• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright (c) 2019 Guo Yejun
2#
3# This file is part of FFmpeg.
4#
5# FFmpeg is free software; you can redistribute it and/or
6# modify it under the terms of the GNU Lesser General Public
7# License as published by the Free Software Foundation; either
8# version 2.1 of the License, or (at your option) any later version.
9#
10# FFmpeg is distributed in the hope that it will be useful,
11# but WITHOUT ANY WARRANTY; without even the implied warranty of
12# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13# Lesser General Public License for more details.
14#
15# You should have received a copy of the GNU Lesser General Public
16# License along with FFmpeg; if not, write to the Free Software
17# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18# ==============================================================================
19
20import tensorflow as tf
21import numpy as np
22import sys, struct
23import convert_header as header
24
25__all__ = ['convert_from_tensorflow']
26
27class Operand(object):
28    IOTYPE_INPUT = 1
29    IOTYPE_OUTPUT = 2
30    IOTYPE_INTERMEDIATE = IOTYPE_INPUT | IOTYPE_OUTPUT
31    DTYPE_FLOAT = 1
32    DTYPE_UINT8 = 4
33    index = 0
34    def __init__(self, name, dtype, dims):
35        self.name = name
36        self.dtype = dtype
37        self.dims = dims
38        self.iotype = 0
39        self.used_count = 0
40        self.index = Operand.index
41        Operand.index = Operand.index + 1
42        self.iotype2str = {Operand.IOTYPE_INPUT: 'in', Operand.IOTYPE_OUTPUT: 'out', Operand.IOTYPE_INTERMEDIATE: 'inout'}
43        self.dtype2str = {Operand.DTYPE_FLOAT: 'DT_FLOAT', Operand.DTYPE_UINT8: 'DT_UINT8'}
44
45    def add_iotype(self, iotype):
46        self.iotype = self.iotype | iotype
47        if iotype == Operand.IOTYPE_INPUT:
48            self.used_count = self.used_count + 1
49
50    def __str__(self):
51        return "{}: (name: {}, iotype: {}, dtype: {}, dims: {}, used_count: {})".format(self.index,
52                            self.name, self.iotype2str[self.iotype], self.dtype2str[self.dtype],
53                            self.dims, self.used_count)
54
55    def __lt__(self, other):
56        return self.index < other.index
57
58class TFConverter:
59    def __init__(self, graph_def, nodes, outfile, dump4tb):
60        self.graph_def = graph_def
61        self.nodes = nodes
62        self.outfile = outfile
63        self.dump4tb = dump4tb
64        self.layer_number = 0
65        self.output_names = []
66        self.name_node_dict = {}
67        self.edges = {}
68        self.conv_activations = {'Relu':0, 'Tanh':1, 'Sigmoid':2, 'None':3, 'LeakyRelu':4}
69        self.conv_paddings = {'VALID':0, 'SAME':1}
70        self.pool_paddings = {'VALID':0, 'SAME':1}
71        self.converted_nodes = set()
72        self.conv2d_scope_names = set()
73        self.conv2d_scopename_inputname_dict = {}
74        self.dense_scope_names = set()
75        self.dense_scopename_inputname_dict = {}
76        self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4,
77                        'MathBinary':5, 'MathUnary':6, 'AvgPool':7, 'MatMul':8}
78        self.mathbin2code = {'Sub':0, 'Add':1, 'Mul':2, 'RealDiv':3, 'Minimum':4, 'FloorMod':5}
79        self.mathun2code  = {'Abs':0, 'Sin':1, 'Cos':2, 'Tan':3, 'Asin':4,
80                'Acos':5, 'Atan':6, 'Sinh':7, 'Cosh':8, 'Tanh':9, 'Asinh':10,
81                'Acosh':11, 'Atanh':12, 'Ceil':13, 'Floor':14, 'Round':15}
82        self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
83        self.name_operand_dict = {}
