Drop-in replacement of torch.optim.AdamW.
- Doesn't support
foreach,fusedargument, as the optimizer is already fused - Doesn't support
amsgrad,maximize,capturable,differentiableargument yet
pip install adam-atan2from adam_atan2 import AdamATan2
# All supported arguments are listed below
optim = AdamATan2(model.parameters(),
lr=1e-3,
weight_decay=0.1,
betas=(0.9, 0.95)
)We tested the consistency against reference AdamW-atan2 PyTorch implementation. To run tests, clone this repository, run pytest:
pip install -e .
pytest