NextCloud/dlib
1
0
Fork 0
mirror of https://github.com/davisking/dlib.git synced 2024-11-01 10:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity
724cb5006f
dlib/tools/convert_dlib_nets_to_caffe/main.cpp

698 lines
28 KiB
C++
Raw Normal View History

Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
#include <dlib/xml_parser.h>
#include <dlib/matrix.h>
#include <fstream>
#include <vector>
#include <stack>
#include <set>
#include <dlib/string.h>
using namespace std;
using namespace dlib;
// ----------------------------------------------------------------------------------------
// Only these computational layers have parameters
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
const std::set<string> comp_tags_with_params = {"fc", "fc_no_bias", "con", "affine_con", "affine_fc", "affine", "prelu"};
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
struct layer
{
string type; // comp, loss, or input
int idx;
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
matrix<long,4,1> output_tensor_shape; // (N,K,NR,NC)
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
string detail_name; // The name of the tag inside the layer tag. e.g. fc, con, max_pool, input_rgb_image.
std::map<string,double> attributes;
matrix<double> params;
long tag_id = -1; // If this isn't -1 then it means this layer was tagged, e.g. wrapped with tag2<> giving tag_id==2
long skip_id = -1; // If this isn't -1 then it means this layer draws its inputs from
// the most recent layer with tag_id==skip_id rather than its immediate predecessor.
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
double attribute (const string& key) const
{
auto i = attributes.find(key);
if (i != attributes.end())
return i->second;
else
throw dlib::error("Layer doesn't have the requested attribute '" + key + "'.");
}
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
string caffe_layer_name() const
{
if (type == "input")
return "data";
else
return detail_name+to_string(idx);
}
};
// ----------------------------------------------------------------------------------------
std::vector<layer> parse_dlib_xml(
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
const matrix<long,4,1>& input_tensor_shape,
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
const string& xml_filename
);
// ----------------------------------------------------------------------------------------
template <typename iterator>
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
const layer& find_layer (
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
iterator i,
long tag_id
)
/*!
requires
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
- i is a reverse iterator pointing to a layer in the list of layers produced by parse_dlib_xml().
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
- i is not an input layer.
ensures
- if (tag_id == -1) then
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
- returns the previous layer (i.e. closer to the input) to layer i.
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
- else
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
- returns the previous layer (i.e. closer to the input) to layer i with the
given tag_id.
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
!*/
{
if (tag_id == -1)
{
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
return *(i-1);
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else
{
while(true)
{
i--;
// if we hit the end of the network before we found what we were looking for
if (i->tag_id == tag_id)
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
return *i;
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
if (i->type == "input")
throw dlib::error("Network definition is bad, a layer wanted to skip back to a non-existing layer.");
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
}
}
template <typename iterator>
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
const layer& find_input_layer (iterator i) { return find_layer(i, i->skip_id); }
template <typename iterator>
string find_layer_caffe_name (
iterator i,
long tag_id
)
{
return find_layer(i,tag_id).caffe_layer_name();
}
template <typename iterator>
string find_input_layer_caffe_name (iterator i) { return find_input_layer(i).caffe_layer_name(); }
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
// ----------------------------------------------------------------------------------------
template <typename EXP>
void print_as_np_array(std::ostream& out, const matrix_exp<EXP>& m)
{
out << "np.array([";
for (auto x : m)
out << x << ",";
out << "], dtype='float32')";
}
// ----------------------------------------------------------------------------------------
Fixed spelling error
2017-05-08 07:40:42 +08:00
void convert_dlib_xml_to_caffe_python_code(
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
const string& xml_filename,
const long N,
const long K,
const long NR,
const long NC
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
)
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
const string out_filename = left_substr(xml_filename,".") + "_dlib_to_caffe_model.py";
cout << "Writing model to " << out_filename << endl;
ofstream fout(out_filename);
fout.precision(9);
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
const auto layers = parse_dlib_xml({N,K,NR,NC}, xml_filename);
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
Added more comments
2017-05-08 05:16:08 +08:00
fout << "#\n";
fout << "# !!! This file was automatically generated by dlib's tools/convert_dlib_nets_to_caffe utility. !!!\n";
fout << "# !!! It contains all the information from a dlib DNN network and lets you save it as a cafe model. !!!\n";
fout << "#\n";
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << "import caffe " << endl;
fout << "from caffe import layers as L, params as P" << endl;
fout << "import numpy as np" << endl;
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
// dlib nets don't commit to a batch size, so just use 1 as the default
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << "\n# Input tensor dimensions" << endl;
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
fout << "input_batch_size = " << N << ";" << endl;
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
if (layers.back().detail_name == "input_rgb_image")
{
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
fout << "input_num_channels = 3;" << endl;
fout << "input_num_rows = "<<NR<<";" << endl;
fout << "input_num_cols = "<<NC<<";" << endl;
if (K != 3)
throw dlib::error("The dlib model requires input tensors with NUM_CHANNELS==3, but the dtoc command line specified NUM_CHANNELS=="+to_string(K));
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (layers.back().detail_name == "input_rgb_image_sized")
{
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
fout << "input_num_channels = 3;" << endl;
fout << "input_num_rows = " << layers.back().attribute("nr") << ";" << endl;
fout << "input_num_cols = " << layers.back().attribute("nc") << ";" << endl;
if (NR != layers.back().attribute("nr"))
Made error message slightly better.
