Performance problem

[zkcys001]

Thank you for the great work!

I have been training the network with your code. In the Pretraining stage, I use 8 GPU to train in the cifar10 task and use this command: python main_supcon.py --batch_size 1024
--learning_rate 0.5
--temp 0.1
--cosine

Then, in the Linear evaluation stage, I also adopt the following code to train:

python main_linear.py --batch_size 512
--learning_rate 5
--ckpt /path/to/model.pth

But the performance is 95.58. But your report performance is 96.0. Could you check it is normal? Thanks~

Kecheng

[HobbitLong]

Hi, @zkcys001 ,

I got two runs, one giving 95.9% and the other 96.0%. Your number might fall into the variation range (I don't have enough data to put a std here tho), but you may try different learning rates in the linear evaluation stage to double check things.

[Hibb-bb]

Hi @HobbitLong , may I ask which checkpoint did you use? I also get a similar accuracy of 95.56% using the checkpoint of the last epoch. The previous checkpoints like 350, 600 showed worse than 95.56%