kai

This is normal. At the initial training stage, the model is noisy and less capable of generating pseudo labels. So in semi-supervised learning, we usually first train supervised data for...

The equation is the hard version of class-aware contrastive matrix without re-weighting. Equation 5 is the select_matrix * mask in line 55 and score_mask is the re-weighting factor. Also note...

Thanks for your notice, sorry for the misleading. i and j mean different indices of columns and rows of the mask matrix. When writing the eq(6) at i=j (same image's...

Thanks for your kind explanation. I think it is a bug in the code because mathematically setting the weight to 1 should be better. After setting the diagonal to 1,...

Here's the CoMatch config on semi-inat. I cannot remember whether it is the exact same config used in paper, but has comparable performance. [comatchccssl_exp512_seminat_b4x16.py.zip](https://github.com/TencentYoutuResearch/Classification-SemiCLS/files/10283403/comatchccssl_exp512_seminat_b4x16.py.zip)

Added. According to issue #4 , the code refactor changed the randomness of the code. Please try different seeds if needed. [config](https://github.com/TencentYoutuResearch/Classification-SemiCLS/blob/main/configs/ccssl/fixmatchccssl_exp512_cifar10_wres_x2_b1x64_l4000_soft.py)

I have tried several tests but didn't get results much better than 80% (In @aeo123 epoch40 got top-1 ACC > 84.4). Please check we are using the same training commands...

> hi @KaiWU5 here is my training log under torch1.6 environment and it seems that the results differ much compared with torch1.1x. I didn't complete the whole training process. Would...

My experiment has the same phenomenon. Only change torch1.6 to torch 1.10 and I got top-1 acc at 82.21 at epoch 28 which means it would be way better than...

Yes, the seeds are 1,2,3,4,5. The performance for each seed on CIFAR100 is label400: 59.79 | 59.54 | 62.83 | 59.55 | 62.65 label2500: 76.33 | 75.07 | 75.7725 |...