-
Notifications
You must be signed in to change notification settings - Fork 19
/
keras_debugger.py
479 lines (394 loc) · 17.1 KB
/
keras_debugger.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
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
from distutils import spawn as sp
import glob
import h5py
import os
import subprocess
import shutil
import sys
class KerasDebugger:
ncnn_prog_template = \
'''
#include "net.h"
#include <algorithm>
#include <stdio.h>
#include <stdlib.h>
float rand_float() { return (float)rand() / (float)RAND_MAX; }
static int rand_mat(ncnn::Mat& m){
for(int i=0; i<3; i++){
float *ptr = m.channel(i);
for(int j=0; j<(m.w * m.h); j++){
ptr[j] = rand_float();
}
}
return 0;
}
static int dump_mat(const ncnn::Mat& m, const char *m_name)
{
char filename[1000] = "";
sprintf(filename, "%s-%d-%d-%d-layer_dump.dat", m_name, m.c, m.h, m.w);
FILE* fp = fopen(filename, "w+");
if(fp == NULL){
return -1;
}
for (int q=0; q<m.c; q++)
{
const float* ptr = m.channel(q);
fwrite(ptr, sizeof(float), m.w * m.h, fp);
}
fclose(fp);
return 0;
}
static int detect_ncnn_net()
{
ncnn::Net ncnn_net;
ncnn_net.load_param("$FILENAME$.param");
ncnn_net.load_model("$FILENAME$.bin");
ncnn::Extractor ex = ncnn_net.create_extractor();
$INSERT$
return 0;
}
int main(int argc, char** argv)
{
detect_ncnn_net();
return 0;
}
'''
input_extractor_template = '\tncnn::Mat $layer_name_rep$_blob;\n'\
'\t$layer_name_rep$_blob.create($input_shape$, 4u);\n'\
'\trand_mat($layer_name_rep$_blob);\n'\
'\tex.input("$layer_name$_blob", $layer_name_rep$_blob);\n\n'
output_extractor_template = '\tncnn::Mat $layer_name_rep$;\n'\
'\tex.extract("$layer_name$_blob", $layer_name_rep$);\n'\
'\tdump_mat($layer_name_rep$, "$layer_name$");\n\n'
tmp_dir = os.path.join('.keras2ncnn_build')
def init_env(self):
if 'win32' in sys.platform:
print('\tThe debugger currently does not support Win32.')
return -1
if not os.path.exists(os.path.join(self.tmp_dir, '.stamp_env_init')):
# Test for compiler
required_utils = [
'git',
'gcc',
'g++',
'make',
'cmake'
]
subprocess.run(['python3', '-m', 'pip', 'install',
'--upgrade', 'virtualenv==16.7.9'])
for util in required_utils:
res = sp.find_executable(util)
if res is None:
print(
'\t%s is not inside PATH, please install it first.' %
util)
return -1
# Setup virtualenv
import virtualenv # pylint: disable=import-outside-toplevel
virtualenv.create_environment(self.tmp_dir)
activator_filename = os.path.join(
self.tmp_dir, 'bin', 'activate_this.py')
exec(open(activator_filename).read(),
{'__file__': activator_filename})
install_pkg_list = [
'numpy',
'tensorflow',
'matplotlib',
'scipy'
]
# Install packages
subprocess.run(
[os.path.join(self.tmp_dir, 'bin', 'pip3'), 'install', '--upgrade', 'pip'])
subprocess.run(
[os.path.join(self.tmp_dir, 'bin', 'pip3'), 'install'] + install_pkg_list)
# Pull ncnn
subprocess.run(['git',
'clone',
'https://github.com/Tencent/ncnn.git',
os.path.join(self.tmp_dir,
'ncnn')])
# setup build
os.mkdir(os.path.join(self.tmp_dir, 'ncnn', 'build'))
open(
os.path.join(
self.tmp_dir,
'ncnn',
'benchmark',
'CMakeLists.txt'),
'a').write(
'add_executable(keras2ncnn keras2ncnn.cpp)\n' +
'target_link_libraries(keras2ncnn PRIVATE ncnn)')
open(
os.path.join(
self.tmp_dir,
'.stamp_env_init'),
'w+').write('Rua!')
