Skip to content

Catalyst 20.09
Compare
Choose a tag to compare
@Scitator Scitator released this 20 Dec 10:10
· 387 commits to master since this release
v20.09
2275ea9
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

[20.09] - 2020-09-07

Added

  • MovieLens dataset loader (#903)
  • force and bert-level keywords to catalyst-data text2embedding (#917)
  • OptunaCallback to catalyst.contrib (#915)
  • DynamicQuantizationCallback and catalyst-dl quantize script for fast quantization of your model (#890)
  • Multi-scheduler support for multi-optimizer case (#923)
  • Native mixed-precision training support (#740)
  • OptiomizerCallback - flag use_fast_zero_grad for faster (and hacky) version of optimizer.zero_grad() (#927)
  • IOptiomizerCallback, ISchedulerCallback, ICheckpointCallback, ILoggerCallback as core abstractions for Callbacks (#933)
  • flag USE_AMP for PyTorch AMP usage (#933)

Changed

  • Pruning moved to catalyst.dl (#933)
  • default USE_APEX changed to 0 (#933)

Removed

Fixed

  • autoresume option for Config API (#907)
  • a few issues with TF projector (#917)
  • batch sampler speed issue (#921)
  • add apex key-value optimizer support (#924)
  • runtime warning for PyTorch 1.6 (920)
  • Apex synbn usage (920)
  • Catalyst dependency on system git (922)
Assets 2
Loading