Jonathan Bratt
Jonathan Bratt
Paused; waiting on custom Keras layer API to be standardized. https://github.com/rstudio/keras/pull/942
Thanks for the note! My best guess is that it's a python version issue and/or a TensorFlow version issue. Unfortunately, RBERT has been in an update limbo between TF versions...
Known bug: - NULL batch_size in run_classifier.R: input_fn_builder()
One possible approach is to write some code to convert PyTorch models into TensorFlow checkpoints, at which point it should be possible to use existing code to load/use. I don't...
Some possibly helpful links: https://forums.fast.ai/t/converting-a-pytorch-model-to-tensorflow-or-keras-for-production/14016/30 https://github.com/onnx/tutorials https://github.com/onnx/tutorials/blob/master/tutorials/PytorchTensorflowMnist.ipynb
Related: SciBERT checkpoints have been released in tensorflow format, so those are already available in RBERT. https://github.com/allenai/scibert
Thanks for the note! I don't think we've tested the multilingual models specifically in RBERT. What's the nature of the failure you encountered? Were you able to get any of...
Possibly! We're working right now on using tensorflow hub (via tfhub package) to load the models, but are open to additional/other ways of proceeding. I haven't had much time lately...
There appear to still be some issues with this function that appear when the q,k,v tensors are not all the same (so the tests miss it). For example: ``` library(torch)...
I think I found the problem. Will submit PR shortly.