NVIDIA-NeMo / NVIDIA-NeMo/RL

Cosine Decay Not Properly Enforced

Open
#1,532 3 comments 0 reactions 1 assignee Claimed by @ashors1 View on GitHub
bug i-nvpr
Dominant language
Python
Stars
2k
Forks
561
Avg merge
4d 5h
Merged PRs (30d)
145

Description

When passing the following hyper-parameters , one expects to see a cosine decay profile:

```
scheduler:
start_weight_decay: ${policy.megatron_cfg.optimizer.weight_decay}
end_weight_decay: ${policy.megatron_cfg.optimizer.weight_decay}
weight_decay_incr_style: "cosine"
lr_decay_style: "cosine"
lr_warmup_iters: 25
lr_warmup_init: 1e-7
```

However, although warmup is properly enforced, learning rate seems to remain constant. Please see the figure below.

Image

We appreciate if there is a solution for actual enforcement of a cosine decay profile.

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.