Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[sched] polynomial_decay_schedule use default power=1.0 #6473

Merged
merged 1 commit into from
Aug 17, 2020

Conversation

stas00
Copy link
Contributor

@stas00 stas00 commented Aug 14, 2020

As discussed in #6361 we weren't sure why fairseq's polynomial_decay_schedule power default was 1.0, and decided to go with 2.0 as the latter does something polynomial.

I got the devs at fairseq to answer this question: facebookresearch/fairseq#2466 (comment)

myleott wrote:
This is based on the original BERT code, which implemented a linear decay via a polynomial schedule with power=1.0: https://github.com/google-research/bert/blob/f39e881b169b9d53bea03d2d341b31707a6c052b/optimization.py#L37

So, perhaps we do the same or we don't.

If we don't - then the doc needs to be fixed that the default is power=2.0 as currently it says 1.0 - my mistake. If we do (this PR), then the doc is already correct.

Thanks.

@JetRunner JetRunner requested a review from sgugger August 17, 2020 03:15
Copy link
Collaborator

@sgugger sgugger left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have no strong opinions on the default, so let's use the same as FAIR.

@sgugger sgugger merged commit 39c3b1d into huggingface:master Aug 17, 2020
@stas00 stas00 deleted the sched-poly-power branch August 17, 2020 15:01
Zigur pushed a commit to Zigur/transformers that referenced this pull request Oct 26, 2020
fabiocapsouza pushed a commit to fabiocapsouza/transformers that referenced this pull request Nov 15, 2020
fabiocapsouza added a commit to fabiocapsouza/transformers that referenced this pull request Nov 15, 2020
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

2 participants