aboutsummaryrefslogtreecommitdiffstats
path: root/tools/python/convert_from_tensorflow.py
blob: 38e64c1c941a06a31996a8fdda85cc60364811ac (plain) (blame)
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
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
# Copyright (c) 2019 Guo Yejun
#
# This file is part of FFmpeg.
#
# FFmpeg is free software; you can redistribute it and/or
# modify it under the terms of the GNU Lesser General Public
# License as published by the Free Software Foundation; either
# version 2.1 of the License, or (at your option) any later version.
#
# FFmpeg is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
# Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public
# License along with FFmpeg; if not, write to the Free Software
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
# ==============================================================================

import tensorflow as tf
import numpy as np
import sys, struct
import convert_header as header

__all__ = ['convert_from_tensorflow']

class Operand(object):
    IOTYPE_INPUT = 1
    IOTYPE_OUTPUT = 2
    IOTYPE_INTERMEDIATE = IOTYPE_INPUT | IOTYPE_OUTPUT
    DTYPE_FLOAT = 1
    DTYPE_UINT8 = 4
    index = 0
    def __init__(self, name, dtype, dims):
        self.name = name
        self.dtype = dtype
        self.dims = dims
        self.iotype = 0
        self.used_count = 0
        self.index = Operand.index
        Operand.index = Operand.index + 1
        self.iotype2str = {Operand.IOTYPE_INPUT: 'in', Operand.IOTYPE_OUTPUT: 'out', Operand.IOTYPE_INTERMEDIATE: 'inout'}
        self.dtype2str = {Operand.DTYPE_FLOAT: 'DT_FLOAT', Operand.DTYPE_UINT8: 'DT_UINT8'}

    def add_iotype(self, iotype):
        self.iotype = self.iotype | iotype
        if iotype == Operand.IOTYPE_INPUT:
            self.used_count = self.used_count + 1

    def __str__(self):
        return "{}: (name: {}, iotype: {}, dtype: {}, dims: {}, used_count: {})".format(self.index,
                            self.name, self.iotype2str[self.iotype], self.dtype2str[self.dtype],
                            self.dims, self.used_count)

    def __lt__(self, other):
        return self.index < other.index

class TFConverter:
    def __init__(self, graph_def, nodes, outfile, dump4tb):
        self.graph_def = graph_def
        self.nodes = nodes
        self.outfile = outfile
        self.dump4tb = dump4tb
        self.layer_number = 0
        self.output_names = []
        self.name_node_dict = {}
        self.edges = {}
        self.conv_activations = {'Relu':0, 'Tanh':1, 'Sigmoid':2, 'None':3, 'LeakyRelu':4}
        self.conv_paddings = {'VALID':0, 'SAME':1}
        self.pool_paddings = {'VALID':0, 'SAME':1}
        self.converted_nodes = set()
        self.conv2d_scope_names = set()
        self.conv2d_scopename_inputname_dict = {}
        self.dense_scope_names = set()
        self.dense_scopename_inputname_dict = {}
        self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4,
                        'MathBinary':5, 'MathUnary':6, 'AvgPool':7, 'MatMul':8}
        self.mathbin2code = {'Sub':0, 'Add':1, 'Mul':2, 'RealDiv':3, 'Minimum':4, 'FloorMod':5}
        self.mathun2code  = {'Abs':0, 'Sin':1, 'Cos':2, 'Tan':3, 'Asin':4,
                'Acos':5, 'Atan':6, 'Sinh':7, 'Cosh':8, 'Tanh':9, 'Asinh':10,
                'Acosh':11, 'Atanh':12, 'Ceil':13, 'Floor':14, 'Round':15,
                'Exp':16}
        self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
        self.name_operand_dict = {}


    def add_operand(self, name, type):
        node = self.name_node_dict[name]
        if name not in self.name_operand_dict:
            dtype = node.attr['dtype'].type
            if dtype == 0:
                dtype = node.attr['T'].type
            dims = [-1,-1,-1,-1]
            if 'shape' in node.attr:
                dims[0] = node.attr['shape'].shape.dim[0].size
                dims[1] = node.attr['shape'].shape.dim[1].size
                dims[2] = node.attr['shape'].shape.dim[2].size
                dims[3] = node.attr['shape'].shape.dim[3].size
            operand = Operand(name, dtype, dims)
            self.name_operand_dict[name] = operand;
        self.name_operand_dict[name].add_iotype(type)
        return self.name_operand_dict[name].index


    def dump_for_tensorboard(self):
        graph = tf.get_default_graph()
        tf.import_graph_def(self.graph_def, name="")
        tf.summary.FileWriter('/tmp/graph', graph)
        print('graph saved, run "tensorboard --logdir=/tmp/graph" to see it')


    def get_conv2d_params(self, conv2d_scope_name):
        knode = self.name_node_dict[conv2d_scope_name + '/kernel']
        bnode = self.name_node_dict[conv2d_scope_name + '/bias']

        if conv2d_scope_name + '/dilation_rate' in self.name_node_dict:
            dnode = self.name_node_dict[conv2d_scope_name + '/dilation_rate']
        else:
            dnode = None

