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[0], self.dims[1], self.dims[2], self.dims[3], 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.converted_nodes = set()
  71         self.conv2d_scope_names = set()
  72         self.conv2d_scopename_inputname_dict = {}
  73         self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4, 'MathBinary':5}
  74         self.mathbin2code = {'Sub':0, 'Add':1, 'Mul':2, 'RealDiv':3}
  75         self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
  76         self.name_operand_dict = {}
  77
  78
  79     def add_operand(self, name, type):
  80         node = self.name_node_dict[name]
  81         if name not in self.name_operand_dict:
  82             dtype = node.attr['dtype'].type
  83             if dtype == 0:
  84                 dtype = node.attr['T'].type
  85             dims = [-1,-1,-1,-1]
  86             if 'shape' in node.attr:
  87                 dims[0] = node.attr['shape'].shape.dim[0].size
  88                 dims[1] = node.attr['shape'].shape.dim[1].size
  89                 dims[2] = node.attr['shape'].shape.dim[2].size
  90                 dims[3] = node.attr['shape'].shape.dim[3].size
  91             operand = Operand(name, dtype, dims)
  92             self.name_operand_dict[name] = operand;
  93         self.name_operand_dict[name].add_iotype(type)
  94         return self.name_operand_dict[name].index
  95
  96
  97     def dump_for_tensorboard(self):
  98         graph = tf.get_default_graph()
  99         tf.import_graph_def(self.graph_def, name="")
 100         tf.summary.FileWriter('/tmp/graph', graph)
 101         print('graph saved, run "tensorboard --logdir=/tmp/graph" to see it')
 102
 103
 104     def get_conv2d_params(self, conv2d_scope_name):
 105         knode = self.name_node_dict[conv2d_scope_name + '/kernel']
 106         bnode = self.name_node_dict[conv2d_scope_name + '/bias']
 107
 108         if conv2d_scope_name + '/dilation_rate' in self.name_node_dict:
 109             dnode = self.name_node_dict[conv2d_scope_name + '/dilation_rate']
 110         else:
 111             dnode = None
 112
 113         # the BiasAdd name is possible be changed into the output name,
 114         # if activation is None, and BiasAdd.next is the last op which is Identity
 115         if conv2d_scope_name + '/BiasAdd' in self.edges:
 116             anode = self.edges[conv2d_scope_name + '/BiasAdd'][0]
 117             if anode.op not in self.conv_activations:
 118                 anode = None
 119         else:
 120             anode = None
 121         return knode, bnode, dnode, anode
 122
 123
 124     def dump_complex_conv2d_to_file(self, node, f):
 125         assert(node.op == 'Conv2D')
 126         self.layer_number = self.layer_number + 1
 127         self.converted_nodes.add(node.name)
 128
 129         scope_name = TFConverter.get_scope_name(node.name)
 130         #knode for kernel, bnode for bias, dnode for dilation, anode for activation
 131         knode, bnode, dnode, anode = self.get_conv2d_params(scope_name)
 132
 133         if dnode is not None:
 134             dilation = struct.unpack('i', dnode.attr['value'].tensor.tensor_content[0:4])[0]
 135         else:
 136             dilation = 1
 137
 138         if anode is not None:
 139             activation = anode.op
 140         else:
 141             activation = 'None'
 142
 143         padding = node.attr['padding'].s.decode("utf-8")
 144         # conv2d with dilation > 1 generates tens of nodes, not easy to parse them, so use this tricky method.
 145         if dilation > 1 and scope_name + '/stack' in self.name_node_dict:
 146             if self.name_node_dict[scope_name + '/stack'].op == "Const":
 147                 padding = 'SAME'
 148         padding = self.conv_paddings[padding]
 149
 150         ktensor = knode.attr['value'].tensor
 151         filter_height = ktensor.tensor_shape.dim[0].size
 152         filter_width = ktensor.tensor_shape.dim[1].size
 153         in_channels = ktensor.tensor_shape.dim[2].size
 154         out_channels = ktensor.tensor_shape.dim[3].size
 155         kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
 156         kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels)
 157         kernel = np.transpose(kernel, [3, 0, 1, 2])
 158
 159         has_bias = 1
 160         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)
 161         kernel.tofile(f)
 162
 163         btensor = bnode.attr['value'].tensor
 164         if btensor.tensor_shape.dim[0].size == 1:
 165             bias = struct.pack("f", btensor.float_val[0])
 166         else:
 167             bias = btensor.tensor_content
 168         f.write(bias)
 169
 170         input_name = self.conv2d_scopename_inputname_dict[scope_name]
 171         input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
 172
 173         if anode is not None:
 174             output_operand_index = self.add_operand(anode.name, Operand.IOTYPE_OUTPUT)
 175         else:
 176             output_operand_index = self.add_operand(self.edges[bnode.name][0].name, Operand.IOTYPE_OUTPUT)
 177         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 178
 179
 180     def dump_simple_conv2d_to_file(self, node, f):
 181         assert(node.op == 'Conv2D')
 182         self.layer_number = self.layer_number + 1
 183         self.converted_nodes.add(node.name)
 184
 185         node0 = self.name_node_dict[node.input[0]]
 186         node1 = self.name_node_dict[node.input[1]]
 187         if node0.op == 'Const':
 188             knode = node0
 189             input_name = node.input[1]
 190         else:
 191             knode = node1
 192             input_name = node.input[0]
 193
 194         ktensor = knode.attr['value'].tensor
 195         filter_height = ktensor.tensor_shape.dim[0].size
 196         filter_width = ktensor.tensor_shape.dim[1].size
 197         in_channels = ktensor.tensor_shape.dim[2].size
 198         out_channels = ktensor.tensor_shape.dim[3].size
 199         if filter_height * filter_width * in_channels * out_channels == 1:
 200             kernel = np.float32(ktensor.float_val[0])
 201         else:
 202             kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
 203         kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels)
 204         kernel = np.transpose(kernel, [3, 0, 1, 2])
 205
 206         has_bias = 0
 207         dilation = 1
 208         padding = node.attr['padding'].s.decode("utf-8")
 209         np.array([self.op2code[node.op], dilation, self.conv_paddings[padding], self.conv_activations['None'],
 210                   in_channels, out_channels, filter_height, has_bias], dtype=np.uint32).tofile(f)
 211         kernel.tofile(f)
 212
 213         input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
 214         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 215         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 216
 217
 218     def dump_depth2space_to_file(self, node, f):
 219         assert(node.op == 'DepthToSpace')
 220         self.layer_number = self.layer_number + 1
 221         block_size = node.attr['block_size'].i
 222         np.array([self.op2code[node.op], block_size], dtype=np.uint32).tofile(f)
 223         self.converted_nodes.add(node.name)
 224         input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
 225         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 226         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 227
 228
 229     def dump_mirrorpad_to_file(self, node, f):
 230         assert(node.op == 'MirrorPad')
 231         self.layer_number = self.layer_number + 1
 232         mode = node.attr['mode'].s
 233         mode = self.mirrorpad_mode[mode.decode("utf-8")]
 234         np.array([self.op2code[node.op], mode], dtype=np.uint32).tofile(f)
 235         pnode = self.name_node_dict[node.input[1]]
 236         self.converted_nodes.add(pnode.name)
 237         paddings = pnode.attr['value'].tensor.tensor_content
 238         f.write(paddings)
 239         self.converted_nodes.add(node.name)
 240         input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
