/a/ Reduces mem-consumption, which could be interesting if you run REINFORCE on for under-powered GPUs
/b/ Must give a speed-up on modern GPUs (eg Volta) (however, I didn't test that)
How Has This Been Tested?
/1/ Manually checked convergence on the channel (zipfian) game
/2/ Checked that it unlocks ~2x larger batch size on language_bottleneck/image_classification
Enables mixed-precision training when
--fp16
is specified.Description
Uses pytorch.cuda.amp primitives in Trainer.
Related Issue (if any)
https://github.com/facebookresearch/EGG/issues/165
Motivation and Context
/a/ Reduces mem-consumption, which could be interesting if you run REINFORCE on for under-powered GPUs /b/ Must give a speed-up on modern GPUs (eg Volta) (however, I didn't test that)
How Has This Been Tested?
/1/ Manually checked convergence on the channel (zipfian) game /2/ Checked that it unlocks ~2x larger batch size on language_bottleneck/image_classification