Inference results in weird values about imgclsmob HOT 6 CLOSED

Repro:
wrong https://github.com/artyompal/google_landmark_2019/blob/master/imagenet_classifier.py
right https://github.com/artyompal/google_landmark_2019/blob/master/imagenet_classifier2.py

from imgclsmob.

If I use pretrainedmodels.se_resnext50_32x4d instead, it also works. So I think there's a problem with models in this repo, specifically with the last FC layer. Probably no-one noticed it because nobody uses the last layer, it's usually all about finetuning.

from imgclsmob.

What if you try model = get_model('seresnext50_32x4d', pretrained=True).cuda()?

from imgclsmob.

If the advice does not help, please reopen the issue.

from imgclsmob.

Oh thanks, that was actually a pretty stupid bug.

from imgclsmob.

You're welcome:)

from imgclsmob.