[FFmpeg-devel] [PATCH 1/2] libavfilter/dnn: add script to convert TensorFlow model (.pb) to native model (.model)
Guo, Yejun
yejun.guo at intel.com
Tue May 28 15:06:38 EEST 2019
> -----Original Message-----
> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces at ffmpeg.org] On Behalf Of
> Liu Steven
> Sent: Tuesday, May 28, 2019 6:00 PM
> To: FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
> Cc: Liu Steven <lq at chinaffmpeg.org>; Guo, Yejun <yejun.guo at intel.com>
> Subject: Re: [FFmpeg-devel] [PATCH 1/2] libavfilter/dnn: add script to convert
> TensorFlow model (.pb) to native model (.model)
>
>
>
> > 在 2019年5月28日，下午4:01，Guo, Yejun <yejun.guo at intel.com> 写
> 道：
> >
> > 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_a
> nd_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.
> >
> > Signed-off-by: Guo, Yejun <yejun.guo at intel.com>
> > ---
> > libavfilter/dnn/python/convert.py | 52 ++++++
> > libavfilter/dnn/python/convert_from_tensorflow.py | 200
> ++++++++++++++++++++++
> What about move them into ./tools/ ?
yes, this is another feasible option. My idea is to put all the dnn stuffs together,
other dnn .h/.c files will be at libavfilter/dnn/
>
> > 2 files changed, 252 insertions(+)
> > create mode 100644 libavfilter/dnn/python/convert.py
> > create mode 100644 libavfilter/dnn/python/convert_from_tensorflow.py
> >
> > diff --git a/libavfilter/dnn/python/convert.py
> b/libavfilter/dnn/python/convert.py
> > new file mode 100644
> > index 0000000..662b429
> > --- /dev/null
> > +++ b/libavfilter/dnn/python/convert.py
> > @@ -0,0 +1,52 @@
> > +# Copyright (c) 2019 Guo Yejun
> > +#
> > +# This file is part of FFmpeg.
> > +#
> > +# FFmpeg is free software; you can redistribute it and/or
> > +# modify it under the terms of the GNU Lesser General Public
> > +# License as published by the Free Software Foundation; either
> > +# version 2.1 of the License, or (at your option) any later version.
> > +#
> > +# FFmpeg is distributed in the hope that it will be useful,
> > +# but WITHOUT ANY WARRANTY; without even the implied warranty of
> > +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> GNU
> > +# Lesser General Public License for more details.
> > +#
> > +# You should have received a copy of the GNU Lesser General Public
> > +# License along with FFmpeg; if not, write to the Free Software
> > +# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301
> USA
> > +#
> ================================================================
> ==============
> > +
> > +# verified with Python 3.5.2 on Ubuntu 16.04
> > +import argparse
> > +import os
> > +from convert_from_tensorflow import *
> > +
> > +def get_arguments():
> > + parser = argparse.ArgumentParser(description='generate native mode
> model with weights from deep learning model')
> > + parser.add_argument('--outdir', type=str, default='./', help='where to
> put generated files')
> > + parser.add_argument('--infmt', type=str, default='tensorflow',
> help='format of the deep learning model')
> > + parser.add_argument('infile', help='path to the deep learning model
> with weights')
> > +
> > + return parser.parse_args()
> > +
> > +def main():
> > + args = get_arguments()
> > +
> > + if not os.path.isfile(args.infile):
> > + print('the specified input file %s does not exist' % args.infile)
> > + exit(1)
> > +
> > + if not os.path.exists(args.outdir):
> > + print('create output directory %s' % args.outdir)
> > + os.mkdir(args.outdir)
> > +
> > + basefile = os.path.split(args.infile)[1]
> > + basefile = os.path.splitext(basefile)[0]
> > + outfile = os.path.join(args.outdir, basefile) + '.model'
> > +
> > + if args.infmt == 'tensorflow':
> > + convert_from_tensorflow(args.infile, outfile)
> > +
> > +if __name__ == '__main__':
> > + main()
> > diff --git a/libavfilter/dnn/python/convert_from_tensorflow.py
> b/libavfilter/dnn/python/convert_from_tensorflow.py
> > new file mode 100644
> > index 0000000..436ec0e
> > --- /dev/null
> > +++ b/libavfilter/dnn/python/convert_from_tensorflow.py
> > @@ -0,0 +1,200 @@
> > +# Copyright (c) 2019 Guo Yejun
> > +#
> > +# This file is part of FFmpeg.
> > +#
> > +# FFmpeg is free software; you can redistribute it and/or
> > +# modify it under the terms of the GNU Lesser General Public
> > +# License as published by the Free Software Foundation; either
> > +# version 2.1 of the License, or (at your option) any later version.
> > +#
> > +# FFmpeg is distributed in the hope that it will be useful,
> > +# but WITHOUT ANY WARRANTY; without even the implied warranty of
> > +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> GNU
> > +# Lesser General Public License for more details.
> > +#
> > +# You should have received a copy of the GNU Lesser General Public
> > +# License along with FFmpeg; if not, write to the Free Software
> > +# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301
> USA
> > +#
> ================================================================
> ==============
> > +
> > +import tensorflow as tf
> > +import numpy as np
> > +import sys, struct
> > +
> > +__all__ = ['convert_from_tensorflow']
> > +
> > +# as the first step to be compatible with vf_sr, it is not general.
> > +# it will be refined step by step.
> > +
> > +class TFConverter:
> > + def __init__(self, graph_def, nodes, outfile):
> > + self.graph_def = graph_def
> > + self.nodes = nodes
> > + self.outfile = outfile
> > + self.layer_number = 0
> > + self.output_names = []
> > + self.name_node_dict = {}
> > + self.edges = {}
> > + self.conv_activations = {'Relu':0, 'Tanh':1, 'Sigmoid':2,
> 'LeakyRelu':4}
> > + self.conv_paddings = {'VALID':2, 'SAME':1}
> > + self.converted_nodes = set()
> > +
> > +
> > + def dump_for_tensorboard(self):
> > + graph = tf.get_default_graph()
> > + tf.import_graph_def(self.graph_def, name="")
> > + # tensorboard --logdir=/tmp/graph
> > + tf.summary.FileWriter('/tmp/graph', graph)
> > +
> > +
> > + def get_conv2d_params(self, node):
> > + knode = self.name_node_dict[node.input[1]]
> > + bnode = None
> > + activation = 'None'
> > + next = self.edges[node.name][0]
> > + if next.op == 'BiasAdd':
> > + self.converted_nodes.add(next.name)
> > + bnode = self.name_node_dict[next.input[1]]
> > + next = self.edges[next.name][0]
> > + if next.op in self.conv_activations:
> > + self.converted_nodes.add(next.name)
> > + activation = next.op
> > + return knode, bnode, activation
> > +
> > +
> > + def dump_conv2d_to_file(self, node, f):
> > + assert(node.op == 'Conv2D')
> > + self.layer_number = self.layer_number + 1
> > + self.converted_nodes.add(node.name)
> > + knode, bnode, activation = self.get_conv2d_params(node)
> > +
> > + dilation = node.attr['dilations'].list.i[0]
> > + padding = node.attr['padding'].s
> > + padding = self.conv_paddings[padding.decode("utf-8")]
> > +
> > + ktensor = knode.attr['value'].tensor
> > + filter_height = ktensor.tensor_shape.dim[0].size
> > + filter_width = ktensor.tensor_shape.dim[1].size
> > + in_channels = ktensor.tensor_shape.dim[2].size
> > + out_channels = ktensor.tensor_shape.dim[3].size
> > + kernel = np.frombuffer(ktensor.tensor_content, dtype=np.float32)
> > + kernel = kernel.reshape(filter_height, filter_width, in_channels,
> out_channels)
> > + kernel = np.transpose(kernel, [3, 0, 1, 2])
> > +
> > + np.array([1, dilation, padding, self.conv_activations[activation],
> in_channels, out_channels, filter_height], dtype=np.uint32).tofile(f)
> > + kernel.tofile(f)
> > +
> > + btensor = bnode.attr['value'].tensor
> > + if btensor.tensor_shape.dim[0].size == 1:
> > + bias = struct.pack("f", btensor.float_val[0])
> > + else:
> > + bias = btensor.tensor_content
> > + f.write(bias)
> > +
> > +
> > + def dump_depth2space_to_file(self, node, f):
> > + assert(node.op == 'DepthToSpace')
> > + self.layer_number = self.layer_number + 1
> > + block_size = node.attr['block_size'].i
> > + np.array([2, block_size], dtype=np.uint32).tofile(f)
> > + self.converted_nodes.add(node.name)
> > +
> > +
> > + def generate_layer_number(self):
> > + # in current hard code implementation, the layer number is the
> first data written to the native model file
> > + # it is not easy to know it at the beginning time in the general
> converter, so first do a dry run for compatibility
> > + # will be refined later.