        # the BiasAdd name is possible be changed into the output name,
        # if activation is None, and BiasAdd.next is the last op which is Identity
        if conv2d_scope_name + '/BiasAdd' in self.edges:
            anode = self.edges[conv2d_scope_name + '/BiasAdd'][0]
            if anode.op not in self.conv_activations:
                anode = None
        else:
            anode = None
        return knode, bnode, dnode, anode


    def get_dense_params(self, dense_scope_name):
        knode = self.name_node_dict[dense_scope_name + '/kernel']
        bnode = self.name_node_dict.get(dense_scope_name + '/bias')
        # the BiasAdd name is possible be changed into the output name,
        # if activation is None, and BiasAdd.next is the last op which is Identity
        anode = None
        if bnode:
            if dense_scope_name + '/BiasAdd' in self.edges:
                anode = self.edges[dense_scope_name + '/BiasAdd'][0]
                if anode.op not in self.conv_activations:
                    anode = None
        else:
            anode = None
        return knode, bnode, anode


    def dump_complex_conv2d_to_file(self, node, f):
        assert(node.op == 'Conv2D')
        self.layer_number = self.layer_number + 1
        self.converted_nodes.add(node.name)

        scope_name = TFConverter.get_scope_name(node.name)
        #knode for kernel, bnode for bias, dnode for dilation, anode for activation
        knode, bnode, dnode, anode = self.get_conv2d_params(scope_name)

        if dnode is not None:
            dilation = struct.unpack('i', dnode.attr['value'].tensor.tensor_content[0:4])[0]
        else:
            dilation = 1

        if anode is not None:
            activation = anode.op
        else:
            activation = 'None'

        padding = node.attr['padding'].s.decode("utf-8")
        # conv2d with dilation > 1 generates tens of nodes, not easy to parse them, so use this tricky method.
        if dilation > 1 and scope_name + '/stack' in self.name_node_dict:
            if self.name_node_dict[scope_name + '/stack'].op == "Const":
                padding = 'SAME'
        padding = self.conv_paddings[padding]

        ktensor = knode.attr['value'].tensor
        filter_height = ktensor.tensor_shape.dim[0].size
        filter_width = ktensor.tensor_shape.dim[1].size
        in_channels = ktensor.tensor_shape.dim[2].size
        out_channels = ktensor.tensor_shape.dim[3].size
        kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
        kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels)
        kernel = np.transpose(kernel, [3, 0, 1, 2])

        has_bias = 1
        np.array([self.op2code[node.op], dilation, padding, self.conv_activations[activation], in_channels, out_channels, filter_height, has_bias], dtype=np.uint32).tofile(f)
        kernel.tofile(f)

        btensor = bnode.attr['value'].tensor
        if btensor.tensor_shape.dim[0].size == 1:
            bias = struct.pack("f", btensor.float_val[0])
        else:
            bias = btensor.tensor_content
        f.write(bias)

        input_name = self.conv2d_scopename_inputname_dict[scope_name]
        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)

        if anode is not None:
            output_operand_index = self.add_operand(anode.name, Operand.IOTYPE_OUTPUT)
        else:
            output_operand_index = self.add_operand(self.edges[bnode.name][0].name, Operand.IOTYPE_OUTPUT)
        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)

    def dump_dense_to_file(self, node, f):
        assert(node.op == 'MatMul')
        self.layer_number = self.layer_number + 1
        self.converted_nodes.add(node.name)

        scope_name = TFConverter.get_scope_name(node.name)
        #knode for kernel, bnode for bias, anode for activation
        knode, bnode, anode = self.get_dense_params(scope_name.split('/')[0])

        if bnode is not None:
            has_bias = 1
            btensor = bnode.attr['value'].tensor
            if btensor.tensor_shape.dim[0].size == 1:
                bias = struct.pack("f", btensor.float_val[0])
            else:
                bias = btensor.tensor_content
        else:
            has_bias = 0

        if anode is not None:
            activation = anode.op
        else:
            activation = 'None'