84
85
86    def add_operand(self, name, type):
87        node = self.name_node_dict[name]
88        if name not in self.name_operand_dict:
89            dtype = node.attr['dtype'].type
90            if dtype == 0:
91                dtype = node.attr['T'].type
92            dims = [-1,-1,-1,-1]
93            if 'shape' in node.attr:
94                dims[0] = node.attr['shape'].shape.dim[0].size
95                dims[1] = node.attr['shape'].shape.dim[1].size
96                dims[2] = node.attr['shape'].shape.dim[2].size
97                dims[3] = node.attr['shape'].shape.dim[3].size
98            operand = Operand(name, dtype, dims)
99            self.name_operand_dict[name] = operand;
100        self.name_operand_dict[name].add_iotype(type)
101        return self.name_operand_dict[name].index
102
103
104    def dump_for_tensorboard(self):
105        graph = tf.get_default_graph()
106        tf.import_graph_def(self.graph_def, name="")
107        tf.summary.FileWriter('/tmp/graph', graph)
108        print('graph saved, run "tensorboard --logdir=/tmp/graph" to see it')
109
110
111    def get_conv2d_params(self, conv2d_scope_name):
112        knode = self.name_node_dict[conv2d_scope_name + '/kernel']
113        bnode = self.name_node_dict[conv2d_scope_name + '/bias']
114
115        if conv2d_scope_name + '/dilation_rate' in self.name_node_dict:
116            dnode = self.name_node_dict[conv2d_scope_name + '/dilation_rate']
117        else:
118            dnode = None
119
120        # the BiasAdd name is possible be changed into the output name,
121        # if activation is None, and BiasAdd.next is the last op which is Identity
122        if conv2d_scope_name + '/BiasAdd' in self.edges:
123            anode = self.edges[conv2d_scope_name + '/BiasAdd'][0]
124            if anode.op not in self.conv_activations:
125                anode = None
126        else:
127            anode = None
128        return knode, bnode, dnode, anode
129
130
131    def get_dense_params(self, dense_scope_name):
132        knode = self.name_node_dict[dense_scope_name + '/kernel']
133        bnode = self.name_node_dict.get(dense_scope_name + '/bias')
134        # the BiasAdd name is possible be changed into the output name,
135        # if activation is None, and BiasAdd.next is the last op which is Identity
136        anode = None
137        if bnode:
138            if dense_scope_name + '/BiasAdd' in self.edges:
139                anode = self.edges[dense_scope_name + '/BiasAdd'][0]
140                if anode.op not in self.conv_activations:
141                    anode = None
142        else:
143            anode = None
144        return knode, bnode, anode
145
146
147    def dump_complex_conv2d_to_file(self, node, f):
148        assert(node.op == 'Conv2D')
149        self.layer_number = self.layer_number + 1
150        self.converted_nodes.add(node.name)
151
152        scope_name = TFConverter.get_scope_name(node.name)
153        #knode for kernel, bnode for bias, dnode for dilation, anode for activation
154        knode, bnode, dnode, anode = self.get_conv2d_params(scope_name)
155
156        if dnode is not None:
157            dilation = struct.unpack('i', dnode.attr['value'].tensor.tensor_content[0:4])[0]
158        else:
159            dilation = 1
160
161        if anode is not None:
162            activation = anode.op
163        else:
164            activation = 'None'
165
166        padding = node.attr['padding'].s.decode("utf-8")
167        # conv2d with dilation > 1 generates tens of nodes, not easy to parse them, so use this tricky method.
