Closed Tverous closed 4 years ago
From the original paper
Learning to Learn from Noisy Labeled Data
The algorithm states that the meta loss should be calculated after every consistency loss is accumulated:
While in the original code, the consistency loss does not accumulate before divide the number of mini-batch
This issue was mentioned in #7 before, but if the consistency_loss gradient is accumulated across M synthetic data
Then it should not be divided by M( args.num_fast in the source) every time after generating noisy labels
This pull request should fix the problems that appear in #8 and #9 as well.
The GPU memory usage problems should be resolved after decreasing the number in args.num_fast
Thanks for the update! I have merged the pull request.
From the original paper
Learning to Learn from Noisy Labeled Data
The algorithm states that the meta loss should be calculated after every consistency loss is accumulated:
While in the original code, the consistency loss does not accumulate before divide the number of mini-batch
This issue was mentioned in #7 before, but if the consistency_loss gradient is accumulated across M synthetic data
Then it should not be divided by M( args.num_fast in the source) every time after generating noisy labels
This pull request should fix the problems that appear in #8 and #9 as well.
The GPU memory usage problems should be resolved after decreasing the number in args.num_fast