flat
net.params[out_layer][1].data.flat = softmax_b
def dump_tower(sess, net, from_layer, tower_name, tower_layers, operation='create'):
for tower_layer in tower_layers:
tower_layer = '{}/{}'.format(tower_name, tower_layer)
if 'pool' in tower_layer:
dump_pool(sess, net, from_layer, tower_layer, operation)
else:
dump_convbn(sess, net, from_layer, tower_layer, operation)
from_layer = tower_layer
def dump_inception(sess, net, inception_name, tower_names, operation='create', final=True):
if operation == 'create':
towers_layers = []
for tower_name in tower_names:
tower_name = '{}/{}'.format(inception_name, tower_name)
assert tower_name in net.keys(), tower_name
towers_layers.append(net[tower_name])
if final:
inception_name = '{}/join'.format(inception_name)
net[inception_name] = L.Concat(*towers_layers, axis=1)
def run_inference_on_image(image):
if not gfile.Exists(image):
tf.logging.fatal('File does not exist %s', image)
image_data = gfile.FastGFile(image).read()
# Creates graph from saved GraphDef.
create_graph()
# sess = tf.InteractiveSession(config=tf.ConfigProto(
# allow_soft_placement=True))
sess = tf.InteractiveSession()
ops = sess.graph.get_operations()
for op in ops:
print(op.name)
# Run the graph until softmax
# start = datetime.now()
data_tensor = sess.graph.get_tensor_by_name('Mul:0')
softmax_tensor = sess.graph.get_tensor_by_name('softmax:0')
data = sess.run(data_tensor, {'DecodeJpeg/contents:0': image_data})
predictions = sess.run(softmax_tensor,
{'DecodeJpeg/contents:0': image_data})
# time_len = datetime.now() - start
# print(time_len.microseconds / 1000)
# print predictions indices and values
predictions = np.squeeze(predictions)
top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1]
for p in top_k:
print(get_labelname(p), predictions[p])
sess.close()
deploy_net_file = 'models/inception_v3/inception_v3_deploy.prototxt'
model_file = 'models/inception_v3/inception_v3.caffemodel'
net = caffe.Net(deploy_net_file, model_file, caffe.TEST)
net.blobs['data'].reshape(1, 3, 299, 299)
data = data.transpose(0, 3, 1, 2)
net.blobs['data'].data.flat = data.flat
output = net.forward()
predictions = output['softmax_prob']
predictions = np.squeeze(predictions)
top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1]
for p in top_k:
print(get_labelname(p), predictions[p])
def dump_model(operation='create', redo=False):
# Creates graph from saved GraphDef.
create_graph()
sess = tf.InteractiveSession()
# Creates caffe model.
deploy_net_file = 'models/inception_v3/inception_v3_deploy.prototxt'
model_file = 'models/inception_v3/inception_v3.caffemodel'
net = []
if operation == 'create' and (not os.path.exists(deploy_net_file) or redo):
net = caffe.NetSpec()
elif operation == 'save' and (not os.path.exists(model_file) or redo):
caffe.set_device(1)
caffe.set_mode_gpu()
net = caffe.Net(deploy_net_file, caffe.TEST)
else:
return
# dump the preprocessing parameters
dump_inputlayer(sess, net, operation)
# dump the filters
dump_convbn(sess, net, 'data', 'conv', operation)
dump_convbn(sess, net, 'conv', 'conv_1', operation)
dump_convbn(sess, net, 'conv_1', 'conv_2', operation)
dump_pool(sess, net, 'conv_2', 'pool', operation)
dump_convbn(sess, net, 'pool', 'conv_3', operation)
dump_convbn(sess, net, 'conv_3', 'conv_4', operation)
dump_pool(sess, net, 'conv_4', 'pool_1', operation)
# inceptions with 1x1, 3x3, 5x5 convolutions
from_layer = 'pool_1'
for inception_id in xrange(0, 3):
if inception_id == 0:
out_layer = 'mixed'
else:
out_layer = 'mixed_{}'.format(inception_id)
dump_tower(sess, net, from_layer, out_layer,
['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1'], operation)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1', 'conv_2'], operation)
dump_tower(sess, net, from_layer, '{}/tower_2'.format(out_layer),
['pool', 'conv'], operation)
dump_inception(sess, net, out_layer,
['conv', 'tower/conv_1', 'tower_1/conv_2', 'tower_2/conv'], operation)
from_layer = '{}/join'.format(out_layer)
# inceptions with 1x1, 3x3(in sequence) convolutions
out_layer = 'mixed_3'
dump_tower(sess, net, from_layer, out_layer,
['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1', 'conv_2'], operation)