168        if dilation > 1 and scope_name + '/stack' in self.name_node_dict:
169            if self.name_node_dict[scope_name + '/stack'].op == "Const":
170                padding = 'SAME'
171        padding = self.conv_paddings[padding]
172
173        ktensor = knode.attr['value'].tensor
174        filter_height = ktensor.tensor_shape.dim[0].size
175        filter_width = ktensor.tensor_shape.dim[1].size
176        in_channels = ktensor.tensor_shape.dim[2].size
177        out_channels = ktensor.tensor_shape.dim[3].size
178        kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
179        kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels)
180        kernel = np.transpose(kernel, [3, 0, 1, 2])
181
182        has_bias = 1
183        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)
184        kernel.tofile(f)
185
186        btensor = bnode.attr['value'].tensor
187        if btensor.tensor_shape.dim[0].size == 1:
188            bias = struct.pack("f", btensor.float_val[0])
189        else:
190            bias = btensor.tensor_content
191        f.write(bias)
192
193        input_name = self.conv2d_scopename_inputname_dict[scope_name]
194        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
195
196        if anode is not None:
197            output_operand_index = self.add_operand(anode.name, Operand.IOTYPE_OUTPUT)
198        else:
199            output_operand_index = self.add_operand(self.edges[bnode.name][0].name, Operand.IOTYPE_OUTPUT)
200        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
201
202    def dump_dense_to_file(self, node, f):
203        assert(node.op == 'MatMul')
204        self.layer_number = self.layer_number + 1
205        self.converted_nodes.add(node.name)
206
207        scope_name = TFConverter.get_scope_name(node.name)
208        #knode for kernel, bnode for bias, anode for activation
209        knode, bnode, anode = self.get_dense_params(scope_name.split('/')[0])
210
211        if bnode is not None:
212            has_bias = 1
213            btensor = bnode.attr['value'].tensor
214            if btensor.tensor_shape.dim[0].size == 1:
215                bias = struct.pack("f", btensor.float_val[0])
216            else:
217                bias = btensor.tensor_content
218        else:
219            has_bias = 0
220
221        if anode is not None:
222            activation = anode.op
223        else:
224            activation = 'None'
225
226        ktensor = knode.attr['value'].tensor
227        in_channels = ktensor.tensor_shape.dim[0].size
228        out_channels = ktensor.tensor_shape.dim[1].size
229        if in_channels * out_channels == 1:
230            kernel = np.float32(ktensor.float_val[0])
231        else:
232            kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
233        kernel = kernel.reshape(in_channels, out_channels)
234        kernel = np.transpose(kernel, [1, 0])
235
236        np.array([self.op2code[node.op], self.conv_activations[activation], in_channels, out_channels, has_bias], dtype=np.uint32).tofile(f)
237        kernel.tofile(f)
238        if has_bias:
239            f.write(bias)
240
241        input_name = self.dense_scopename_inputname_dict[scope_name.split('/')[0]]
242        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
243
244        if anode is not None:
245            output_operand_index = self.add_operand(anode.name, Operand.IOTYPE_OUTPUT)
246        else:
247            if bnode is not None:
248                output_operand_index = self.add_operand(self.edges[bnode.name][0].name, Operand.IOTYPE_OUTPUT)
249            else:
250                output_operand_index = self.add_operand(self.edges[scope_name+'/concat_1'][0].name, Operand.IOTYPE_OUTPUT)
251        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
252
253
254    def dump_simple_conv2d_to_file(self, node, f):
255        assert(node.op == 'Conv2D')
256        self.layer_number = self.layer_number + 1
257        self.converted_nodes.add(node.name)
258
259        node0 = self.name_node_dict[node.input[0]]
260        node1 = self.name_node_dict[node.input[1]]
261        if node0.op == 'Const':
262            knode = node0
263            input_name = node.input[1]
264        else:
265            knode = node1
266            input_name = node.input[0]
267
268        ktensor = knode.attr['value'].tensor
269        filter_height = ktensor.tensor_shape.dim[0].size
270        filter_width = ktensor.tensor_shape.dim[1].size
271        in_channels = ktensor.tensor_shape.dim[2].size
272        out_channels = ktensor.tensor_shape.dim[3].size
273        if filter_height * filter_width * in_channels * out_channels == 1:
274            kernel = np.float32(ktensor.float_val[0])
275        else:
276            kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
277        kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels)
278        kernel = np.transpose(kernel, [3, 0, 1, 2])
279
280        has_bias = 0
281        dilation = 1
282        padding = node.attr['padding'].s.decode("utf-8")
283        np.array([self.op2code[node.op], dilation, self.conv_paddings[padding], self.conv_activations['None'],
284                  in_channels, out_channels, filter_height, has_bias], dtype=np.uint32).tofile(f)
285        kernel.tofile(f)
286
287        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
288        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
289        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
290
291
292    def dump_depth2space_to_file(self, node, f):
293        assert(node.op == 'DepthToSpace')
294        self.layer_number = self.layer_number + 1
295        block_size = node.attr['block_size'].i
296        np.array([self.op2code[node.op], block_size], dtype=np.uint32).tofile(f)
297        self.converted_nodes.add(node.name)
298        input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
299        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
300        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
301
302
303    def dump_mirrorpad_to_file(self, node, f):
304        assert(node.op == 'MirrorPad')
305        self.layer_number = self.layer_number + 1
306        mode = node.attr['mode'].s
307        mode = self.mirrorpad_mode[mode.decode("utf-8")]
308        np.array([self.op2code[node.op], mode], dtype=np.uint32).tofile(f)
309        pnode = self.name_node_dict[node.input[1]]
310        self.converted_nodes.add(pnode.name)
311        paddings = pnode.attr['value'].tensor.tensor_content
312        f.write(paddings)
313        self.converted_nodes.add(node.name)
314        input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
315        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
316        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
317
318
319    def dump_maximum_to_file(self, node, f):
320        assert(node.op == 'Maximum')
321        self.layer_number = self.layer_number + 1
322        ynode = self.name_node_dict[node.input[1]]
323        y = ynode.attr['value'].tensor.float_val[0]
324        np.array([self.op2code[node.op]], dtype=np.uint32).tofile(f)
325        np.array([y], dtype=np.float32).tofile(f)
326        self.converted_nodes.add(node.name)
327        input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
328        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
329        np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
330
331
332    def dump_mathbinary_to_file(self, node, f):
333        self.layer_number = self.layer_number + 1
334        self.converted_nodes.add(node.name)
335        i0_node = self.name_node_dict[node.input[0]]
336        i1_node = self.name_node_dict[node.input[1]]
337        np.array([self.op2code['MathBinary'], self.mathbin2code[node.op]], dtype=np.uint32).tofile(f)
338        if i0_node.op == 'Const':
339            scalar = i0_node.attr['value'].tensor.float_val[0]
340            np.array([1], dtype=np.uint32).tofile(f)            # broadcast: 1
341            np.array([scalar], dtype=np.float32).tofile(f)
342            np.array([0], dtype=np.uint32).tofile(f)            # broadcast: 0
343            input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
344            np.array([input_operand_index], dtype=np.uint32).tofile(f)
345        elif i1_node.op == 'Const':
346            scalar = i1_node.attr['value'].tensor.float_val[0]
347            np.array([0], dtype=np.uint32).tofile(f)
348            input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
349            np.array([input_operand_index], dtype=np.uint32).tofile(f)
350            np.array([1], dtype=np.uint32).tofile(f)
351            np.array([scalar], dtype=np.float32).tofile(f)
352        else:
353            np.array([0], dtype=np.uint32).tofile(f)
354            input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
355            np.array([input_operand_index], dtype=np.uint32).tofile(f)
356            np.array([0], dtype=np.uint32).tofile(f)
357            input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
358            np.array([input_operand_index], dtype=np.uint32).tofile(f)
359        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
360        np.array([output_operand_index], dtype=np.uint32).tofile(f)
361
362
363    def dump_mathunary_to_file(self, node, f):
364        self.layer_number = self.layer_number + 1
365        self.converted_nodes.add(node.name)
366        i0_node = self.name_node_dict[node.input[0]]
367        np.array([self.op2code['MathUnary'], self.mathun2code[node.op]], dtype=np.uint32).tofile(f)
368        input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
369        np.array([input_operand_index], dtype=np.uint32).tofile(f)
370        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
371        np.array([output_operand_index],dtype=np.uint32).tofile(f)
372
373
374    def dump_avg_pool_to_file(self, node, f):
375        assert(node.op == 'AvgPool')
376        self.layer_number = self.layer_number + 1
377        self.converted_nodes.add(node.name)
378        node0 = self.name_node_dict[node.input[0]]
379        strides = node.attr['strides']
380
381        # Tensorflow do not support pooling strides in batch dimension and
382        # current native NN do not support pooling strides in channel dimension, added assert() here.
