Create your own GitHub profile
Sign up for your own profile on GitHub, the best place to host code, manage projects, and build software alongside 40 million developers.
Sign up
Pro
Popular repositories
trax-robot doesn’t have any public repositories yet.
319 contributions in the last year
Contribution activity
February 2020
- Remove Python 2 compatibility code from layers package.
- Fast inference helper for autoregressive models.
- 3rd batch of new tf-numpy ops, and a revamp of the dtype inference/promotion system.
- Continue design/implementation of the new supervised training loop:
- Small changes carved off from larger ongoing work:
- Fix issue #301 (thanks matwilso!):
- Make awr_trainer_test smaller/faster - hopefully.
- Change from py_test to jax_test, in order to run on accelerators.
- Speed up mnist_test by training for 1000 steps rather than 2000.
- Fix typo in config
- AWR implementation in Trax
- Update Reformer configs for WMT en-de
- Remove TPU warning now that bug is fixed
- Separate ff_dropout hparam to control Reformer relu dropout
- Tighten down on the possible outputs of shapes.signature.
- Serialization refactoring part two: implement a uniform interface for raw and serialized policies. Not plugged into the RL pipelines yet.
- Start making ValueError messages more consistent.
- Copy play_env_problem_with_policy from tensor2tensor to Trax for future tweaking.
- Fast inference support for Reformer encoder-decoder model
- Caching layer to use in Transformer for decoding.
- Fast inference for new efficient SelfAttention class
- Rename _forward_internal to pure_fn and make public to ease purely functional model use.
- Heart of design for simplified set of classes for supervised training. Main parts:
- Small corrections: L2 loss, add Conv1d, allow to checkpoint highest and lowest metric.
- Support dropout and rngs in new efficient attention code
- Some pull requests not shown.