models returned are generic functions, not Flux models

[amckay1]

Perhaps I'm missing something big, but the models I'm loading are generic functions, not Flux models, with no trainable params. I tried this on resnet and shufflenet since someone had mentioned in issue #25 that they were able to load that model into Flux but it failed later, but I wasn't able to even get that far.

What am I doing wrong? Running Julia v1.1.1, Flux v0.8.3, ONNX v0.1.1, and following the README:

using Flux, ONNX
ONNX.load_model("model.onnx")
weights = ONNX.load_weights("weights.bson")
model = include("model.jl")

Thank you for your help!

[ayush-1506]

Sorry for not getting back at the issues raised over the past week on time. Yes, the model from your above code is just a function, not specifically a Flux model. (params(model) will return Params([])). Hence, at the moment the ONNX models can be used only for inference. However, some of the ONNX models have been ported to Metalhead.jl, which maybe be what you're looking for.

[amckay1]

Thanks for the clarification Ayush! Just needed a sanity check, and maybe we should adjust the README.md since by my reading it implied a returned Flux model. I also wonder if with the newest Zygote.jl advances in AD we could make these models trainable, but certainly not a priority :)