This is an example of a clean, reproducible, and boilerplate-free deep learning project that was developed as part of the Groomata deep learning course.
This project is organized using pytorch-lightning
, and all configurations and artifacts can be uploaded to wandb
without any compromise. You can see an example wandb workspace here. All configurations are programmatically generated and maintained by hydra
and hydra-zen
.
docker run \
--gpus=all \
--ipc=host \
--volume=/path/to/volume:/vision/.cache \
--env-file=/path/to/.env \
--tty \
groomata/vision \
# Override any configurations you want
optimizer.lr=0.0001 \
datamodule.dataloader.batch_size=64 \
trainer.max_epochs=100 \
trainer.gradient_clip_algorithm="norm" \
trainer.gradient_clip_val=1.0