luatorch-onnx-convert

Overview

This repository provides an utility to extract model(s) from a serialized lua-torch model (.t7 file) and convert it/them to onnx format.

Requirements

• You need first to install lua torch as well as the library used in the model that are necessary to load the model from torch (for instance, nn, nngraph, onmt or your own libraries).
• Install protobuf lua library from this repository https://github.com/jsenellart/protobuf-lua (this version is mandatory since it fixes some issues with the original implementation).

You can regenerate the lua interpreter of the onnx proto file by doing:

protoc --lua_out=. onnx.proto

This will generate an updated version of onnx_pb.lua

Extracting model

$ th convert.lua -t7 test/data/model1.t7 -require nn  -force

convert.lua goes through the serialized torch object and looks for supported modules or models. Each of them is converted into test/data/model1.onnxdir directory with the corresponding name:

model.linear-bias.onnx
model.linear-nobias.onnx

The option -require nn indicates which library are necessary to deserialize the model. GPU model needs cunn installation to be read.

Convertors

To perform conversion to onnx, each of the used module must be described with onnx operators. New convertors for specific library can be added in convertors. Their name must match onnx_class where class is the torch classname of the modules.