Memory requirement for training on the conll-2012 corpus

[thomwolf]

Hi, I am trying to train your model on a AWS p2x instance (with a 12 Go K80 GPU) on the Conll-2012 corpus (2802 documents in the training dataset). The training eats all (RAM) memory (64 Go) in less than 30 % of the first epoch and gets killed before finishing it.

I was wondering on what type of machine you trained it ? Is 64 Go of RAM too small for training on the conll corpus ?

[julien-c]

+1

[clarkkev]

Huh, that strange. I trained the model on a 128G machine, but I don't think it should use more than 10G RAM. Do you know what's taking up all the memory? What is the size of the data/features directory created during preprocessing?