2017-05-23 07:11:00 +08:00
throw dlib::error("The dlib model requires input tensors with NUM_ROWS=="+to_string((long)layers.back().attribute("nr"))+", but the dtoc command line specified NUM_ROWS=="+to_string(NR));
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
if (NC != layers.back().attribute("nc"))
Made error message slightly better.
2017-05-23 07:11:00 +08:00
throw dlib::error("The dlib model requires input tensors with NUM_COLUMNS=="+to_string((long)layers.back().attribute("nc"))+", but the dtoc command line specified NUM_COLUMNS=="+to_string(NC));
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
if (K != 3)
throw dlib::error("The dlib model requires input tensors with NUM_CHANNELS==3, but the dtoc command line specified NUM_CHANNELS=="+to_string(K));
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (layers.back().detail_name == "input")
{
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
fout << "input_num_channels = 1;" << endl;
fout << "input_num_rows = "<<NR<<";" << endl;
fout << "input_num_cols = "<<NC<<";" << endl;
if (K != 1)
throw dlib::error("The dlib model requires input tensors with NUM_CHANNELS==1, but the dtoc command line specified NUM_CHANNELS=="+to_string(K));
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else
{
throw dlib::error("No known transformation from dlib's " + layers.back().detail_name + " layer to caffe.");
}
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << endl;
Added more comments
2017-05-08 05:16:08 +08:00
fout << "# Call this function to write the dlib DNN model out to file as a pair of caffe\n";
fout << "# definition and weight files. You can then use the network by loading it with\n";
fout << "# this statement: \n";
fout << "# net = caffe.Net(def_file, weights_file, caffe.TEST);\n";
fout << "#\n";
fout << "def save_as_caffe_model(def_file, weights_file):\n";
fout << " with open(def_file, 'w') as f: f.write(str(make_netspec()));\n";
fout << " net = caffe.Net(def_file, caffe.TEST);\n";
fout << " set_network_weights(net);\n";
fout << " net.save(weights_file);\n\n";
fout << "###############################################################################\n";
fout << "# EVERYTHING BELOW HERE DEFINES THE DLIB MODEL PARAMETERS #\n";
fout << "###############################################################################\n\n\n";
// -----------------------------------------------------------------------------------
// The next block of code outputs python code that defines the network architecture.
// -----------------------------------------------------------------------------------
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << "def make_netspec():" << endl;
fout << " # For reference, the only \"documentation\" about caffe layer parameters seems to be this page:\n";
fout << " # https://github.com/BVLC/caffe/blob/master/src/caffe/proto/caffe.proto\n" << endl;
fout << " n = caffe.NetSpec(); " << endl;
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
fout << " n.data,n.label = L.MemoryData(batch_size=input_batch_size, channels=input_num_channels, height=input_num_rows, width=input_num_cols, ntop=2)" << endl;
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
// iterate the layers starting with the input layer
for (auto i = layers.rbegin(); i != layers.rend(); ++i)
{
// skip input and loss layers
if (i->type == "loss" || i->type == "input")
continue;
if (i->detail_name == "con")
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.Convolution(n." << find_input_layer_caffe_name(i);
fout << ", num_output=" << i->attribute("num_filters");
fout << ", kernel_w=" << i->attribute("nc");
fout << ", kernel_h=" << i->attribute("nr");
fout << ", stride_w=" << i->attribute("stride_x");
fout << ", stride_h=" << i->attribute("stride_y");
fout << ", pad_w=" << i->attribute("padding_x");
fout << ", pad_h=" << i->attribute("padding_y");
fout << ");\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (i->detail_name == "relu")
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.ReLU(n." << find_input_layer_caffe_name(i);
fout << ");\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
Added missing sig and prelu layer support.
