Skip to content
This repository has been archived by the owner on Aug 3, 2021. It is now read-only.

OpenSeq2Seq v0.3

Pre-release
Pre-release
Compare
Choose a tag to compare
@Kipok Kipok released this 31 May 17:34
· 1217 commits to master since this release
v0.3
4f054b1
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Verified
Learn about vigilant mode.

New models:

  • Added ResNet model and ImageNet data layer.
  • Improved DeepSpeech-2 models and reached 4.59% WER.
  • Added Transformers model.

New features:

  • Implemented evaluation in Horovod mode.
  • Added mixed precision support for Horovod mode.
  • Fixed evaluation in multi-GPU mode.
  • All string/numerical config parameters can now be rewritten from command line (nested dicts are separated with "/").
  • Moved start_experiment.sh functionality to run.py (--enable_logs parameter). Additionally it now logs exact command line arguments used to invoke the script.
  • Added new benchmarking functionality: now models can also report number of objects (e..g tokens or images) per second.
  • Added more summaries/parameters for different models.

API changes:

  • Replaced Seq2Seq class with EncoderDecoderModel to support arbitrary models that can be expressed in encoder-decoder-loss paradigm.
  • Changed data layer API to only work with tf.data (dropped placeholders support).
  • Hid Horovod/non Horovod differences from users (no need to take care about that when creating new models / data layers).

Other changes:

  • Improved unit tests and documentation.
Assets 2
Loading