> > + with open('/tmp/tmp.model', 'wb') as f:
> > + self.dump_layers_to_file(f)
> > + self.converted_nodes.clear()
> > +
> > +
> > + def dump_layers_to_file(self, f):
> > + for node in self.nodes:
> > + if node.name in self.converted_nodes:
> > + continue
> > + if node.op == 'Conv2D':
> > + self.dump_conv2d_to_file(node, f)
> > + elif node.op == 'DepthToSpace':
> > + self.dump_depth2space_to_file(node, f)
> > +
> > +
> > + def dump_to_file(self):
> > + self.generate_layer_number()
> > + with open(self.outfile, 'wb') as f:
> > + np.array([self.layer_number], dtype=np.uint32).tofile(f)
> > + self.dump_layers_to_file(f)
> > +
> > +
> > + def generate_name_node_dict(self):
> > + for node in self.nodes:
> > + self.name_node_dict[node.name] = node
> > +
> > +
> > + def generate_output_names(self):
> > + used_names = []
> > + for node in self.nodes:
> > + for input in node.input:
> > + used_names.append(input)
> > +
> > + for node in self.nodes:
> > + if node.name not in used_names:
> > + self.output_names.append(node.name)
> > +
> > +
> > + def remove_identity(self):
> > + id_nodes = []
> > + id_dict = {}
> > + for node in self.nodes:
> > + if node.op == 'Identity':
> > + name = node.name
> > + input = node.input[0]
> > + id_nodes.append(node)
> > + # do not change the output name
> > + if name in self.output_names:
> > + self.name_node_dict[input].name = name
> > + self.name_node_dict[name] =
> self.name_node_dict[input]
> > + del self.name_node_dict[input]
> > + else:
> > + id_dict[name] = input
> > +
> > + for idnode in id_nodes:
> > + self.nodes.remove(idnode)
> > +
> > + for node in self.nodes:
> > + for i in range(len(node.input)):
> > + input = node.input[i]
> > + if input in id_dict:
> > + node.input[i] = id_dict[input]
> > +
> > +
> > + def generate_edges(self):
> > + for node in self.nodes:
> > + for input in node.input:
> > + if input in self.edges:
> > + self.edges[input].append(node)
> > + else:
> > + self.edges[input] = [node]
> > +
> > +
> > + def run(self):
> > + self.generate_name_node_dict()
> > + self.generate_output_names()
> > + self.remove_identity()
> > + self.generate_edges()
> > +
> > + #check the graph with tensorboard with human eyes
> > + #self.dump_for_tensorboard()
> > +
> > + self.dump_to_file()
> > +
> > +
> > +def convert_from_tensorflow(infile, outfile):
> > + with open(infile, 'rb') as f:
> > + # read the file in .proto format
> > + graph_def = tf.GraphDef()
> > + graph_def.ParseFromString(f.read())
> > + nodes = graph_def.node
> > +
> > + converter = TFConverter(graph_def, nodes, outfile)
> > + converter.run()
> > --
> > 2.7.4
> >
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >
> > To unsubscribe, visit link above, or email
> > ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>
>
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
More information about the ffmpeg-devel
mailing list