raphaelschwinger
commented
1 month ago @nhaH-luaP
Hey Paul,
On our hardware 1500 worked well and fast, but I guess 300 is a more sensible and compatible choice, thanks for pointing this out!
Greetings
Raphael
https://github.com/DBD-research-group/BirdSet/blob/23381a561d37a2fc446f6358db53230b54310bcc/birdset/datamodule/birdset_datamodule.py#L153-L162
With one-hot encoding of the labels, the batch_size of 1500 seems to be unnecessarily large, especially when compared to mapping the data a few lines above, where it is only 300. I had several runs on a cluster stuck without any error messages at the one-hot encoding step without knowing exactly why, until I manually reduced the BatchSize to 300. Are there any reasons for this high batch_size here? Otherwise i would recommend reducing it to 300.
Greetings, Paul