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}
  74         self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
  75         self.name_operand_dict = {}
  76
  77
  78     def add_operand(self, name, type):
  79         node = self.name_node_dict[name]
  80         if name not in self.name_operand_dict:
  81             dtype = node.attr['dtype'].type
  82             if dtype == 0:
  83                 dtype = node.attr['T'].type
  84             dims = [-1,-1,-1,-1]
  85             if 'shape' in node.attr:
  86                 dims[0] = node.attr['shape'].shape.dim[0].size
  87                 dims[1] = node.attr['shape'].shape.dim[1].size
  88                 dims[2] = node.attr['shape'].shape.dim[2].size
  89                 dims[3] = node.attr['shape'].shape.dim[3].size
  90             operand = Operand(name, dtype, dims)
  91             self.name_operand_dict[name] = operand;
  92         self.name_operand_dict[name].add_iotype(type)
  93         return self.name_operand_dict[name].index
  94
  95
  96     def dump_for_tensorboard(self):
  97         graph = tf.get_default_graph()
  98         tf.import_graph_def(self.graph_def, name="")
  99         tf.summary.FileWriter('/tmp/graph', graph)
 100         print('graph saved, run "tensorboard --logdir=/tmp/graph" to see it')
 101
 102
 103     def get_conv2d_params(self, conv2d_scope_name):
 104         knode = self.name_node_dict[conv2d_scope_name + '/kernel']
 105         bnode = self.name_node_dict[conv2d_scope_name + '/bias']
 106
 107         if conv2d_scope_name + '/dilation_rate' in self.name_node_dict:
 108             dnode = self.name_node_dict[conv2d_scope_name + '/dilation_rate']
 109         else:
 110             dnode = None
 111
 112         # the BiasAdd name is possible be changed into the output name,
 113         # if activation is None, and BiasAdd.next is the last op which is Identity
 114         if conv2d_scope_name + '/BiasAdd' in self.edges:
 115             anode = self.edges[conv2d_scope_name + '/BiasAdd'][0]
 116         else:
 117             anode = None
 118         return knode, bnode, dnode, anode
 119
 120
 121     def dump_conv2d_to_file(self, node, f):
 122         assert(node.op == 'Conv2D')
 123         self.layer_number = self.layer_number + 1
 124         self.converted_nodes.add(node.name)
 125
 126         scope_name = TFConverter.get_scope_name(node.name)
 127         #knode for kernel, bnode for bias, dnode for dilation, anode for activation
 128         knode, bnode, dnode, anode = self.get_conv2d_params(scope_name)
 129
 130         if dnode is not None:
 131             dilation = struct.unpack('i', dnode.attr['value'].tensor.tensor_content[0:4])[0]
 132         else:
 133             dilation = 1
 134
 135         if anode is not None:
 136             activation = anode.op
 137         else:
 138             activation = 'None'
 139
 140         padding = node.attr['padding'].s.decode("utf-8")
 141         # conv2d with dilation > 1 generates tens of nodes, not easy to parse them, so use this tricky method.
 142         if dilation > 1 and scope_name + '/stack' in self.name_node_dict:
 143             if self.name_node_dict[scope_name + '/stack'].op == "Const":
 144                 padding = 'SAME'
 145         padding = self.conv_paddings[padding]
 146
 147         ktensor = knode.attr['value'].tensor
 148         filter_height = ktensor.tensor_shape.dim[0].size
 149         filter_width = ktensor.tensor_shape.dim[1].size
 150         in_channels = ktensor.tensor_shape.dim[2].size
 151         out_channels = ktensor.tensor_shape.dim[3].size
 152         kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
 153         kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels)
 154         kernel = np.transpose(kernel, [3, 0, 1, 2])
 155
 156         np.array([self.op2code[node.op], dilation, padding, self.conv_activations[activation], in_channels, out_channels, filter_height], dtype=np.uint32).tofile(f)
 157         kernel.tofile(f)
 158
 159         btensor = bnode.attr['value'].tensor
 160         if btensor.tensor_shape.dim[0].size == 1:
 161             bias = struct.pack("f", btensor.float_val[0])
 162         else:
 163             bias = btensor.tensor_content
 164         f.write(bias)
 165
 166         input_name = self.conv2d_scopename_inputname_dict[scope_name]
 167         input_operand_index = self.add_operand(input_name, Operand.IOTYPE_INPUT)
 168
 169         if anode is not None:
 170             output_operand_index = self.add_operand(anode.name, Operand.IOTYPE_OUTPUT)
 171         else:
 172             output_operand_index = self.add_operand(self.edges[bnode.name][0].name, Operand.IOTYPE_OUTPUT)
 173         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 174
 175
 176     def dump_depth2space_to_file(self, node, f):
 177         assert(node.op == 'DepthToSpace')
 178         self.layer_number = self.layer_number + 1
 179         block_size = node.attr['block_size'].i
 180         np.array([self.op2code[node.op], block_size], dtype=np.uint32).tofile(f)
 181         self.converted_nodes.add(node.name)
 182         input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
 183         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 184         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 185
 186
 187     def dump_mirrorpad_to_file(self, node, f):
 188         assert(node.op == 'MirrorPad')
 189         self.layer_number = self.layer_number + 1
 190         mode = node.attr['mode'].s
 191         mode = self.mirrorpad_mode[mode.decode("utf-8")]
 192         np.array([self.op2code[node.op], mode], dtype=np.uint32).tofile(f)
 193         pnode = self.name_node_dict[node.input[1]]
 194         self.converted_nodes.add(pnode.name)
 195         paddings = pnode.attr['value'].tensor.tensor_content
 196         f.write(paddings)
 197         self.converted_nodes.add(node.name)
 198         input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
 199         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 200         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 201
