Camelyon16: pretrained embedders

[Bontempogianpaolo1]

Hi @binli123,
Given the data obtained as #39, I extracted the features using both model-v0 and model-v2. The differences between their performances on the downstream task are evident. Here is the AUC:

Looking at #12 you say they differ for batch size and training time. Could you be more specific?

[binli123]

In my case both are converging, v2 converges significantly slower at around 100 epoch. If you load "init.pth" it would be much faster.

[binli123]

I incorporated a simple weights initialization method which helps stabilize the training in the latest commit.