Awesome

Code to test AdaBelief in Transformer on ISWLT14DE-EN

install adableief

pip install adabelief-pytorch

set up fairseq

pip install --editable .

prepare data

sh prepare-iwslt14.sh

run code

sh config/adabelief.sh

Results

The BLEU score on my local machine (PyTorch 1.1, CUDA 9.0) is roughly: AdamW: 35.60 RAdam: 35.51 AdaBelief: 35.85 The result could vary with rnadomness, however they are all above 35.

Incompatibility between fairseq implementation here and new version PyTorch

When I tested AdaBelief in PyTorch 1.4 and PyTorch 1.6, the BLEU score is always below 30. Furthremore, the gradient norm in PyTorch 1.1 is always below 1.0, while with higher version PyTorch the grad explodes to 2 or more.

This seems to be a problem of the version incompatibility between fairseq here (<=0.8) and PyTorch.<br> The code here works fine with PyTorch 1.1.<br> When using PyTorch 1.6, AdaBelief (same code as here) works fine with latest fairseq implementation.<br> Code for transformer to work with PyTorch 1.6 is at https://github.com/juntang-zhuang/fairseq-adabelief