bbuing9/darp

Code for the paper "Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning" (NeurIPS 20)

darp's Issues

inconsistent results of ReMixMatch

--

Code Release

I read the interesting work, DARP.
Thank you for the awesome study.

Could you let me know when you release the official code?

Precision Calculation

Hi, Thanks for the great work, I am really interested in your paper and recently tried to reimplement the codes. However, I am a bit confused about the metric calculation, especially balanced accuracy (bACC), I understand that bACC = (Sensitivity + Specificity) / 2, but it seems like the way you are calculating the bACC is just the top-1 accuracy. Is this a mistake?

Mixmatch on cifar-100 and stl-10

It is very nice of you to provide the results on cifar-10, can you please provide the python code for Mixmatch on cifar-100 and stl-10? Because we are not able to reproduce the results on these two datasets. Thank you so much!

CIfar100_Codes

Hi,

Could you also share the codes on how you split cifar 100 and stl-10 dataset?

Many thanks,
Jeff

optimize.newton fail to converge

hello, i have used darp on my own dataset, but while training, in opt_solver, it raise an error that Failed to converge after 30 iterations, value is 466315.4184961797. How can i solve this? Thank you.

Respect authors:

Hello, I have a question about the code. where is the file that is named 'progress.bar'? You mentioned the file name in the 'utils/init.py'.

Reproducing ReMixMatch + DARP + cRT on CIFAR-10

Thanks for the great work. I have a question regarding reproducing the results in table 6 in the supplementary material.

I was trying to reproduce the 79.7% bACC from ReMixMatch + DARP + cRT on CIFAR-10 with the imbalance ratio of 150. According to appendix E, I re-initialized the linear classifier and re-trained it for 20 epochs while freezing other parameters. The re-weighting loss I used is the standard re-weighting by frequency (your paper refers to the loss from [24], maybe mistaken?). The classifier was trained on the labeled dataset.

Other hyper-parameters are: SGD with a learning rate 0.1, momentum 0.9, and weight decay 0.0005. The learning rate of SGD decays by 0.01 at the 16th and 18th epoch.

Despite the random initialization, the best bACC I got is 76.94%. I wonder where I did wrong. Could you please share your implementation details for this experiment?

Thanks!

interesting work

Hi, when could you release the codes?