Tags: TaoLv/sockeye
Tags
Update to MXNet 1.3.0 (awslabs#534) Update to MXNet 1.3.0 to make use of the following new features: - use of MXNet unravel_index & topk - use of MXNet logical operators - topk is now a HybridBlock. - Inference HybridBlocks use static_alloc and static_shape
No longer relying on external translation id. (awslabs#508) * No longer relying on external translation id. * Making mypy happy. * removing chunk_id from input tests. * fix branch without source eos * update docstring * Sentence ids can now be strings * addressed comment
Ensure recent numpy version to avoid MKL conflicts. (awslabs#494) * Making sure to use numpy>=1.14 to avoid MKL conflicts. * changelog * Update CHANGELOG.md * Update __init__.py * Update CHANGELOG.md * Update __init__.py
Make rouge metrics available in sockeye.evaluate CLI (awslabs#471) * Make rouge metrics available in sockeye.evaluate CLI * Add chrf to optimizable metrics * fix * compute rouge in system tests * fix
Fixing the pyyaml version. (awslabs#458) The new pyyaml version is introducing some issues, which we should fix. For now let's depend on the old version to unblock Travis.
Fix for transformer-with-conv-embed encoder. (awslabs#434) * Fix for transformer-with-conv-embed encoder. * changelog * version fix
Refactor beam search to use gluon hybrid blocks (awslabs#425) This refactors beam search to group fixed-size operations in beam search into cached ops through Gluon HybridBlocks. My testing showed ~3% speed improvement. Not much, but consistent. Another change that is included here is to not use columns from `sequences` to pass into the decoder step module, but use `best_word_indices` from the previous iteration. NDArray indexing seems expensive and ideally we should aim for avoiding all indexing ops in an iteration.
Allow specifying different transformer parameters for encoder and dec… …oder (awslabs#419)
PreviousNext