        ktensor = knode.attr['value'].tensor
        in_channels = ktensor.tensor_shape.dim[0].size
        out_channels = ktensor.tensor_shape.dim[1].size
        if in_channels * out_channels == 1:
            kernel = np.float32(ktensor.float_val[0])
        else:
            kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
        kernel = kernel.reshape(in_channels, out_channels)
        kernel = np.transpose(kernel, [1, 0])

        np.array([self.op2code[node.op], self.conv_activations[activation], in_channels, out_channels, has_bias], dtype=np.uint32).tofile(f)
        kernel.tofile(f)
        if has_bias:
            f.write(bias)

        input_name = self.dense_scopename_inputname_dict[scope_name.split('/')[0]]
        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)

        if anode is not None:
            output_operand_index = self.add_operand(anode.name, Operand.IOTYPE_OUTPUT)
        else:
            if bnode is not None:
                output_operand_index = self.add_operand(self.edges[bnode.name][0].name, Operand.IOTYPE_OUTPUT)
            else:
                output_operand_index = self.add_operand(self.edges[scope_name+'/concat_1'][0].name, Operand.IOTYPE_OUTPUT)
        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)


    def dump_simple_conv2d_to_file(self, node, f):
        assert(node.op == 'Conv2D')
        self.layer_number = self.layer_number + 1
        self.converted_nodes.add(node.name)

        node0 = self.name_node_dict[node.input[0]]
        node1 = self.name_node_dict[node.input[1]]
        if node0.op == 'Const':
            knode = node0
            input_name = node.input[1]
        else:
            knode = node1
            input_name = node.input[0]

        ktensor = knode.attr['value'].tensor
        filter_height = ktensor.tensor_shape.dim[0].size
        filter_width = ktensor.tensor_shape.dim[1].size
        in_channels = ktensor.tensor_shape.dim[2].size
        out_channels = ktensor.tensor_shape.dim[3].size
        if filter_height * filter_width * in_channels * out_channels == 1:
            kernel = np.float32(ktensor.float_val[0])
        else:
            kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
        kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels)
        kernel = np.transpose(kernel, [3, 0, 1, 2])

        has_bias = 0
        dilation = 1
        padding = node.attr['padding'].s.decode("utf-8")
        np.array([self.op2code[node.op], dilation, self.conv_paddings[padding], self.conv_activations['None'],
                  in_channels, out_channels, filter_height, has_bias], dtype=np.uint32).tofile(f)
        kernel.tofile(f)

        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)


    def dump_depth2space_to_file(self, node, f):
        assert(node.op == 'DepthToSpace')
        self.layer_number = self.layer_number + 1
        block_size = node.attr['block_size'].i
        np.array([self.op2code[node.op], block_size], dtype=np.uint32).tofile(f)
        self.converted_nodes.add(node.name)
        input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)


    def dump_mirrorpad_to_file(self, node, f):
        assert(node.op == 'MirrorPad')
        self.layer_number = self.layer_number + 1
        mode = node.attr['mode'].s
        mode = self.mirrorpad_mode[mode.decode("utf-8")]
        np.array([self.op2code[node.op], mode], dtype=np.uint32).tofile(f)
        pnode = self.name_node_dict[node.input[1]]
        self.converted_nodes.add(pnode.name)
        paddings = pnode.attr['value'].tensor.tensor_content
        f.write(paddings)
        self.converted_nodes.add(node.name)
        input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)


    def dump_maximum_to_file(self, node, f):
        assert(node.op == 'Maximum')
        self.layer_number = self.layer_number + 1
        ynode = self.name_node_dict[node.input[1]]
        y = ynode.attr['value'].tensor.float_val[0]
        np.array([self.op2code[node.op]], dtype=np.uint32).tofile(f)
        np.array([y], dtype=np.float32).tofile(f)
        self.converted_nodes.add(node.name)
        input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)


    def dump_mathbinary_to_file(self, node, f):
        self.layer_number = self.layer_number + 1
        self.converted_nodes.add(node.name)
        i0_node = self.name_node_dict[node.input[0]]
        i1_node = self.name_node_dict[node.input[1]]
        np.array([self.op2code['MathBinary'], self.mathbin2code[node.op]], dtype=np.uint32).tofile(f)
        if i0_node.op == 'Const':
            scalar = i0_node.attr['value'].tensor.float_val[0]
            np.array([1], dtype=np.uint32).tofile(f)            # broadcast: 1
            np.array([scalar], dtype=np.float32).tofile(f)
            np.array([0], dtype=np.uint32).tofile(f)            # broadcast: 0
            input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
            np.array([input_operand_index], dtype=np.uint32).tofile(f)
        elif i1_node.op == 'Const':
            scalar = i1_node.attr['value'].tensor.float_val[0]
            np.array([0], dtype=np.uint32).tofile(f)
            input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
            np.array([input_operand_index], dtype=np.uint32).tofile(f)
            np.array([1], dtype=np.uint32).tofile(f)
            np.array([scalar], dtype=np.float32).tofile(f)
        else:
            np.array([0], dtype=np.uint32).tofile(f)
            input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
            np.array([input_operand_index], dtype=np.uint32).tofile(f)
            np.array([0], dtype=np.uint32).tofile(f)
            input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
            np.array([input_operand_index], dtype=np.uint32).tofile(f)
        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
        np.array([output_operand_index], dtype=np.uint32).tofile(f)


    def dump_mathunary_to_file(self, node, f):
        self.layer_number = self.layer_number + 1
        self.converted_nodes.add(node.name)
        i0_node = self.name_node_dict[node.input[0]]
        np.array([self.op2code['MathUnary'], self.mathun2code[node.op]], dtype=np.uint32).tofile(f)
        input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
        np.array([input_operand_index], dtype=np.uint32).tofile(f)
        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
        np.array([output_operand_index],dtype=np.uint32).tofile(f)


    def dump_avg_pool_to_file(self, node, f):
        assert(node.op == 'AvgPool')
        self.layer_number = self.layer_number + 1
        self.converted_nodes.add(node.name)
        node0 = self.name_node_dict[node.input[0]]
        strides = node.attr['strides']

        # Tensorflow do not support pooling strides in batch dimension and
        # current native NN do not support pooling strides in channel dimension, added assert() here.
        assert(strides.list.i[1]==strides.list.i[2])
        assert(strides.list.i[0]==1)
        assert(strides.list.i[3]==1)
        strides = strides.list.i[1]
        filter_node = node.attr['ksize']
        input_name = node.input[0]

        # Tensorflow do not support pooling ksize in batch dimension and channel dimension.
        assert(filter_node.list.i[0]==1)
        assert(filter_node.list.i[3]==1)
        filter_height = filter_node.list.i[1]
        filter_width = filter_node.list.i[2]

        padding = node.attr['padding'].s.decode("utf-8")
        np.array([self.op2code[node.op], strides, self.pool_paddings[padding], filter_height],
                 dtype=np.uint32).tofile(f)

        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
        np.array([input_operand_index, output_operand_index],dtype=np.uint32).tofile(f)


    def dump_layers_to_file(self, f):
        for node in self.nodes:
            if node.name in self.converted_nodes:
                continue

            # conv2d with dilation generates very complex nodes, so handle it in special
            if self.in_conv2d_scope(node.name):
                if node.op == 'Conv2D':
                    self.dump_complex_conv2d_to_file(node, f)
                continue
            if self.in_dense_scope(node.name):
                if node.op == 'MatMul':
                    self.dump_dense_to_file(node, f)
                continue


            if node.op == 'Conv2D':
                self.dump_simple_conv2d_to_file(node, f)
                continue
            if node.name in self.output_names:
                input_name = self.id_different_scope_dict[node.name]
                if TFConverter.get_scope_name(input_name)!=TFConverter.get_scope_name(node.name):
                    continue
            if node.op == 'AvgPool':
                self.dump_avg_pool_to_file(node, f)
            elif node.op == 'DepthToSpace':
                self.dump_depth2space_to_file(node, f)
            elif node.op == 'MirrorPad':
                self.dump_mirrorpad_to_file(node, f)
            elif node.op == 'Maximum':
                self.dump_maximum_to_file(node, f)
            elif node.op in self.mathbin2code:
                self.dump_mathbinary_to_file(node, f)
            elif node.op in self.mathun2code:
                self.dump_mathunary_to_file(node, f)


    def dump_operands_to_file(self, f):
            operands = sorted(self.name_operand_dict.values())
            for operand in operands:
                #print('{}'.format(operand))
                np.array([operand.index, len(operand.name)], dtype=np.uint32).tofile(f)
                f.write(operand.name.encode('utf-8'))
                np.array([operand.iotype, operand.dtype], dtype=np.uint32).tofile(f)
                np.array(operand.dims, dtype=np.uint32).tofile(f)


    def dump_to_file(self):
        with open(self.outfile, 'wb') as f:
            f.write(header.str.encode('utf-8'))
            np.array([header.major, header.minor], dtype=np.uint32).tofile(f)
            self.dump_layers_to_file(f)
            self.dump_operands_to_file(f)
            np.array([self.layer_number, len(self.name_operand_dict)], dtype=np.uint32).tofile(f)


    def generate_name_node_dict(self):
        for node in self.nodes:
            self.name_node_dict[node.name] = node


    def generate_output_names(self):
        used_names = []
        for node in self.nodes:
            for input in node.input:
                used_names.append(input)

        for node in self.nodes:
            if node.name not in used_names:
                self.output_names.append(node.name)


    def remove_identity(self):
        self.id_different_scope_dict = {}
        id_nodes = []
        id_dict = {}
        for node in self.nodes:
            if node.op == 'Identity':
                name = node.name
                input = node.input[0]
                id_nodes.append(node)
                # do not change the output name
                if name in self.output_names:
                    self.name_node_dict[input].name = name
                    self.name_node_dict[name] = self.name_node_dict[input]
                    del self.name_node_dict[input]
                    self.id_different_scope_dict[name] = input
                else:
                    id_dict[name] = input

        for idnode in id_nodes:
            self.nodes.remove(idnode)

        for node in self.nodes:
            for i in range(len(node.input)):
                input = node.input[i]
                if input in id_dict:
                    node.input[i] = id_dict[input]


    def generate_edges(self):
        for node in self.nodes:
            for input in node.input:
                if input in self.edges:
                    self.edges[input].append(node)
                else:
                    self.edges[input] = [node]


    @staticmethod
    def get_scope_name(name):
        index = name.rfind('/')
        if index == -1:
            return ""
        return name[0:index]


    def in_conv2d_scope(self, name):
        inner_scope = TFConverter.get_scope_name(name)
        if inner_scope == "":
            return False;
        for scope in self.conv2d_scope_names:
            index = inner_scope.find(scope)
            if index == 0:
                return True
        return False


    def in_dense_scope(self, name):
        inner_scope = TFConverter.get_scope_name(name)
        if inner_scope == "":
            return False;
        for scope in self.dense_scope_names:
            index = inner_scope.find(scope)
            if index == 0:
                return True
        return False

    def generate_sub_block_op_scope_info(self):
        # mostly, conv2d/dense is a sub block in graph, get the scope name
        for node in self.nodes:
            if node.op == 'Conv2D':
                scope = TFConverter.get_scope_name(node.name)
                # for the case tf.nn.conv2d is called directly
                if scope == '':
                    continue
                # for the case tf.nn.conv2d is called within a scope
                if scope + '/kernel' not in self.name_node_dict:
                    continue
                self.conv2d_scope_names.add(scope)
            elif node.op == 'MatMul':
                scope = TFConverter.get_scope_name(node.name)
                # for the case tf.nn.dense is called directly
                if scope == '':
                    continue
                # for the case tf.nn.dense is called within a scope
                if scope + '/kernel' not in self.name_node_dict and scope.split('/Tensordot')[0] + '/kernel' not in self.name_node_dict:
                    continue
                self.dense_scope_names.add(scope.split('/Tensordot')[0])

        # get the input name to the conv2d/dense sub block
        for node in self.nodes:
            scope = TFConverter.get_scope_name(node.name)
            if scope in self.conv2d_scope_names:
                if node.op == 'Conv2D' or node.op == 'Shape':
                    for inp in node.input:
                        if TFConverter.get_scope_name(inp) != scope:
                            self.conv2d_scopename_inputname_dict[scope] = inp
            elif scope in self.dense_scope_names:
                if node.op == 'MatMul' or node.op == 'Shape':
                    for inp in node.input:
                        if TFConverter.get_scope_name(inp) != scope:
                            self.dense_scopename_inputname_dict[scope] = inp
            elif scope.split('/Tensordot')[0] in self.dense_scope_names:
                if node.op == 'Transpose':
                    for inp in node.input:
                        if TFConverter.get_scope_name(inp).find(scope)<0 and TFConverter.get_scope_name(inp).find(scope.split('/')[0])<0:
                            self.dense_scopename_inputname_dict[scope.split('/Tensordot')[0]] = inp


    def run(self):
        self.generate_name_node_dict()
        self.generate_output_names()
        self.remove_identity()
        self.generate_edges()
        self.generate_sub_block_op_scope_info()

        if self.dump4tb:
            self.dump_for_tensorboard()

        self.dump_to_file()


def convert_from_tensorflow(infile, outfile, dump4tb):
    with open(infile, 'rb') as f:
        # read the file in .proto format
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
        nodes = graph_def.node

    converter = TFConverter(graph_def, nodes, outfile, dump4tb)
    converter.run()