else:
activator_filename = os.path.join(
self.tmp_dir, 'bin', 'activate_this.py')
exec(open(activator_filename).read(),
{'__file__': activator_filename})
return 0
def emit_file(self, file_name, ncnn_graph, keras_graph, graph_seq):
def replace_bs(x): return x.replace('/', '_')
extractor_list = []
c_payload = self.ncnn_prog_template
for layer_name in graph_seq:
layer_type = ncnn_graph.get_node_attr(layer_name)['type']
if layer_type == 'Input':
op_shape = keras_graph.get_node_attr(
layer_name)
if op_shape == None or None in op_shape:
print('Input has undetermind shape in W/H/C, default to 224,224,3')
op_shape = [3, 224, 224]
else:
op_shape = op_shape['layer']['config']['batch_input_shape'][1:]
extractor_list.append(
self.input_extractor_template.replace(
'$layer_name$',
layer_name).replace(
'$layer_name_rep$',
replace_bs(layer_name)).replace(
'$input_shape$',
', '.join(list(map(str, op_shape)))
))
extractor_list.append(
self.output_extractor_template.replace(
'$layer_name$',
layer_name).replace(
'$layer_name_rep$',
replace_bs(layer_name)))
c_payload = c_payload.replace('$FILENAME$', file_name)
c_payload = c_payload.replace('$INSERT$', '\n'.join(extractor_list))
open(
os.path.join(
self.tmp_dir,
'ncnn',
'benchmark',
'keras2ncnn.cpp'),
'w+').write(c_payload)
try:
os.mkdir(os.path.join(self.tmp_dir, 'ncnn', 'build', 'benchmark'))
except BaseException:
pass
shutil.copy(
os.path.join(
self.tmp_dir,
file_name + '.param'),
os.path.join(
self.tmp_dir,
'ncnn',
'build',
'benchmark'))
shutil.copy(
os.path.join(
self.tmp_dir,
file_name + '.bin'),
os.path.join(
self.tmp_dir,
'ncnn',
'build',
'benchmark'))
def run_debug(self):
for f in glob.glob(
os.path.join(
self.tmp_dir,
'ncnn',
'build',
'benchmark',
'*.dat')):
try:
os.remove(f)
except BaseException:
pass
try:
os.remove(
os.path.join(
self.tmp_dir,
'ncnn',
'build',
'benchmark',
'keras2ncnn'))
except BaseException:
pass
subprocess.run(
'cd %s; cmake -DNCNN_BENCHMARK=ON ..; cd benchmark; make -j' %
os.path.join(
self.tmp_dir,
'ncnn',
'build'),
shell=True)
subprocess.run(
'cd %s; chmod +x keras2ncnn; ./keras2ncnn' %
(os.path.join(
self.tmp_dir,
'ncnn',
'build',
'benchmark')),
shell=True)
def decode(self, h5_file, keras_graph, graph_seq):
import numpy as np # pylint: disable=import-outside-toplevel
from tensorflow.python import keras # pylint: disable=import-outside-toplevel
from tensorflow.python.keras import backend as K # pylint: disable=import-outside-toplevel
from tensorflow.python.keras.models import model_from_json
K.set_learning_phase(0)
from PIL import Image # pylint: disable=import-outside-toplevel
from matplotlib import pyplot as plt # pylint: disable=import-outside-toplevel
from numpy import linalg # pylint: disable=import-outside-toplevel
from scipy.spatial import distance # pylint: disable=import-outside-toplevel
# Read results from ncnn log file
def strip_filename(x): return (os.path.split(x)[-1])
def is_log_file(x): return '.dat' in x
ncnn_det_out = {}
log_files = filter(
is_log_file,
os.listdir(
os.path.join(
self.tmp_dir,
'ncnn',
'build',
'benchmark')))
for mat_file in log_files:
mat_sp = strip_filename(mat_file).split('-')
ncnn_det_out[mat_sp[0]] = np.fromfile(os.path.join(self.tmp_dir, 'ncnn', 'build', 'benchmark', mat_file),
dtype='float32').reshape(*list(map(int, mat_sp[1:4])))
# Inference using keras
f = h5py.File(h5_file, mode='r')
model_config_raw = f.attrs.get('model_config')
model = model_from_json(model_config_raw)
model.load_weights(h5_file)
output_node_names = [[], ]
output_nodes = [[], ]
input_images = []
inputs = [model.inputs]
for layer_idx in range(len(model.layers)):
if 'Model' in str(type(model.layers[layer_idx])):
output_node_names.append([])
output_nodes.append([])
# input_images.append([])
inputs.append(model.layers[layer_idx].inputs)
i = 0
while True:
try:
layer = model.layers[layer_idx].get_layer(index=i)
node_name = layer.name
if node_name in ncnn_det_out.keys():
if keras_graph.get_node_attr(
node_name)['layer']['class_name'] == 'InputLayer':
input_images.append(
ncnn_det_out[node_name].transpose(
1, 2, 0)[
np.newaxis, ...])
else:
output_node_names[-1].append(node_name)
output_nodes[-1].append(layer.output)
i = i + 1
except BaseException:
break
else:
node_name = model.layers[layer_idx].name
if node_name in ncnn_det_out.keys():
if keras_graph.get_node_attr(
node_name)['layer']['class_name'] == 'InputLayer':
input_images.append(
ncnn_det_out[node_name].transpose(
1, 2, 0)[
np.newaxis, ...])