383        assert(strides.list.i[1]==strides.list.i[2])
384        assert(strides.list.i[0]==1)
385        assert(strides.list.i[3]==1)
386        strides = strides.list.i[1]
387        filter_node = node.attr['ksize']
388        input_name = node.input[0]
389
390        # Tensorflow do not support pooling ksize in batch dimension and channel dimension.
391        assert(filter_node.list.i[0]==1)
392        assert(filter_node.list.i[3]==1)
393        filter_height = filter_node.list.i[1]
394        filter_width = filter_node.list.i[2]
395
396        padding = node.attr['padding'].s.decode("utf-8")
397        np.array([self.op2code[node.op], strides, self.pool_paddings[padding], filter_height],
398                 dtype=np.uint32).tofile(f)
399
400        input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
401        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
402        np.array([input_operand_index, output_operand_index],dtype=np.uint32).tofile(f)
403
404
405    def dump_layers_to_file(self, f):
406        for node in self.nodes:
407            if node.name in self.converted_nodes:
408                continue
409
410            # conv2d with dilation generates very complex nodes, so handle it in special
411            if self.in_conv2d_scope(node.name):
412                if node.op == 'Conv2D':
413                    self.dump_complex_conv2d_to_file(node, f)
414                continue
415            if self.in_dense_scope(node.name):
416                if node.op == 'MatMul':
417                    self.dump_dense_to_file(node, f)
418                continue
419
420
421            if node.op == 'Conv2D':
422                self.dump_simple_conv2d_to_file(node, f)
423                continue
424            if node.name in self.output_names:
425                input_name = self.id_different_scope_dict[node.name]
426                if TFConverter.get_scope_name(input_name)!=TFConverter.get_scope_name(node.name):
427                    continue
428            if node.op == 'AvgPool':
429                self.dump_avg_pool_to_file(node, f)
430            elif node.op == 'DepthToSpace':
431                self.dump_depth2space_to_file(node, f)
432            elif node.op == 'MirrorPad':
433                self.dump_mirrorpad_to_file(node, f)
434            elif node.op == 'Maximum':
435                self.dump_maximum_to_file(node, f)
436            elif node.op in self.mathbin2code:
437                self.dump_mathbinary_to_file(node, f)
438            elif node.op in self.mathun2code:
439                self.dump_mathunary_to_file(node, f)
440
441
442    def dump_operands_to_file(self, f):
443            operands = sorted(self.name_operand_dict.values())
444            for operand in operands:
445                #print('{}'.format(operand))
446                np.array([operand.index, len(operand.name)], dtype=np.uint32).tofile(f)
447                f.write(operand.name.encode('utf-8'))
448                np.array([operand.iotype, operand.dtype], dtype=np.uint32).tofile(f)
449                np.array(operand.dims, dtype=np.uint32).tofile(f)
450
451
452    def dump_to_file(self):
453        with open(self.outfile, 'wb') as f:
454            f.write(header.str.encode('utf-8'))
455            np.array([header.major, header.minor], dtype=np.uint32).tofile(f)
456            self.dump_layers_to_file(f)
457            self.dump_operands_to_file(f)
458            np.array([self.layer_number, len(self.name_operand_dict)], dtype=np.uint32).tofile(f)
459
460
461    def generate_name_node_dict(self):
462        for node in self.nodes:
463            self.name_node_dict[node.name] = node
464
465
466    def generate_output_names(self):
467        used_names = []
468        for node in self.nodes:
469            for input in node.input:
470                used_names.append(input)
471
472        for node in self.nodes:
473            if node.name not in used_names:
474                self.output_names.append(node.name)
475
476
477    def remove_identity(self):
478        self.id_different_scope_dict = {}
479        id_nodes = []
480        id_dict = {}
481        for node in self.nodes:
482            if node.op == 'Identity':
483                name = node.name
484                input = node.input[0]
485                id_nodes.append(node)
486                # do not change the output name
487                if name in self.output_names:
488                    self.name_node_dict[input].name = name
489                    self.name_node_dict[name] = self.name_node_dict[input]
490                    del self.name_node_dict[input]
491                    self.id_different_scope_dict[name] = input
