dimensionality-driven-learning
dimensionality-driven-learning copied to clipboard
xingjunm
pytorch version
I reworked the code into a new pytorch version and reproduced the results.
The following two figures show the results of LID and train/test accuracy throughout training with clean(first figure) and noisy(second figure) labels:
I used six different training criteria(cross-entropy,forward,backward,boot-hard,boot-soft,D2L) to train the model.
The figures below show the trend of test accuracy and subspace dimensionality on CIFAR-10 with 40% noisy labels:
The figures below show the trend of test accuracy and subspace dimensionality on CIFAR-10 with 60% noisy labels:
