search

bl0 / PIC

Parametric Instance Classification for Unsupervised Visual Feature Learning, NeurIPS 2020
https://arxiv.org/abs/2006.14618v1
51 stars 4 forks source link

k = x.shape[1] #4

Open techmonsterwang opened 2 years ago

techmonsterwang commented 2 years ago

Hi! A line of code in PIC.py is: k = x.shape[1].

I wondor what does k means here? Maybe image channels? But why is each channel handled separately?

bl0 commented 2 years ago

To achieve better performance, our PIC adopts a multi-crop strategy, which is proposed in SwAV. In each iteration, one 160 x 160 crop and three 96 x 96 crops of an image are fed into the model. With similar memory and compute requirements, PIC could achieve better performance than the original PIC model.

techmonsterwang commented 2 years ago

Thanks a lot. I have understand k meaning. But why choose--num-crop 1 \ --num-crop2 3 \?

bl0 commented 2 years ago

No special reason, just practice choices.