else:
output_node_names[0].append(node_name)
output_nodes[0].append(model.layers[layer_idx].output)
keras_layer_dumps_list = []
for func_idx in range(len(output_node_names)):
functor = K.function(inputs[func_idx], output_nodes[func_idx])
layer_outs = functor(input_images + [1, ])
keras_layer_dumps_list.append(
dict(zip(output_node_names[func_idx], layer_outs)))
keras_layer_dumps = {
k: v for d in keras_layer_dumps_list for k,
v in d.items()}
output_node_names = [j for i in output_node_names for j in i]
for output_node_name in graph_seq:
if output_node_name not in output_node_names:
print(output_node_name)
continue
print('==================================')
print(
'Layer Name: %s, Layer Shape: keras->%s ncnn->%s' %
(output_node_name, str(
keras_layer_dumps[output_node_name].shape), str(
ncnn_det_out[output_node_name].shape)))
print(
'Max: \tkeras->%.03f ncnn->%.03f \tMin: keras->%.03f ncnn->%.03f' %
(keras_layer_dumps[output_node_name].flatten().max(),
ncnn_det_out[output_node_name].flatten().max(),
keras_layer_dumps[output_node_name].flatten().min(),
ncnn_det_out[output_node_name].flatten().min()))
print(
'Mean: \tkeras->%.03f ncnn->%.03f \tVar: keras->%.03f ncnn->%.03f' %
(keras_layer_dumps[output_node_name].flatten().mean(),
ncnn_det_out[output_node_name].flatten().mean(),
keras_layer_dumps[output_node_name].flatten().std(),
ncnn_det_out[output_node_name].flatten().std()))
if keras_layer_dumps[output_node_name][0].ndim == 3:
if keras_layer_dumps[output_node_name].size != ncnn_det_out[output_node_name].size:
print('Size not matched, not able to calculate similarity.')
else:
print(
'Cosine Similarity: %.05f' %
distance.cosine(
keras_layer_dumps[output_node_name][0].transpose(
(2, 0, 1)).flatten(), ncnn_det_out[output_node_name].flatten()))
print('Keras Feature Map: \t%s' % np.array2string(
keras_layer_dumps[output_node_name][0][0:10, 0, 0], suppress_small=True, precision=3))
print('Ncnn Feature Map: \t%s' % np.array2string(
ncnn_det_out[output_node_name][0, 0:10, 0], suppress_small=True, precision=3))
elif keras_layer_dumps[output_node_name][0].ndim == 2:
if keras_layer_dumps[output_node_name].size != ncnn_det_out[output_node_name].size:
print('Size not matched, not able to calculate similarity.')
else:
print(
'Cosine Similarity: %.05f' %
distance.cosine(
keras_layer_dumps[output_node_name][0].transpose(
(1, 0)).flatten(), ncnn_det_out[output_node_name].flatten()))
print('Keras Feature Map: \t%s' % np.array2string(
keras_layer_dumps[output_node_name][0][0:10, 0], suppress_small=True, precision=3))
print('Ncnn Feature Map: \t%s' % np.array2string(
ncnn_det_out[output_node_name][0, 0:10], suppress_small=True, precision=3))
elif keras_layer_dumps[output_node_name][0].ndim == 1\
and (ncnn_det_out[output_node_name].shape[:2] == (1, 1)
or ncnn_det_out[output_node_name].ndim == 1):
print(
'Cosine Similarity: %.05f' %
distance.cosine(
keras_layer_dumps[output_node_name][0].flatten(),
ncnn_det_out[output_node_name].flatten()))
print('Keras Feature Map: \t%s' % np.array2string(
keras_layer_dumps[output_node_name][0][0:10], suppress_small=True, precision=3))
if ncnn_det_out[output_node_name].ndim == 3:
print('Ncnn Feature Map: \t%s' % np.array2string(
ncnn_det_out[output_node_name][0, 0, 0:10], suppress_small=True, precision=3))
keras_index = keras_layer_dumps[output_node_name][0].argsort(
)[-5:][::-1]
keras_top_value = keras_layer_dumps[output_node_name][0][keras_index]
keras_topk = dict(zip(keras_index, keras_top_value))
if ncnn_det_out[output_node_name].ndim == 3:
ncnn_index = ncnn_det_out[output_node_name][0, 0].argsort(
)[-5:][::-1]
ncnn_top_value = ncnn_det_out[output_node_name][0,
0][ncnn_index]
ncnn_topk = dict(zip(ncnn_index, ncnn_top_value))
if os.path.exists('./ImageNetLabels.txt'):
labels = open('ImageNetLabels.txt').readlines()
keras_topk_str = ", ".join(
("%s:%.03f" % (labels[i[0] + 1].strip(), i[1]) for i in keras_topk.items()))
ncnn_topk_str = ", ".join(
("%s:%.03f" % (labels[i[0] + 1].strip(), i[1]) for i in ncnn_topk.items()))
else:
keras_topk_str = ", ".join(
("%d:%.03f" % i for i in keras_topk.items()))
ncnn_topk_str = ", ".join(
("%d:%.03f" % i for i in ncnn_topk.items()))
print(
'Top-k:\nKeras Top-k: \t%s\nncnn Top-k: \t%s' %
(keras_topk_str, ncnn_topk_str))