search

udibr / noisy_labels

TRAINING DEEP NEURAL-NETWORKS USING A NOISE ADAPTATION LAYER
118 stars 38 forks source link

Implementation of Reed baseline does not use predictions? #1

Closed ghost closed 7 years ago

ghost commented 7 years ago

Hi, thanks for sharing your code! Much appreciated.

I'm curious about your implementation of the soft/hard loss by Reed et al. It seems that the fix_output function just copies the label when used for training the model (see https://github.com/udibr/noisy_labels/blob/master/jacob-reed.py#L780). However, the whole idea of the paper by Reed et al. is to use the model's predictions as an extra 'bootstrapped' label, causing the model to train more consistently in the context of noisy labels.

ghost commented 7 years ago

Nevermind, I see that in your loss function you never actually use the label.