joshvarty / cancerdetection Goto Github PK

30 epochs @ slice(1e-6,1e-2)) (Cancer Normalization)

10 epochs @ slice(1e-5,2e-2) (ImageNet normalization)

15 epochs @ slice(1e-6, 1e-2) (Imagenet normalization)

15 epochs @ slice(1e-6, 1e-2) (Updated cancer normalization)

15 epochs @ slice(1e-6, 1e-2) + 5 epochs @ (1e-7, 1e-3) (Updated cancer normalization)

25 epochs @ slice(1e-6, 1e-2) (Imagnet normalization) More transformations

Apparently I have been using the wrong metric. Let's fix that.

• Choose sensible transformations
• Set up training for various models
• Begin remote training

We could try using weight decay to reduce overfitting.

We only care about tumors in the 32x32 center of the image. Maybe we should try cropping to 48x48.

See: https://docs.fast.ai/basic_train.html#TTA

We need to create a proper validation set by removing sets of images that came from the same original slide picture. After creating the validation set we should train a few networks on it and compare the validation AUC score to what we get on the leaderboard.

When we increase validation score, does our LB usually improve?

It might help to have a custom head with an extra fully connected layer.

We should use:

• flip_vert
• increase rotation

It's possible normalization made our performance worse.