git.sesse.net Git - ffmpeg/blob - tools/python/convert_from_tensorflow.py

   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
  20 import tensorflow as tf
  21 import numpy as np
  22 import sys, struct
  23 import convert_header as header
  24
  25 __all__ = ['convert_from_tensorflow']
  26
  27 class 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
  58 class 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
 598 def 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()