 241         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 242         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 243
 244
 245     def dump_maximum_to_file(self, node, f):
 246         assert(node.op == 'Maximum')
 247         self.layer_number = self.layer_number + 1
 248         ynode = self.name_node_dict[node.input[1]]
 249         y = ynode.attr['value'].tensor.float_val[0]
 250         np.array([self.op2code[node.op]], dtype=np.uint32).tofile(f)
 251         np.array([y], dtype=np.float32).tofile(f)
 252         self.converted_nodes.add(node.name)
 253         input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
 254         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 255         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 256
 257
 258     def dump_mathbinary_to_file(self, node, f):
 259         self.layer_number = self.layer_number + 1
 260         self.converted_nodes.add(node.name)
 261         i0_node = self.name_node_dict[node.input[0]]
 262         i1_node = self.name_node_dict[node.input[1]]
 263         np.array([self.op2code['MathBinary'], self.mathbin2code[node.op]], dtype=np.uint32).tofile(f)
 264         if i0_node.op == 'Const':
 265             scalar = i0_node.attr['value'].tensor.float_val[0]
 266             np.array([1], dtype=np.uint32).tofile(f)            # broadcast: 1
 267             np.array([scalar], dtype=np.float32).tofile(f)
 268             np.array([0], dtype=np.uint32).tofile(f)            # broadcast: 0
 269             input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
 270             np.array([input_operand_index], dtype=np.uint32).tofile(f)
 271         elif i1_node.op == 'Const':
 272             scalar = i1_node.attr['value'].tensor.float_val[0]
 273             np.array([0], dtype=np.uint32).tofile(f)
 274             input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
 275             np.array([input_operand_index], dtype=np.uint32).tofile(f)
 276             np.array([1], dtype=np.uint32).tofile(f)
 277             np.array([scalar], dtype=np.float32).tofile(f)
 278         else:
 279             np.array([0], dtype=np.uint32).tofile(f)
 280             input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
 281             np.array([input_operand_index], dtype=np.uint32).tofile(f)
 282             np.array([0], dtype=np.uint32).tofile(f)
 283             input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
 284             np.array([input_operand_index], dtype=np.uint32).tofile(f)
 285         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 286         np.array([output_operand_index], dtype=np.uint32).tofile(f)
 287
 288
 289     def dump_layers_to_file(self, f):
 290         for node in self.nodes:
 291             if node.name in self.converted_nodes:
 292                 continue
 293
 294             # conv2d with dilation generates very complex nodes, so handle it in special
 295             if self.in_conv2d_scope(node.name):
 296                 if node.op == 'Conv2D':
 297                     self.dump_complex_conv2d_to_file(node, f)
 298                 continue
 299
 300             if node.op == 'Conv2D':
 301                 self.dump_simple_conv2d_to_file(node, f)
 302             elif node.op == 'DepthToSpace':
 303                 self.dump_depth2space_to_file(node, f)
 304             elif node.op == 'MirrorPad':
 305                 self.dump_mirrorpad_to_file(node, f)
 306             elif node.op == 'Maximum':
 307                 self.dump_maximum_to_file(node, f)
 308             elif node.op == 'Sub':
 309                 self.dump_mathbinary_to_file(node, f)
 310             elif node.op == 'Add':
 311                 self.dump_mathbinary_to_file(node, f)
 312             elif node.op == 'Mul':
 313                 self.dump_mathbinary_to_file(node, f)
 314             elif node.op == 'RealDiv':
 315                 self.dump_mathbinary_to_file(node, f)
 316
 317     def dump_operands_to_file(self, f):
 318             operands = sorted(self.name_operand_dict.values())
 319             for operand in operands:
 320                 #print('{}'.format(operand))
 321                 np.array([operand.index, len(operand.name)], dtype=np.uint32).tofile(f)
 322                 f.write(operand.name.encode('utf-8'))
 323                 np.array([operand.iotype, operand.dtype], dtype=np.uint32).tofile(f)
 324                 np.array([operand.dims[0], operand.dims[1], operand.dims[2], operand.dims[3]], dtype=np.uint32).tofile(f)
 325
 326
 327     def dump_to_file(self):
