UNICON-Noisy-Label
UNICON-Noisy-Label copied to clipboard
About calculate noise rate by using clean eval-dataset
Thank you very much for your interesting work. I have a question about this line: in noise learning, is it a normal way to use a clean eval-set to calculate the noise rate (or something similar)? (If you have a test set to do the final test) I'm just new to this field and don't know the rules very well. Looking forward to your answer.
Add a little bit of my understanding: assuming a manually set ratio, the trainer can adjust the hyperparameter using the result of eval. Your work just automates the process. Is this the correct understanding?
To your first question, the noise rate is calculated from the training set. We use a clean eval-set for calculating the validation accuracy.
To your second question, yes, we do not adjust the hyperparameters for different noise rates. Same HP for all noise rates.