2017-05-08 04:23:35 +08:00
else if (i->detail_name == "sig")
{
fout << " n." << i->caffe_layer_name() << " = L.Sigmoid(n." << find_input_layer_caffe_name(i);
fout << ");\n";
}
else if (i->detail_name == "prelu")
{
fout << " n." << i->caffe_layer_name() << " = L.PReLU(n." << find_input_layer_caffe_name(i);
fout << ", channel_shared=True";
fout << ");\n";
}
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
else if (i->detail_name == "max_pool")
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.Pooling(n." << find_input_layer_caffe_name(i);
fout << ", pool=P.Pooling.MAX";
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
if (i->attribute("nc")==0)
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << ", global_pooling=True";
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
}
else
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << ", kernel_w=" << i->attribute("nc");
fout << ", kernel_h=" << i->attribute("nr");
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
}
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << ", stride_w=" << i->attribute("stride_x");
fout << ", stride_h=" << i->attribute("stride_y");
fout << ", pad_w=" << i->attribute("padding_x");
fout << ", pad_h=" << i->attribute("padding_y");
fout << ");\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (i->detail_name == "avg_pool")
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.Pooling(n." << find_input_layer_caffe_name(i);
fout << ", pool=P.Pooling.AVE";
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
if (i->attribute("nc")==0)
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << ", global_pooling=True";
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
}
else
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << ", kernel_w=" << i->attribute("nc");
fout << ", kernel_h=" << i->attribute("nr");
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
}
if (i->attribute("padding_x") != 0 || i->attribute("padding_y") != 0)
{
throw dlib::error("dlib and caffe implement pooling with non-zero padding differently, so you can't convert a "
"network with such pooling layers.");
}
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << ", stride_w=" << i->attribute("stride_x");
fout << ", stride_h=" << i->attribute("stride_y");
fout << ", pad_w=" << i->attribute("padding_x");
fout << ", pad_h=" << i->attribute("padding_y");
fout << ");\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (i->detail_name == "fc")
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.InnerProduct(n." << find_input_layer_caffe_name(i);
fout << ", num_output=" << i->attribute("num_outputs");
fout << ", bias_term=True";
fout << ");\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (i->detail_name == "fc_no_bias")
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.InnerProduct(n." << find_input_layer_caffe_name(i);
fout << ", num_output=" << i->attribute("num_outputs");
fout << ", bias_term=False";
fout << ");\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
else if (i->detail_name == "bn_con" || i->detail_name == "bn_fc")
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
{
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
throw dlib::error("Conversion from dlib's batch norm layers to caffe's isn't supported. Instead, "
Fixed affine_fc and clarified error message.
2017-05-08 04:57:34 +08:00
"you should put your dlib network into 'test mode' by switching batch norm layers to affine layers. "
"Then you can convert that 'test mode' network to caffe.");
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
else if (i->detail_name == "affine_con")
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.Scale(n." << find_input_layer_caffe_name(i);
fout << ", bias_term=True";
fout << ");\n";
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
}
else if (i->detail_name == "affine_fc")
{
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " n." << i->caffe_layer_name() << " = L.Scale(n." << find_input_layer_caffe_name(i);
fout << ", bias_term=True";
fout << ");\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (i->detail_name == "add_prev")
{
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
auto in_shape1 = find_input_layer(i).output_tensor_shape;
auto in_shape2 = find_layer(i,i->attribute("tag")).output_tensor_shape;
if (in_shape1 != in_shape2)
{
// if only the number of channels differs then we will use a dummy layer to
// pad with zeros. But otherwise we will throw an error.
