The performance after linear evaluation on ImageNet
yxgeee opened this issue · comments
Hi,
Thanks for your nice implementation! I am wondering about the performance after training a linear classifier on ImageNet. I noticed a result in your blog (57.7%), but there seems no explanation about the dataset and detailed evaluation method on it.
Looking forward to your reply! Thanks.
Hi @yxgeee , thanks for the note!
The 57.7% result in the table is linear evaluation accuracy on the STL10 test set. Embeddings from half of the test set are used to fit an sklearn
linear classifier, and the other half is used for validation. We find that this tracks well with more robust linear evaluation protocols (training a linear layer with SGD). I've added a footnote to make this more clear. Thanks for the comment!