search

Yunfan-Li / Contrastive-Clustering

Code for the paper "Contrastive Clustering" (AAAI 2021)
MIT License
297 stars 91 forks source link

weight_decay #22

Open YIBO-0805 opened 2 years ago

YIBO-0805 commented 2 years ago

Why the parameter 'weight_decay' set in code is 0.

Yunfan-Li commented 2 years ago

We didn't tune that hyper-parameter during the experiments. Perhaps a proper weight decay could further improve the performance.

YIBO-0805 commented 2 years ago

But I get only 72.3% accuracy on cifar10 dataset. I run it in pytorch1.9.0+cuda11.4.

Yunfan-Li commented 2 years ago

Maybe it is due to the different environment and random seed. The config we provided is exactly the one we used to produce the result reported in the paper. You may check the code again to make sure no additional modification is made and rerun the code with a different random seed.

mark375chen commented 2 years ago

We didn't tune that hyper-parameter during the experiments. Perhaps a proper weight decay could further improve the performance.

What might be an appropriate weight decay for CIFAR10?

Yunfan-Li commented 2 years ago

We didn't tune that hyper-parameter in our experiments. You may try the commonly used 1e-3 or 1e-4.