if (in_shape1(0) == in_shape2(0) &&
in_shape1(2) == in_shape2(2) &&
in_shape1(3) == in_shape2(3))
{
fout << " n." << i->caffe_layer_name() << "_zeropad = L.DummyData(num=" << in_shape1(0);
fout << ", channels="<<std::abs(in_shape1(1)-in_shape2(1));
fout << ", height="<<in_shape1(2);
fout << ", width="<<in_shape1(3);
fout << ");\n";
string smaller_layer = find_input_layer_caffe_name(i);
string bigger_layer = find_layer_caffe_name(i, i->attribute("tag"));
if (in_shape1(1) > in_shape2(1))
swap(smaller_layer, bigger_layer);
fout << " n." << i->caffe_layer_name() << "_concat = L.Concat(n." << smaller_layer;
fout << ", n." << i->caffe_layer_name() << "_zeropad";
fout << ");\n";
fout << " n." << i->caffe_layer_name() << " = L.Eltwise(n." << i->caffe_layer_name() << "_concat";
fout << ", n." << bigger_layer;
fout << ", operation=P.Eltwise.SUM";
fout << ");\n";
}
else
{
std::ostringstream sout;
sout << "The dlib network contained an add_prev layer (layer idx " << i->idx << ") that adds two previous ";
sout << "layers with different output tensor dimensions. Caffe's equivalent layer, Eltwise, doesn't support ";
sout << "adding layers together with different dimensions. In the special case where the only difference is ";
sout << "in the number of channels, this converter program will add a dummy layer that outputs a tensor full of zeros ";
sout << "and concat it appropriately so this will work. However, this network you are converting has tensor dimensions ";
sout << "different in values other than the number of channels. In particular, here are the two tensor shapes (batch size, channels, rows, cols): ";
std::ostringstream sout2;
sout2 << wrap_string(sout.str()) << endl;
sout2 << trans(in_shape1);
sout2 << trans(in_shape2);
throw dlib::error(sout2.str());
}
}
else
{
fout << " n." << i->caffe_layer_name() << " = L.Eltwise(n." << find_input_layer_caffe_name(i);
fout << ", n." << find_layer_caffe_name(i, i->attribute("tag"));
fout << ", operation=P.Eltwise.SUM";
fout << ");\n";
}
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else
{
throw dlib::error("No known transformation from dlib's " + i->detail_name + " layer to caffe.");
}
}
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " return n.to_proto();\n\n" << endl;
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
Added missing sig and prelu layer support.
2017-05-08 04:23:35 +08:00
// -----------------------------------------------------------------------------------
// The next block of code outputs python code that populates all the filter weights.
// -----------------------------------------------------------------------------------
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << "def set_network_weights(net):\n";
fout << " # populate network parameters\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
// iterate the layers starting with the input layer
for (auto i = layers.rbegin(); i != layers.rend(); ++i)
{
// skip input and loss layers
if (i->type == "loss" || i->type == "input")
continue;
if (i->detail_name == "con")
{
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
const long num_filters = i->attribute("num_filters");
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
matrix<double> weights = trans(rowm(i->params,range(0,i->params.size()-num_filters-1)));
matrix<double> biases = trans(rowm(i->params,range(i->params.size()-num_filters, i->params.size()-1)));
// main filter weights
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " p = "; print_as_np_array(fout,weights); fout << ";\n";
fout << " p.shape = net.params['"<<i->caffe_layer_name()<<"'][0].data.shape;\n";
fout << " net.params['"<<i->caffe_layer_name()<<"'][0].data[:] = p;\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
// biases
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " p = "; print_as_np_array(fout,biases); fout << ";\n";
fout << " p.shape = net.params['"<<i->caffe_layer_name()<<"'][1].data.shape;\n";
fout << " net.params['"<<i->caffe_layer_name()<<"'][1].data[:] = p;\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (i->detail_name == "fc")
{
matrix<double> weights = trans(rowm(i->params, range(0,i->params.nr()-2)));
matrix<double> biases = rowm(i->params, i->params.nr()-1);
// main filter weights
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " p = "; print_as_np_array(fout,weights); fout << ";\n";
fout << " p.shape = net.params['"<<i->caffe_layer_name()<<"'][0].data.shape;\n";
fout << " net.params['"<<i->caffe_layer_name()<<"'][0].data[:] = p;\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
// biases
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " p = "; print_as_np_array(fout,biases); fout << ";\n";
fout << " p.shape = net.params['"<<i->caffe_layer_name()<<"'][1].data.shape;\n";
fout << " net.params['"<<i->caffe_layer_name()<<"'][1].data[:] = p;\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
else if (i->detail_name == "fc_no_bias")
{
matrix<double> weights = trans(i->params);
// main filter weights
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " p = "; print_as_np_array(fout,weights); fout << ";\n";
fout << " p.shape = net.params['"<<i->caffe_layer_name()<<"'][0].data.shape;\n";
fout << " net.params['"<<i->caffe_layer_name()<<"'][0].data[:] = p;\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
else if (i->detail_name == "affine_con" || i->detail_name == "affine_fc")
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
{
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
const long dims = i->params.size()/2;
matrix<double> gamma = trans(rowm(i->params,range(0,dims-1)));
matrix<double> beta = trans(rowm(i->params,range(dims, 2*dims-1)));
// set gamma weights
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " p = "; print_as_np_array(fout,gamma); fout << ";\n";
fout << " p.shape = net.params['"<<i->caffe_layer_name()<<"'][0].data.shape;\n";
fout << " net.params['"<<i->caffe_layer_name()<<"'][0].data[:] = p;\n";
Added support for all the other dlib layers that make sense.