dump_tower(sess, net, from_layer, out_layer,
['pool'], operation)
dump_inception(sess, net, out_layer,
['conv', 'tower/conv_2', 'pool'], operation)
from_layer = '{}/join'.format(out_layer)
# inceptions with 1x1, 7x1, 1x7 convolutions
for inception_id in xrange(4, 8):
out_layer = 'mixed_{}'.format(inception_id)
dump_tower(sess, net, from_layer, out_layer,
['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1', 'conv_2'], operation)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1', 'conv_2', 'conv_3', 'conv_4'], operation)
dump_tower(sess, net, from_layer, '{}/tower_2'.format(out_layer),
['pool', 'conv'], operation)
dump_inception(sess, net, out_layer,
['conv', 'tower/conv_2', 'tower_1/conv_4', 'tower_2/conv'], operation)
from_layer = '{}/join'.format(out_layer)
# inceptions with 1x1, 3x3, 1x7, 7x1 filters
out_layer = 'mixed_8'
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv', 'conv_1'], operation)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1', 'conv_2', 'conv_3'], operation)
dump_tower(sess, net, from_layer, out_layer,
['pool'], operation)
dump_inception(sess, net, out_layer,
['tower/conv_1', 'tower_1/conv_3', 'pool'], operation)
from_layer = '{}/join'.format(out_layer)
for inception_id in xrange(9, 11):
out_layer = 'mixed_{}'.format(inception_id)
dump_tower(sess, net, from_layer, out_layer,
['conv'], operation)
dump_tower(sess, net, from_layer, '{}/tower'.format(out_layer),
['conv'], operation)
dump_tower(sess, net, '{}/tower/conv'.format(out_layer),
'{}/tower/mixed'.format(out_layer), ['conv'], operation)
dump_tower(sess, net, '{}/tower/conv'.format(out_layer),
'{}/tower/mixed'.format(out_layer), ['conv_1'], operation)
dump_inception(sess, net, '{}/tower/mixed'.format(out_layer),
['conv', 'conv_1'], operation, False)
dump_tower(sess, net, from_layer, '{}/tower_1'.format(out_layer),
['conv', 'conv_1'], operation)
dump_tower(sess, net, '{}/tower_1/conv_1'.format(out_layer),
'{}/tower_1/mixed'.format(out_layer), ['conv'], operation)
dump_tower(sess, net, '{}/tower_1/conv_1'.format(out_layer),
'{}/tower_1/mixed'.format(out_layer), ['conv_1'], operation)
dump_inception(sess, net, '{}/tower_1/mixed'.format(out_layer),
['conv', 'conv_1'], operation, False)
dump_tower(sess, net, from_layer, '{}/tower_2'.format(out_layer),
['pool', 'conv'], operation)
dump_inception(sess, net, out_layer,
['conv', 'tower/mixed', 'tower_1/mixed', 'tower_2/conv'], operation)
from_layer = '{}/join'.format(out_layer)
dump_pool(sess, net, from_layer, 'pool_3', operation)
dump_softmax(sess, net, 'pool_3', 'softmax', operation)
if operation == 'create' and (not os.path.exists(deploy_net_file) or redo):
model_dir = os.path.dirname(deploy_net_file)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
with open(deploy_net_file, 'w') as f:
print('name: "inception_v3_deploy"', file=f)
print(net.to_proto(), file=f)
elif operation == 'save' and (not os.path.exists(model_file) or redo):
net.save(model_file)
sess.close()
def maybe_download_and_extract():
"""Download and extract model tar file."""
dest_directory = FLAGS.model_dir
if not os.path.exists(dest_directory):
os.makedirs(dest_directory)
filename = DATA_URL.split('/')[-1]
filepath = os.path.join(dest_directory, filename)
if not os.path.exists(filepath):
def _progress(count, block_size, total_size):
sys.stdout.write('\r>> Downloading %s %.1f%%' % (
filename, float(count * block_size) / float(total_size) * 100.0))
sys.stdout.flush()
filepath, _ = urllib.request.urlretrieve(DATA_URL, filepath,
reporthook=_progress)
print()
statinfo = os.stat(filepath)
print('Succesfully downloaded', filename, statinfo.st_size, 'bytes.')
modelfilepath = os.path.join(dest_directory, 'classify_image_graph_def.pb')
if not os.path.exists(modelfilepath):
tarfile.open(filepath, 'r:gz').extractall(dest_directory)
def main(_):
maybe_download_and_extract()
redo = True
operations = ['create', 'save']
for operation in operations:
dump_model(operation, redo)
eval = True
if eval:
image = (FLAGS.image_file if FLAGS.image_file else
os.path.join(FLAGS.model_dir, 'cropped_panda.jpg'))
run_inference_on_image(image)
if __name__ == '__main__':
tf.app.run()