SamplePairing

[prakashjayy]

Paper

• image A with label x (after augmentation) and image B with label y (after augmentations) are simply averaged and fed into the network as image C with label x.
• The network is initially trained for few epochs (1 for imagenet and 100 for other datasets) without SamplePairing. Then trained for 8 epochs with SamplePairing and 2 epochs without SamplePairing (for imagenet 3L images and 2L images). At the end, the network is fine tuned without SamplePairing.
• we can see large scale fluctuations in the loss whenever SamplePairing is enabled and this is normal. This acts as a strong regularization for the network.