Open nps1ngh opened 1 year ago
Thanks for the report @nps1ngh . Good catch!
For all torchivsion models, the correct reference for accuracies should be this table: https://pytorch.org/vision/main/models.html#table-of-all-available-classification-weights (or the specific model pages like the one you gave above, which has the same info). Unfortunately, the ones reported on the torchhub website may go out of date. I opened https://github.com/pytorch/hub/issues/318 to try to think of solutions / mitigations.
Meanwhile, it's probably worth updating the VGG table with the latest correct values. LMK if you would you like to open a PR for that. Thank you!
Sure, I'll open one!
The
hub
documentation here: https://pytorch.org/hub/pytorch_vision_vgg/ reports thevgg11_bn
top-1 error to be26.70
. That is, a top-1 accuracy of73.30
. This is almost as good as theresnet34
reported here: https://pytorch.org/vision/main/models/generated/torchvision.models.resnet34.htmlBut the main problem is that
torchvision
reports a top-1 accuracy of70.37
for the it here: https://pytorch.org/vision/main/models/generated/torchvision.models.vgg11_bn.htmlI guess the latter reported accuracy is correct (
70.37
) and the one reported here byhub
is incorrect?Or am I missing something? Is this possibly related to https://github.com/pytorch/vision/issues/223? (It's a really old issue though.)
Thanks!