butchland/fastai_xla_extensions

TODO

Closed this issue · 0 comments

tyoc213 commented
  • fit one cycle / fit
  • what happens with lr finder?
  • unfreeze model makes the gradients explode
  • batch transforms are slow :) WHY!
  • change default_device to check GPU, then TPU then set CPU.
    - [ ] add patch to LMLearner.load_pretrainedand fix method load_ignore_keys to work on TPUs
    - [ ] textlearner (language_model_learner) is very slow in running fit_one_cycle (havent finished 1 epoch)
  • test collab learner
  • test tabular learner