fine-tuning scan phase on custom dataset

[noamberg]

Hello,
Thanks for sharing your clean and well-explained implementation.

Is it possible to fine-tune scan/self-labeling phases (or even pretext) based on trained simclr_cifar-10.pth (or other trained weights) you have provided here?
Technically it should work, but will the learned features from different datasets you published can help in other animal-based custom dataset?

When training from zero on my custom dataset on simclr pretext, knn evaluations are nowhere near yours.
It makes me wonder whether it s possible to transfer some knowledge from pretrained weights of different datasets into my experiments.

Thanks