Closed ayush1999 closed 6 years ago
Thanks Ayush, I love this.
The only thing I'd like to think about is where to store the weights. Should we store it in this repo? A separate repo like the builders? Or a single repo for all Flux pre-trained weights -- maybe the model zoo project itself?
Also, wherever we store it, I'd like to use the BinaryBuilder infrastructure to download it to the client.
cc: @MikeInnes for thoughts?
RE build, in other projects we've just attached weights to a Github release and downloaded it from there.
@ayush1999 let's get Keras.jl to emit Julia code for the model, like ONNX.jl, and avoid the Keras dependency.
@aviks Any idea why the build might be failing?
The tests are failing only because TextAnalysis master needs Languages master. I can handle that locally, no worries.
The one change that might be useful is if the hd5 file can be converted to BSON. The we can remove the HDF5 dependency, and add a BSON dependency. Is that possible?
Ayush, thanks for this, I'll take it from here. Really appreciate your effort in getting this done.
Closed in favour of #79
The model was trained in Keras and loaded into Flux using Keras.jl. The current usage is:
The binaries will be downloaded if not found.