Low Accuracy Than Reported

Yunfan-Li / Contrastive-Clustering

Code for the paper "Contrastive Clustering" (AAAI 2021)

MIT License

297 stars 91 forks source link

Open wangtuo13 opened 3 years ago

wangtuo13 commented 3 years ago

I run the code in win10+python3.7+pytorch1.8.1+cu111 and get 21% accuracy on Cifar10 dataset.

Yunfan-Li commented 3 years ago

How many epochs have you trained?

haoweiclouds1 commented 2 years ago

i also have this problem , i trained with 1000 epoch as the same as you. please give some suggestion,thank you very much!

haoweiclouds1 commented 2 years ago

results as follows: ( run cluster.py) Features shape (60000,) NMI = 0.1276 ARI = 0.0685 F = 0.1795 ACC = 0.2311

Yunfan-Li commented 2 years ago

Please make sure that you reload the model trained for 1000 epoch. Perhaps the model at epoch 0 is loaded for clustering.

haoweiclouds1 commented 2 years ago

thank you i will try again

haoweiclouds1 commented 2 years ago

sorry ,I return to ask you some questions.

I have trained for 1000 epoch and the model has saved,

how can i load in cluster.py in your code (eg :i find the path i have saved : Contrastive-Clustering-main/save/CIFAR-10/checkpoint_990.tar)

thank you very much!

Yunfan-Li commented 2 years ago

Changing the ''start epoch'' number in the config file to 990 will do.

haoweiclouds1 commented 2 years ago

hh thank you ,i find !

haoweiclouds1 commented 2 years ago

your work is very wonderful!!

Yunfan-Li commented 2 years ago

Thank you!

Yunfan-Li commented 2 years ago

您好，可以先试一下用release的模型测试看下结果是否正确，然后确保您没有改动训练的代码和config中的超参数，重新训练看下结果，因为软硬件环境的不同，根据目前的反馈可能有1-2个点的偏差，但是0.5一定是哪里出了问题。

haoweiclouds1 commented 2 years ago

好的我试试感谢