 328         with open(self.outfile, 'wb') as f:
 329             f.write(header.str.encode('utf-8'))
 330             np.array([header.major, header.minor], dtype=np.uint32).tofile(f)
 331             self.dump_layers_to_file(f)
 332             self.dump_operands_to_file(f)
 333             np.array([self.layer_number, len(self.name_operand_dict)], dtype=np.uint32).tofile(f)
 334
 335
 336     def generate_name_node_dict(self):
 337         for node in self.nodes:
 338             self.name_node_dict[node.name] = node
 339
 340
 341     def generate_output_names(self):
 342         used_names = []
 343         for node in self.nodes:
 344             for input in node.input:
 345                 used_names.append(input)
 346
 347         for node in self.nodes:
 348             if node.name not in used_names:
 349                 self.output_names.append(node.name)
 350
 351
 352     def remove_identity(self):
 353         id_nodes = []
 354         id_dict = {}
 355         for node in self.nodes:
 356             if node.op == 'Identity':
 357                 name = node.name
 358                 input = node.input[0]
 359                 id_nodes.append(node)
 360                 # do not change the output name
 361                 if name in self.output_names:
 362                     self.name_node_dict[input].name = name
 363                     self.name_node_dict[name] = self.name_node_dict[input]
 364                     del self.name_node_dict[input]
 365                 else:
 366                     id_dict[name] = input
 367
 368         for idnode in id_nodes:
 369             self.nodes.remove(idnode)
 370
 371         for node in self.nodes:
 372             for i in range(len(node.input)):
 373                 input = node.input[i]
 374                 if input in id_dict:
 375                     node.input[i] = id_dict[input]
 376
 377
 378     def generate_edges(self):
 379         for node in self.nodes:
 380             for input in node.input:
 381                 if input in self.edges:
 382                     self.edges[input].append(node)
 383                 else:
 384                     self.edges[input] = [node]
 385
 386
 387     @staticmethod
 388     def get_scope_name(name):
 389         index = name.rfind('/')
 390         if index == -1:
 391             return ""
 392         return name[0:index]
 393
 394
 395     def in_conv2d_scope(self, name):
 396         inner_scope = TFConverter.get_scope_name(name)
 397         if inner_scope == "":
 398             return False;
 399         for scope in self.conv2d_scope_names:
 400             index = inner_scope.find(scope)
 401             if index == 0:
 402                 return True
 403         return False
 404
 405
 406     def generate_conv2d_scope_info(self):
 407         # mostly, conv2d is a sub block in graph, get the scope name
 408         for node in self.nodes:
 409             if node.op == 'Conv2D':
 410                 scope = TFConverter.get_scope_name(node.name)
 411                 # for the case tf.nn.conv2d is called directly
 412                 if scope == '':
 413                     continue
 414                 # for the case tf.nn.conv2d is called within a scope
 415                 if scope + '/kernel' not in self.name_node_dict:
 416                     continue
 417                 self.conv2d_scope_names.add(scope)
 418
 419         # get the input name to the conv2d sub block
 420         for node in self.nodes:
 421             scope = TFConverter.get_scope_name(node.name)
 422             if scope in self.conv2d_scope_names:
 423                 if node.op == 'Conv2D' or node.op == 'Shape':
 424                     for inp in node.input:
 425                         if TFConverter.get_scope_name(inp) != scope:
 426                             self.conv2d_scopename_inputname_dict[scope] = inp
 427
 428
 429     def run(self):
 430         self.generate_name_node_dict()
 431         self.generate_output_names()
 432         self.remove_identity()
 433         self.generate_edges()
 434         self.generate_conv2d_scope_info()
 435
 436         if self.dump4tb:
 437             self.dump_for_tensorboard()
 438
 439         self.dump_to_file()
 440
 441
 442 def convert_from_tensorflow(infile, outfile, dump4tb):
 443     with open(infile, 'rb') as f:
 444         # read the file in .proto format
 445         graph_def = tf.GraphDef()
 446         graph_def.ParseFromString(f.read())
 447         nodes = graph_def.node
 448
 449     converter = TFConverter(graph_def, nodes, outfile, dump4tb)
 450     converter.run()