For example, given TensorFlow model file espcn.pb, to generate native model file espcn.model, just run: python convert.py espcn.pb In current implementation, the native model file is generated for specific dnn network with hard-code python scripts maintained out of ffmpeg. For example, srcnn network used by vf_sr is generated with https://github.com/HighVoltageRocknRoll/sr/blob/master/generate_header_and_model.py#L85 In this patch, the script is designed as a general solution which converts general TensorFlow model .pb file into .model file. The script now has some tricky to be compatible with current implemention, will be refined step by step. The script is also added into ffmpeg source tree. It is expected there will be many more patches and community needs the ownership of it. Another technical direction is to do the conversion in c/c++ code within ffmpeg source tree. While .pb file is organized with protocol buffers, it is not easy to do such work with tiny c/c++ code, see more discussion at http://ffmpeg.org/pipermail/ffmpeg-devel/2019-May/244496.html. So, choose the python script.
Showing the changes to tools/python (2 of 3 files). Show all
Unified · Split
| @@ -0,0 +1,52 @@ | |||
| 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 | # verified with Python 3.5.2 on Ubuntu 16.04 | ||
| 21 | import argparse | ||
| 22 | import os | ||
| 23 | from convert_from_tensorflow import * | ||
| 24 | |||
| 25 | def get_arguments(): | ||
| 26 | parser = argparse.ArgumentParser(description='generate native mode model with weights from deep learning model') | ||
| 27 | parser.add_argument('--outdir', type=str, default='./', help='where to put generated files') | ||
| 28 | parser.add_argument('--infmt', type=str, default='tensorflow', help='format of the deep learning model') | ||
| 29 | parser.add_argument('infile', help='path to the deep learning model with weights') | ||
| 30 | |||
| 31 | return parser.parse_args() | ||
| 32 | |||
| 33 | def main(): | ||
| 34 | args = get_arguments() | ||
| 35 | |||
| 36 | if not os.path.isfile(args.infile): | ||
| 37 | print('the specified input file %s does not exist' % args.infile) | ||
| 38 | exit(1) | ||
| 39 | |||
| 40 | if not os.path.exists(args.outdir): | ||
| 41 | print('create output directory %s' % args.outdir) | ||
| 42 | os.mkdir(args.outdir) | ||
| 43 | |||
| 44 | basefile = os.path.split(args.infile)[1] | ||
| 45 | basefile = os.path.splitext(basefile)[0] | ||
| 46 | outfile = os.path.join(args.outdir, basefile) + '.model' | ||
| 47 | |||
| 48 | if args.infmt == 'tensorflow': | ||
| 49 | convert_from_tensorflow(args.infile, outfile) | ||
| 50 | |||
| 51 | if __name__ == '__main__': | ||
| 52 | main() | ||
| @@ -0,0 +1,201 @@ | |||
| 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 | |||
| 24 | __all__ = ['convert_from_tensorflow'] | ||
| 25 | |||
| 26 | # as the first step to be compatible with vf_sr, it is not general. | ||
| 27 | # it will be refined step by step. | ||
| 28 | |||
| 29 | class TFConverter: | ||
| 30 | def __init__(self, graph_def, nodes, outfile): | ||
| 31 | self.graph_def = graph_def | ||
| 32 | self.nodes = nodes | ||
| 33 | self.outfile = outfile | ||
| 34 | self.layer_number = 0 | ||
| 35 | self.output_names = [] | ||
