Releases: hfxunlp/transformer
Pre-Release v0.1.6
Support Transparent Attention;
Accelerate the decoding of test set through sorting.
Pre-Release v0.1.5
Add support for original paper described transformer.
Pre-Release v0.1.4
fix potential break with multi-gpu decoding with 1 as batch size
Pre-Release v0.1.3
fixes
Pre-Release v0.1.2
A ranking tool is added to help data selection;
Averaging of probabilities rather than log probabilities in Ensemble;
Extracting and preventing generation of source side only words is supported for shared vocabulary (with tools/fbindexes.py);
A tool (mkcy.py) is added to convert python modules into C libraries, but no additional code level improvement for Cython;
Several typos are fixed with above tool.
Pre-Release v0.1.1
Bias in MultiHeadAttn is removed in this release;
Only parameters of the trained model is saved, rather than state dict, trained model with v0.1.0 is not loadable as fine_tune_m without additional convertion;
Try to support RNMT, but recurrent is slow due to less efficient parallelization.