Currently I am using TensorBoardLogger for all my needs and it's perfect, but i do not like how it handles checkpoint naming. I'd prefer to be able to specify the filename and the folder where to put the checkpoint manually, how should i do that?
Yes, it is possible thanks to the ModelCheckPoint callback:
from pytorch_lightning.callbacks import ModelCheckpoint
checkpoint_callback = ModelCheckpoint(
dirpath="best_models",
filename='{epoch}-{val_loss:.2f}-{other_metric:.2f}'
)
trainer = Trainer(callbacks=[checkpoint_callback])
will create a checkpoint in the directory best_models/epoch=2-val_loss=0.02-other_metric=0.03.ckpt for example