 202
 203     def dump_maximum_to_file(self, node, f):
 204         assert(node.op == 'Maximum')
 205         self.layer_number = self.layer_number + 1
 206         ynode = self.name_node_dict[node.input[1]]
 207         y = ynode.attr['value'].tensor.float_val[0]
 208         np.array([self.op2code[node.op]], dtype=np.uint32).tofile(f)
 209         np.array([y], dtype=np.float32).tofile(f)
 210         self.converted_nodes.add(node.name)
 211         input_operand_index = self.add_operand(node.input[0], Operand.IOTYPE_INPUT)
 212         output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
 213         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 214
 215
 216     def dump_layers_to_file(self, f):
 217         for node in self.nodes:
 218             if node.name in self.converted_nodes:
 219                 continue
 220
 221             # conv2d with dilation generates very complex nodes, so handle it in special
 222             scope_name = TFConverter.get_scope_name(node.name)
 223             if scope_name in self.conv2d_scope_names:
 224                 if node.op == 'Conv2D':
 225                     self.dump_conv2d_to_file(node, f)
 226                 continue
 227
 228             if node.op == 'DepthToSpace':
 229                 self.dump_depth2space_to_file(node, f)
 230             elif node.op == 'MirrorPad':
 231                 self.dump_mirrorpad_to_file(node, f)
 232             elif node.op == 'Maximum':
 233                 self.dump_maximum_to_file(node, f)
 234
 235
 236     def dump_operands_to_file(self, f):
 237             operands = sorted(self.name_operand_dict.values())
 238             for operand in operands:
 239                 #print('{}'.format(operand))
 240                 np.array([operand.index, len(operand.name)], dtype=np.uint32).tofile(f)
 241                 f.write(operand.name.encode('utf-8'))
 242                 np.array([operand.iotype, operand.dtype], dtype=np.uint32).tofile(f)
 243                 np.array([operand.dims[0], operand.dims[1], operand.dims[2], operand.dims[3]], dtype=np.uint32).tofile(f)
 244
 245
 246     def dump_to_file(self):
 247         with open(self.outfile, 'wb') as f:
 248             f.write(header.str.encode('utf-8'))
 249             np.array([header.major, header.minor], dtype=np.uint32).tofile(f)
 250             self.dump_layers_to_file(f)
 251             self.dump_operands_to_file(f)
 252             np.array([self.layer_number, len(self.name_operand_dict)], dtype=np.uint32).tofile(f)
 253
 254
 255     def generate_name_node_dict(self):
 256         for node in self.nodes:
 257             self.name_node_dict[node.name] = node
 258
 259
 260     def generate_output_names(self):
 261         used_names = []
 262         for node in self.nodes:
 263             for input in node.input:
 264                 used_names.append(input)
 265
 266         for node in self.nodes:
 267             if node.name not in used_names:
 268                 self.output_names.append(node.name)
 269
 270
 271     def remove_identity(self):
 272         id_nodes = []
 273         id_dict = {}
 274         for node in self.nodes:
 275             if node.op == 'Identity':
 276                 name = node.name
 277                 input = node.input[0]
 278                 id_nodes.append(node)
 279                 # do not change the output name
 280                 if name in self.output_names:
 281                     self.name_node_dict[input].name = name
 282                     self.name_node_dict[name] = self.name_node_dict[input]
 283                     del self.name_node_dict[input]
 284                 else:
 285                     id_dict[name] = input
 286
 287         for idnode in id_nodes:
 288             self.nodes.remove(idnode)
 289
 290         for node in self.nodes:
 291             for i in range(len(node.input)):
 292                 input = node.input[i]
 293                 if input in id_dict:
 294                     node.input[i] = id_dict[input]
 295
 296
 297     def generate_edges(self):
 298         for node in self.nodes:
 299             for input in node.input:
 300                 if input in self.edges:
 301                     self.edges[input].append(node)
 302                 else:
 303                     self.edges[input] = [node]
 304
 305
 306     @staticmethod
 307     def get_scope_name(name):
 308         index = name.rfind('/')
 309         if index == -1:
 310             return ""
 311         return name[0:index]
 312
 313
 314     def generate_conv2d_scope_info(self):
 315         # conv2d is a sub block in graph, get the scope name
 316         for node in self.nodes:
 317             if node.op == 'Conv2D':
 318                 scope = TFConverter.get_scope_name(node.name)
 319                 self.conv2d_scope_names.add(scope)
 320
 321         # get the input name to the conv2d sub block
 322         for node in self.nodes:
 323             scope = TFConverter.get_scope_name(node.name)
 324             if scope in self.conv2d_scope_names:
 325                 if node.op == 'Conv2D' or node.op == 'Shape':
 326                     for inp in node.input:
 327                         if TFConverter.get_scope_name(inp) != scope:
 328                             self.conv2d_scopename_inputname_dict[scope] = inp
 329
 330
 331     def run(self):
 332         self.generate_name_node_dict()
 333         self.generate_output_names()
 334         self.remove_identity()
 335         self.generate_edges()
 336         self.generate_conv2d_scope_info()
 337
 338         if self.dump4tb:
 339             self.dump_for_tensorboard()
 340
 341         self.dump_to_file()
 342
 343
 344 def convert_from_tensorflow(infile, outfile, dump4tb):
 345     with open(infile, 'rb') as f:
 346         # read the file in .proto format
 347         graph_def = tf.GraphDef()
 348         graph_def.ParseFromString(f.read())
 349         nodes = graph_def.node
 350
 351     converter = TFConverter(graph_def, nodes, outfile, dump4tb)
 352     converter.run()