Repo goes through https://pytorch.org/tutorials/intermediate/char_rnn_classification_tutorial for PyTorch, and then replicates it in PyTorch XLA.
While doing this implementation, I decided to simplify the model, and just do a deep neural network. I did this by substituting the RNN by a sequential neural network, and changing so each name is now represented as a frequency map for each letter. This is not the best translation, but will make due for a quick example.
For the TPU example, the notebooks was translated to a main executable for TPU best translation. Graphs are added to the directory for reference.
This has caused the model to be a lot more volatile between runs. I am doing this case as a basic example, and I am not diving deeper, but the behavior I saw definitely suggests overfitting.
Data pulled from https://download.pytorch.org/tutorial/data.zip
To do this migration, I used https://github.com/pytorch/xla/tree/8b45e5993bc55dec0303ca7c3a84f59326cab181?tab=readme-ov-file#getting-started. This means that we are using the multiprocessing instructions. This has required some minor modifications to avoid issues with multiple files being open.
Furthermore, rather than a jupiter notebook, I am relying in main.py.