Jong Wook Kim

Hi, please try [this one](https://drive.google.com/file/d/1Mj2Em07Lvl3mvDQCCxOYHjPiB-S0WGT1/view?usp=sharing), trained for 500,000 iterations on the MAESTRO dataset. I haven't touched the model in a while, but `torch.load('model-500000.pt’)` should be able to load the PyTorch...

Yeah! I'll need some housekeeping to make the checkpoint work cross-version. PR is welcome! Thanks :D

Hello, try [this checkpoint](https://drive.google.com/file/d/1Mj2Em07Lvl3mvDQCCxOYHjPiB-S0WGT1/view?usp=sharing) which was trained for 500,000 iterations on the MAESTRO dataset. I haven't tried fine-tuning on these models, but it should be theoretically possible. I expect that...

I totally agree that the IDs should be able to represent Any type or Strings at least, but the current implementation chose Int because of a few technical reasons, but...

Hi, it's probably a torch version problem. Can you try with 1.2.0 and see if that still happens?

I don't know about `export.py`, but can you try using `map_location="cpu"` (or `map_location=DEFAULT_DEVICE`) wherever you're loading the model with `torch.load`? Basically the input and the model should be on the...

In an ablation study (that we didn't publish) we found that RoBERTa fine-tunes better than BERT or GPT-2 itself. We expect ELECTRA should work as well.

Can you paste the output of the following bash script - to check your system information? ``` curl https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py | python - ``` I suspect two possibilities: - Your NCCL...

I still suspect the NCCL install is the culprit; realized that that script doesn't check for NCCL version.. which can be checked `torch.cuda.nccl.version().`. Can you run the experiments fine with...

thanks! #2451 fixed this