How hard is it to rebuild a tensorflow model in julia?

It has been a pain for me to get tensorflow working on my arm MacBook to run a recently published denoising model.

Id love to port the model Julia because it would open up some options and make it easier to share with other Mac users at my institute. I also suspect there will be useful data in the derivatives themselves.

I’ve fantasized about porting several tensorflow models in the literature. My instinct is the investment is too high. Just go through conda hell, however painful. But I wanted to ask.

Has anyone reproduced something like a published deep ushaped resnet model in flux.jl or lux.jl? How hard is it, what sort of challenges would I expect? The architecture is known and documented. IMC-Denoise: a content aware denoising pipeline to enhance Imaging Mass Cytometry | Nature Communications .

Their code is here GitHub - PENGLU-WashU/IMC_Denoise: IMC_Denoise: a software package to enhance Imaging Mass Cytometry - Nature Communications .

I am not actively using the neural net frameworks in Julia at the moment, but you can take a look at ONNX.jl, which is an implementation of the ONNX standard for interoperability across languages:


I do all my dnn experiments inJulia. In my opinion, the hard think is to make the pipelines and all the rest, while models are usually easy. It is just manual work. I like to write things myself to understand them.

1 Like