492                else:
493                    id_dict[name] = input
494
495        for idnode in id_nodes:
496            self.nodes.remove(idnode)
497
498        for node in self.nodes:
499            for i in range(len(node.input)):
500                input = node.input[i]
501                if input in id_dict:
502                    node.input[i] = id_dict[input]
503
504
505    def generate_edges(self):
506        for node in self.nodes:
507            for input in node.input:
508                if input in self.edges:
509                    self.edges[input].append(node)
510                else:
511                    self.edges[input] = [node]
512
513
514    @staticmethod
515    def get_scope_name(name):
516        index = name.rfind('/')
517        if index == -1:
518            return ""
519        return name[0:index]
520
521
522    def in_conv2d_scope(self, name):
523        inner_scope = TFConverter.get_scope_name(name)
524        if inner_scope == "":
525            return False;
526        for scope in self.conv2d_scope_names:
527            index = inner_scope.find(scope)
528            if index == 0:
529                return True
530        return False
531
532
533    def in_dense_scope(self, name):
534        inner_scope = TFConverter.get_scope_name(name)
535        if inner_scope == "":
536            return False;
537        for scope in self.dense_scope_names:
538            index = inner_scope.find(scope)
539            if index == 0:
540                return True
541        return False
542
543    def generate_sub_block_op_scope_info(self):
544        # mostly, conv2d/dense is a sub block in graph, get the scope name
545        for node in self.nodes:
546            if node.op == 'Conv2D':
547                scope = TFConverter.get_scope_name(node.name)
548                # for the case tf.nn.conv2d is called directly
549                if scope == '':
550                    continue
551                # for the case tf.nn.conv2d is called within a scope
552                if scope + '/kernel' not in self.name_node_dict:
553                    continue
554                self.conv2d_scope_names.add(scope)
555            elif node.op == 'MatMul':
556                scope = TFConverter.get_scope_name(node.name)
557                # for the case tf.nn.dense is called directly
558                if scope == '':
559                    continue
560                # for the case tf.nn.dense is called within a scope
561                if scope + '/kernel' not in self.name_node_dict and scope.split('/Tensordot')[0] + '/kernel' not in self.name_node_dict:
562                    continue
563                self.dense_scope_names.add(scope.split('/Tensordot')[0])
564
565        # get the input name to the conv2d/dense sub block
566        for node in self.nodes:
567            scope = TFConverter.get_scope_name(node.name)
568            if scope in self.conv2d_scope_names:
569                if node.op == 'Conv2D' or node.op == 'Shape':
570                    for inp in node.input:
571                        if TFConverter.get_scope_name(inp) != scope:
572                            self.conv2d_scopename_inputname_dict[scope] = inp
573            elif scope in self.dense_scope_names:
574                if node.op == 'MatMul' or node.op == 'Shape':
575                    for inp in node.input:
576                        if TFConverter.get_scope_name(inp) != scope:
577                            self.dense_scopename_inputname_dict[scope] = inp
578            elif scope.split('/Tensordot')[0] in self.dense_scope_names:
579                if node.op == 'Transpose':
580                    for inp in node.input:
581                        if TFConverter.get_scope_name(inp).find(scope)<0 and TFConverter.get_scope_name(inp).find(scope.split('/')[0])<0:
582                            self.dense_scopename_inputname_dict[scope.split('/Tensordot')[0]] = inp
583
584
585    def run(self):
586        self.generate_name_node_dict()
587        self.generate_output_names()
588        self.remove_identity()
589        self.generate_edges()
590        self.generate_sub_block_op_scope_info()
591
592        if self.dump4tb:
593            self.dump_for_tensorboard()
594
595        self.dump_to_file()
596
597
598def convert_from_tensorflow(infile, outfile, dump4tb):
599    with open(infile, 'rb') as f:
600        # read the file in .proto format
601        graph_def = tf.GraphDef()
602        graph_def.ParseFromString(f.read())
603        nodes = graph_def.node
604
605    converter = TFConverter(graph_def, nodes, outfile, dump4tb)
606    converter.run()
607