About the self-supervised model on cifar10

[luan-g]

Hi,
I have trained self-supervised model on cifar10 with PreAct-ResNet18 and ResNet18. On the 90% noise rate, I achieve the performance about 93.46 with ResNet18, but get a worse result about 92.8 with PreAct-ResNet18. However, the performance with PreAct-ResNet18 in the paper is 93.57. And the available self-supervised model for cifar10 in Google Drive is ResNet18, I achieve the performance about 93.54 by using this model.
I notice that the performance with PreAct-ResNet18 in the paper is actually more similar to ResNet18. Is there any mistake? Or how can I reproduce the performance with PreAct-ResNet18?