2017-05-08 03:26:18 +08:00
// set beta weights
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
fout << " p = "; print_as_np_array(fout,beta); fout << ";\n";
fout << " p.shape = net.params['"<<i->caffe_layer_name()<<"'][1].data.shape;\n";
fout << " net.params['"<<i->caffe_layer_name()<<"'][1].data[:] = p;\n";
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
Added missing sig and prelu layer support.
2017-05-08 04:23:35 +08:00
else if (i->detail_name == "prelu")
{
const double param = i->params(0);
// main filter weights
fout << " tmp = net.params['"<<i->caffe_layer_name()<<"'][0].data.view();\n";
fout << " tmp.shape = 1;\n";
fout << " tmp[0] = "<<param<<";\n";
}
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
}
}
// ----------------------------------------------------------------------------------------
int main(int argc, char** argv) try
{
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
if (argc != 6)
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
{
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
cout << "To use this program, give it an xml file generated by dlib::net_to_xml() " << endl;
cout << "and then 4 numbers that indicate the input tensor size. It will convert " << endl;
cout << "the xml file into a python file that outputs a caffe model containing the dlib model." << endl;
cout << "For example, you might run this program like this: " << endl;
cout << " ./dtoc lenet.xml 1 1 28 28" << endl;
cout << "would convert the lenet.xml model into a caffe model with an input tensor of shape(1,1,28,28)" << endl;
cout << "where the shape values are (num samples in batch, num channels, num rows, num columns)." << endl;
Cleaned up how the output is output.
2017-05-08 03:49:19 +08:00
return 0;
}
Changed caffe converter to require the user to specify the input tensor size when the converter is run.
2017-05-23 07:06:55 +08:00
const long N = sa = argv[2];
const long K = sa = argv[3];
const long NR = sa = argv[4];
const long NC = sa = argv[5];
convert_dlib_xml_to_caffe_python_code(argv[1], N, K, NR, NC);
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
return 0;
}
catch(std::exception& e)
{
cout << "\n\n*************** ERROR CONVERTING TO CAFFE ***************\n" << e.what() << endl;
return 1;
}
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
class doc_handler : public document_handler
{
public:
std::vector<layer> layers;
bool seen_first_tag = false;
layer next_layer;
std::stack<string> current_tag;
long tag_id = -1;
virtual void start_document (
)
{
layers.clear();
seen_first_tag = false;
tag_id = -1;
}
virtual void end_document (
) { }
virtual void start_element (
const unsigned long line_number,
const std::string& name,
const dlib::attribute_list& atts
)
{
if (!seen_first_tag)
{
if (name != "net")
throw dlib::error("The top level XML tag must be a 'net' tag.");
seen_first_tag = true;
}
if (name == "layer")
{
next_layer = layer();
if (atts["type"] == "skip")
{
// Don't make a new layer, just apply the tag id to the previous layer
if (layers.size() == 0)
throw dlib::error("A skip layer was found as the first layer, but the first layer should be an input layer.");
layers.back().skip_id = sa = atts["id"];
// We intentionally leave next_layer empty so the end_element() callback
// don't add it as another layer when called.
}
else if (atts["type"] == "tag")
{
// Don't make a new layer, just remember the tag id so we can apply it on
// the next layer.
tag_id = sa = atts["id"];
// We intentionally leave next_layer empty so the end_element() callback
// don't add it as another layer when called.
}
else
{
next_layer.idx = sa = atts["idx"];
next_layer.type = atts["type"];
if (tag_id != -1)
{
next_layer.tag_id = tag_id;
tag_id = -1;
}
}
}
else if (current_tag.size() != 0 && current_tag.top() == "layer")
{
next_layer.detail_name = name;
// copy all the XML tag's attributes into the layer struct
atts.reset();
while (atts.move_next())
next_layer.attributes[atts.element().key()] = sa = atts.element().value();
}
current_tag.push(name);
}
virtual void end_element (
const unsigned long line_number,
const std::string& name
)
{
current_tag.pop();
if (name == "layer" && next_layer.type.size() != 0)
layers.push_back(next_layer);
}
virtual void characters (
const std::string& data
)
{
if (current_tag.size() == 0)
return;
if (comp_tags_with_params.count(current_tag.top()) != 0)
{
istringstream sin(data);
sin >> next_layer.params;
}
}
virtual void processing_instruction (
const unsigned long line_number,
const std::string& target,
const std::string& data
)
{
}
};
// ----------------------------------------------------------------------------------------
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
void compute_output_tensor_shapes(const matrix<long,4,1>& input_tensor_shape, std::vector<layer>& layers)
{
DLIB_CASSERT(layers.back().type == "input");
layers.back().output_tensor_shape = input_tensor_shape;
for (auto i = ++layers.rbegin(); i != layers.rend(); ++i)
{
const auto input_shape = find_input_layer(i).output_tensor_shape;
if (i->type == "comp")
{
if (i->detail_name == "fc" || i->detail_name == "fc_no_bias")
{
long num_outputs = i->attribute("num_outputs");
i->output_tensor_shape = {input_shape(0), num_outputs, 1, 1};
}
else if (i->detail_name == "con")
{
long num_filters = i->attribute("num_filters");
long filter_nc = i->attribute("nc");
long filter_nr = i->attribute("nr");
long stride_x = i->attribute("stride_x");
long stride_y = i->attribute("stride_y");
long padding_x = i->attribute("padding_x");
long padding_y = i->attribute("padding_y");
long nr = 1+(input_shape(2) + 2*padding_y - filter_nr)/stride_y;
long nc = 1+(input_shape(3) + 2*padding_x - filter_nc)/stride_x;
i->output_tensor_shape = {input_shape(0), num_filters, nr, nc};
}
else if (i->detail_name == "max_pool" || i->detail_name == "avg_pool")
{
long filter_nc = i->attribute("nc");
long filter_nr = i->attribute("nr");
long stride_x = i->attribute("stride_x");
long stride_y = i->attribute("stride_y");
long padding_x = i->attribute("padding_x");
long padding_y = i->attribute("padding_y");
long nr = 1+(input_shape(2) + 2*padding_y - filter_nr)/stride_y;
long nc = 1+(input_shape(3) + 2*padding_x - filter_nc)/stride_x;
i->output_tensor_shape = {input_shape(0), input_shape(1), nr, nc};
}
else if (i->detail_name == "add_prev")
{
auto aux_shape = find_layer(i, i->attribute("tag")).output_tensor_shape;
for (long j = 0; j < input_shape.size(); ++j)
i->output_tensor_shape(j) = std::max(input_shape(j), aux_shape(j));
}
else
{
i->output_tensor_shape = input_shape;
}
}
else
{
i->output_tensor_shape = input_shape;
}
}
}
// ----------------------------------------------------------------------------------------
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
std::vector<layer> parse_dlib_xml(
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
const matrix<long,4,1>& input_tensor_shape,
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
const string& xml_filename
)
{
doc_handler dh;
parse_xml(xml_filename, dh);
if (dh.layers.size() == 0)
throw dlib::error("No layers found in XML file!");
if (dh.layers.back().type != "input")
throw dlib::error("The network in the XML file is missing an input layer!");
Made the converter add zero padding layers when needed by Eltwise to replicate the behavior of dlib's add_prev layers.
2017-05-24 19:24:12 +08:00
compute_output_tensor_shapes(input_tensor_shape, dh.layers);
Added a simple tool to convert dlib model files, which are saved as XML files, to caffe models. Not all layers are supported yet, but it works on the net from the first dlib dnn example program. Will add the rest of the layers soon.
2017-05-07 03:02:38 +08:00
return dh.layers;
}
// ----------------------------------------------------------------------------------------
Reference in New Issue Copy Permalink