fistyee / p-diff Goto Github PK

I have found your paper presented the result of "Deep Self-Learning From Noisy Labels"/INCV, if your re-implement its codes??? If possible, could you share me the source code of "Deep Self-Learning From Noisy Labels"/INCV. Thanks.

Hi @fistyee ,

thank you for the very interesting paper.

I stumbled upon one point I didn't quite get and could also not figure out from your implementation, so maybe you can help me out.

When estimating the noise, we are using zeta which is the expected value of |delta| in the distribution.

I believe in your code, this is named train_sat here: https://github.com/fistyee/P-DIFF/blob/master/layer/p_diff_layer.cpp#L199

In the paper, you mention that "once zeta is larger than a threshold [0.9], all samples with delta < 0 are regarded as noisy samples to estimate the noise rate tau."

When I think about it, it seems the noise rate is then always pointing to the bucket referring to delta = 0, so the bucket id will be 100 (when using 200 buckets), so we would still use the threshold delta > 0.

I tried to consult your code to figure this out more in detail, but it seems that you skip the computation when zeta > 0.9:
if (T_ <= Tk_ || train_sat < thred_train_sat_) https://github.com/fistyee/P-DIFF/blob/master/layer/p_diff_layer.cpp#L202 in which case the noise ratio would again be calculated to be the bucket referring to delta = 0?

I probably missed sth, so it would be great if you can point me in the right direction. Thank you!