Skip to content
This repository has been archived by the owner on Sep 8, 2021. It is now read-only.

validation top-1 accuracy for ResNeXt-101 32x{4,16} models #3

Open
a-maci opened this issue Dec 23, 2019 · 1 comment
Open

validation top-1 accuracy for ResNeXt-101 32x{4,16} models #3

a-maci opened this issue Dec 23, 2019 · 1 comment

Comments

@a-maci
Copy link

a-maci commented Dec 23, 2019

I am not getting top-1 accuracies mentioned in the repo for ResNext101 32x4 and 32x16 models.
Here is what I am getting vs. what the repo mentions:

RX101_32x4d: 83.223 (mine) vs 83.4 (repo)
RX101_32x16d: 83.348 (mine) vs 84.8 (repo)
@rtaori
Copy link

rtaori commented Jan 21, 2020

+1 have encountered the same issue. the rest of the models I can get to within 0.1-0.2% of the advertised accuracy, just the resnext101_32x16d_swsl model has a large discrepancy

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants