Butch Landingin

Results 6 comments of Butch Landingin

Focusing on affine transforms (zoom, warp, rotate) + random resize crop GPU - they seem to cause much of the slowdown. Normalize and lighting (contrast and brightness) transforms dont seem...

Will monitoring [this issue filed with the pytorch-xla team](https://github.com/pytorch/xla/issues/2405) as resolving it requires an update to pytorch-xla itself.

Update: as of 2020/12/14, using updated Pytorch 1.7 XLA and latest fastai (2.1.8) and fastai_xla_extensions (0.0.4) packages, training with batch transforms is still slower than training without batch transforms. *...

Partially mitigated by this enhancement: https://github.com/butchland/fastai_xla_extensions/issues/11 Leaving it open awaiting final action for additional lowerings in Pytorch XLA to support batch transforms

The text_classifier_learner using AWD_LSTM as the model actually completes, but is very slow Training an AWD_LSTM is even slower for TPU compared to a CPU (and GPU of course) Notebooks...

Referencing [issue submitted to pytorch xla team for help in resolution](https://github.com/pytorch/xla/issues/2422). Also including `debug_run` logs submitted for the issue: * single_profile_forward - [log file](https://github.com/pytorch/xla/files/5070392/single_profile_forward.tar.gz) : [script](https://gist.github.com/butchland/d32578f8fca278a498db0ae231476765) * single_profile - [log...