search

facebookresearch / semi-supervised-ImageNet1K-models

Semi-supervised ImageNet1K models
Other
241 stars 21 forks source link

Cannot reproduce validation results for SWSL ResNeXt-101 32x16d model #1

Open rwightman opened 5 years ago

rwightman commented 5 years ago

I ran through the models and was able to reproduce the top-1/top-5 scores for all the models but the largest SWSL one (to within typical variations for specific PyTorch version, PIL, etc).

For SWSL ResNeXt-101 32x16d I'm getting 83.338 top-1 and 96.852 top-5.

For comparison, on my setup, the SWSL 8d is 84.294, 4d is 83.234 top-1.

Edit: Tested on two machines, both Pytorch 1.2, Pillow 6.1

fbzekiyalniz commented 5 years ago

Hi Ross, thank you for reporting the issue and waiting patiently. We were able to find out the issue (incorrectly saved model). We are going to update the ResNext-101-32x16 model soon.

rwightman commented 5 years ago

@fbzekiyalniz thanks for the update