ONNXKnet.jl : Read ONNX graphs in Knet

ONNXKnet.jl is an ONNX backend for Knet as it provides model import functionalities for the Knet.jl machine learning framework. This is heavily inspired by ONNX.jl.

Loading ONNX serialized models:

>>> using Knet, ONNXKnet

>>> ONNXKnet.load_model("model.onnx") # produces two files: weights.bson and model.jl

>>> weights = ONNXKnet.load_weights("weights.bson")

>>> model = include("model.jl")

model is the corresponding Knet model, and can be called just like any other Knet model: model(ip), where ip is the input with appropriate dimensions.

To Do:

This package is still under development and quite a few things need to be done here:

1. Add support for other larger models : VGG, face detection, object detection, emotion ferplus and so on.

2. Look into ways of removing code redundacy : Discuss if it's a good idea to add ONNX.jl as a dependency, so that the model interface code can be shared.

Supported operators:

Currently supported operators include:

1. Reshape
2. Conv
3. Concat
4. Add
5. Mul
6. Relu
7. Maxpool
8. AveragePool
9. Matmul
10. GlobalAveragePool

Contributing

Contributors are always welcome. Discussion takes place on Julia Slack, either on the #Knet or #ONNX channels.