dvc has our raw/cache folder. in the cache folder there is a bunch of throwaway files like cache-abc123.arrow, which are created with new name after each training run. Created a new commit and pushed the dvc files. changed dvcignore
max did some stuff with save_to_model_gs and shell scripts for docker
removed the wandb experiments that are commited under tests/ by max. New experiments should be stored under ./cache/experiments? (no need to track them with git / dvc)!
a couple of changes:
pytorch lightning
dataloader classrefactoerd the model into
pytorch lightning
pl.lightningmodule classChanged the tests to work with hydra. created a config dict in the root dir. I think it should I used this: https://github.com/lucmos/nn-template and this https://pytorch-lightning.readthedocs.io/en/latest/notebooks/lightning_examples/text-transformers.html?highlight=self.metric#Transformer-LightningModule template . both had a bit better quality then our blog post.
dvc has our raw/cache folder. in the cache folder there is a bunch of throwaway files like cache-abc123.arrow, which are created with new name after each training run. Created a new commit and pushed the dvc files. changed dvcignore
max did some stuff with save_to_model_gs and shell scripts for docker
removed the wandb experiments that are commited under tests/ by max. New experiments should be stored under ./cache/experiments? (no need to track them with git / dvc)!