| 36 | self.name_node_dict = {} | ||
| 37 | self.edges = {} | ||
| 38 | self.conv_activations = {'Relu':0, 'Tanh':1, 'Sigmoid':2, 'LeakyRelu':4} | ||
| 39 | self.conv_paddings = {'VALID':2, 'SAME':1} | ||
| 40 | self.converted_nodes = set() | ||
| 41 | self.op2code = {'Conv2D':1, 'DepthToSpace':2} | ||
| 42 | |||
| 43 | |||
| 44 | def dump_for_tensorboard(self): | ||
| 45 | graph = tf.get_default_graph() | ||
| 46 | tf.import_graph_def(self.graph_def, name="") | ||
| 47 | # tensorboard --logdir=/tmp/graph | ||
| 48 | tf.summary.FileWriter('/tmp/graph', graph) | ||
| 49 | |||
| 50 | |||
| 51 | def get_conv2d_params(self, node): | ||
| 52 | knode = self.name_node_dict[node.input[1]] | ||
| 53 | bnode = None | ||
| 54 | activation = 'None' | ||
| 55 | next = self.edges[node.name][0] | ||
| 56 | if next.op == 'BiasAdd': | ||
| 57 | self.converted_nodes.add(next.name) | ||
| 58 | bnode = self.name_node_dict[next.input[1]] | ||
| 59 | next = self.edges[next.name][0] | ||
| 60 | if next.op in self.conv_activations: | ||
| 61 | self.converted_nodes.add(next.name) | ||
| 62 | activation = next.op | ||
| 63 | return knode, bnode, activation | ||
| 64 | |||
| 65 | |||
| 66 | def dump_conv2d_to_file(self, node, f): | ||
| 67 | assert(node.op == 'Conv2D') | ||
| 68 | self.layer_number = self.layer_number + 1 | ||
| 69 | self.converted_nodes.add(node.name) | ||
| 70 | knode, bnode, activation = self.get_conv2d_params(node) | ||
| 71 | |||
| 72 | dilation = node.attr['dilations'].list.i[0] | ||
| 73 | padding = node.attr['padding'].s | ||
| 74 | padding = self.conv_paddings[padding.decode("utf-8")] | ||
| 75 | |||
| 76 | ktensor = knode.attr['value'].tensor | ||
| 77 | filter_height = ktensor.tensor_shape.dim[0].size | ||
| 78 | filter_width = ktensor.tensor_shape.dim[1].size | ||
| 79 | in_channels = ktensor.tensor_shape.dim[2].size | ||
| 80 | out_channels = ktensor.tensor_shape.dim[3].size | ||
| 81 | kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32) | ||
| 82 | kernel = kernel.reshape(filter_height, filter_width, in_channels, out_channels) | ||
| 83 | kernel = np.transpose(kernel, [3, 0, 1, 2]) | ||
| 84 | |||
| 85 | np.array([self.op2code[node.op], dilation, padding, self.conv_activations[activation], in_channels, out_channels, filter_height], dtype=np.uint32).tofile(f) | ||
| 86 | kernel.tofile(f) | ||
| 87 | |||
| 88 | btensor = bnode.attr['value'].tensor | ||
| 89 | if btensor.tensor_shape.dim[0].size == 1: | ||
| 90 | bias = struct.pack("f", btensor.float_val[0]) | ||
| 91 | else: | ||
| 92 | bias = btensor.tensor_content | ||
| 93 | f.write(bias) | ||
| 94 | |||
| 95 | |||
| 96 | def dump_depth2space_to_file(self, node, f): | ||
| 97 | assert(node.op == 'DepthToSpace') | ||
| 98 | self.layer_number = self.layer_number + 1 | ||
| 99 | block_size = node.attr['block_size'].i | ||
| 100 | np.array([self.op2code[node.op], block_size], dtype=np.uint32).tofile(f) | ||
| 101 | self.converted_nodes.add(node.name) | ||
| 102 | |||
| 103 | |||
| 104 | def generate_layer_number(self): | ||
| 105 | # in current hard code implementation, the layer number is the first data written to the native model file | ||
| 106 | # it is not easy to know it at the beginning time in the general converter, so first do a dry run for compatibility | ||
