About COCO Suponly

[Luffy03]

Hello, thanks for your inspiring work! I tried to reproduce the experiments of COCO dataset (SupOnly) with your provided code 'supvised.py'. But the results are far from your reported results. Would you please provide your logs and checkpoints for COCO (SupOnly)?

[LiheYoung]

Thank you for your interest.

Indeed we did not run the COCO SupOnly experiments by ourselves. These results are borrowed from PC²Seg. You can run our semi-supervised codes on the COCO. Or you may upload your SupOnly logs first, then I will also try to run them when I am free. :)

[Luffy03]

Thanks for your kind comment. Here is my training log for COCO (SupOnly 1/512) with code 'supervised.py'. I conducted the experiment with 8 3090 GPUs, but the results are very weird. 20221002_162900.txt

[LiheYoung]

Below are some of my explorations with our supervised.py. Overall, the COCO SupOnly experiments should be trained for more epochs, due to the extremely limited data and challenging class space. I use a total batch size of 16 (4 images/gpu on 4 gpus), with a base learning rate 0.004, and lr_multi is set as 10.0 following the original config.	Epochs	1/512	1/256	1/128	1/64	1/32
30	0.86	0.78	6.79	24.94	37.26
120	1.67	15.55	31.77	37.89	43.17
480	18.27	27.08	33.16	39.07	43.14
PC2Seg	22.9	28.0	33.6	37.8	42.2

The reproduced results with more training iterations are close to that reported in PC²Seg. It is recommended to set fixed training iterations across different partitions.

As aforementioned, we directly borrow the SupOnly results from PC²Seg. To reproduce our work, you may try the semi-supervised codes of FixMatch or UniMatch. Thank you.

[Luffy03]

Thanks for your kind comment. Your explorations seem intuitive. By the way, would you please provide the results of FixMatch? It would help me a lot.

[LiheYoung]

Do you mean the FixMatch results on the COCO dataset? If so, you may refer to our Appendix A. We have included FixMatch results there.

[Luffy03]

Sorry for my neglect. Thanks for your kind answer.