e-lab / th2caffe

A torch-nn to caffe converter for specific layers.

#th2caffe

A torch-nn to caffe converter for specific layers.

Contents

• Installation
• Supported layers
• How to run
• Output

Installation

th2caffe requires the following packages.

caffe: To solve the discrepancy between the size of nn.SpatialMaxPooling and caffe:Pooling(MAX), some caffe codes are modified. This special caffe version can be installed either by overwriting some source codes in the original caffe, or by installing a folk version of caffe.

For overwriting files, please run

cp caffe/caffe.proto location/to/caffe/src/caffe/proto/caffe.proto
cp caffe/pooling_layer.cpp location/to/caffe/src/caffe/layers/pooling_layer.cpp
cp caffe/vision_layers.hpp location/to/caffe/include/caffe/vision_layers.hpp
make clean
make all
make test
make runtest

See BVLC/caffe's GitHub for installation. Install pycaffe also to run python scripts.

Please note:

hdf5: Install both versions of python and torch. for python, run

pip install h5py

for torch, install deepmind's version following the instructions.

Supported layers

Currently th2caffe supports the following layers for conversions.

nn.Linear                          (caffe: InnerProduct)
nn.SpatialConvolutionMM            (caffe: Convolution)
nn.SpatialMaxPooling (floor/ceil)  (caffe: Pooling, type MAX)
nn.ReLU                            (caffe: ReLU)
nn.View                            (caffe: Reshape)
nn.Dropout                         (caffe: Dropout)
nn.SoftMax                         (caffe: Softmax)

How to run

cd th2caffe
th th2caffe.lua --nf netFile \
                --name nnName \
                --c numChannels \
                --w width \
                --h height \
                --loc location \
                --caffe caffeloc

with arguments:

--nf netFile       Path to the .net file saving all information for a torch neural network
--name nnName      Name that is specified to the caffe neural network
--c numChannels    Number of input channels
--w width          Width of input image
--h height         Height of input image
--loc location     Location to save the output files (will be created if it doesn't exist)
--caffe caffeloc   Location to the caffe source

Output

The output files include:

location/architecture/deploy.prototxt   Prototxt file saving the architecture
location/params/params.h5               .h5 files saving layer parameters
location/params/params.caffemodel       .caffemodel file that can be loaded by caffe

License

MIT