| 107 | # will be refined later. | ||
| 108 | with open('/tmp/tmp.model', 'wb') as f: | ||
| 109 | self.dump_layers_to_file(f) | ||
| 110 | self.converted_nodes.clear() | ||
| 111 | |||
| 112 | |||
| 113 | def dump_layers_to_file(self, f): | ||
| 114 | for node in self.nodes: | ||
| 115 | if node.name in self.converted_nodes: | ||
| 116 | continue | ||
| 117 | if node.op == 'Conv2D': | ||
| 118 | self.dump_conv2d_to_file(node, f) | ||
| 119 | elif node.op == 'DepthToSpace': | ||
| 120 | self.dump_depth2space_to_file(node, f) | ||
| 121 | |||
| 122 | |||
| 123 | def dump_to_file(self): | ||
| 124 | self.generate_layer_number() | ||
| 125 | with open(self.outfile, 'wb') as f: | ||
| 126 | np.array([self.layer_number], dtype=np.uint32).tofile(f) | ||
| 127 | self.dump_layers_to_file(f) | ||
| 128 | |||
| 129 | |||
| 130 | def generate_name_node_dict(self): | ||
| 131 | for node in self.nodes: | ||
| 132 | self.name_node_dict[node.name] = node | ||
| 133 | |||
| 134 | |||
| 135 | def generate_output_names(self): | ||
| 136 | used_names = [] | ||
| 137 | for node in self.nodes: | ||
| 138 | for input in node.input: | ||
| 139 | used_names.append(input) | ||
| 140 | |||
| 141 | for node in self.nodes: | ||
| 142 | if node.name not in used_names: | ||
| 143 | self.output_names.append(node.name) | ||
| 144 | |||
| 145 | |||
| 146 | def remove_identity(self): | ||
| 147 | id_nodes = [] | ||
| 148 | id_dict = {} | ||
| 149 | for node in self.nodes: | ||
| 150 | if node.op == 'Identity': | ||
| 151 | name = node.name | ||
| 152 | input = node.input[0] | ||
| 153 | id_nodes.append(node) | ||
| 154 | # do not change the output name | ||
| 155 | if name in self.output_names: | ||
| 156 | self.name_node_dict[input].name = name | ||
| 157 | self.name_node_dict[name] = self.name_node_dict[input] | ||
| 158 | del self.name_node_dict[input] | ||
| 159 | else: | ||
| 160 | id_dict[name] = input | ||
| 161 | |||
| 162 | for idnode in id_nodes: | ||
| 163 | self.nodes.remove(idnode) | ||
| 164 | |||
| 165 | for node in self.nodes: | ||
| 166 | for i in range(len(node.input)): | ||
| 167 | input = node.input[i] | ||
| 168 | if input in id_dict: | ||
| 169 | node.input[i] = id_dict[input] | ||
| 170 | |||
| 171 | |||
| 172 | def generate_edges(self): | ||
| 173 | for node in self.nodes: | ||
| 174 | for input in node.input: | ||
| 175 | if input in self.edges: | ||
| 176 | self.edges[input].append(node) | ||
| 177 | else: | ||
| 178 | self.edges[input] = [node] | ||
| 179 | |||
| 180 | |||
| 181 | def run(self): | ||
| 182 | self.generate_name_node_dict() | ||
| 183 | self.generate_output_names() | ||
| 184 | self.remove_identity() | ||
| 185 | self.generate_edges() | ||
| 186 | |||
| 187 | #check the graph with tensorboard with human eyes | ||
| 188 | #self.dump_for_tensorboard() | ||
| 189 | |||
| 190 | self.dump_to_file() | ||
| 191 | |||
| 192 | |||
| 193 | def convert_from_tensorflow(infile, outfile): | ||
| 194 | with open(infile, 'rb') as f: | ||
| 195 | # read the file in .proto format | ||
| 196 | graph_def = tf.GraphDef() | ||
| 197 | graph_def.ParseFromString(f.read()) | ||
| 198 | nodes = graph_def.node | ||
| 199 | |||
| 200 | converter = TFConverter(graph_def, nodes, outfile) | ||
| 201 | converter.run() | ||