cifar10 and cifar100 parameters

[CandiceUIC]

Hey guys, thanks for your nice work.
I am now trying to re-produce your result presented in the paper. However, I could not reach your result on cifar10 and cifar100. cifar10 is just a bit lower (71.8% with 45% pairflip noise), cifar100 is much lower than your presented result (31% with 45%pairflip noise).
In the paper, you mentioned the parameter of batch size, learning rate etc. I am wondering if you have changed settings for different datasets? If so, could please share them?

Cheers
Candice

[devraj89]

This is also exactly what I have found.
Also strangely enough when I ran on PyTorch based on Cuda75 I am getting around ~34% but on Cuda80 I am getting around 31%.
@CandiceUIC Did you manage to find the reason for this ?

Thanks
Devraj

[CandiceUIC]

Cuda75 can reach 34%? That is much exciting. I will try on cuda75.
I don't know exactly what the reason could be for lower performance. But I am thinking about the python/pytorch version. I changed code from python2 to python3, I don't know if that matters. I remembered they presented 34.x in the paper. If so, 34% is close enough I guess.

[devraj89]

Hi @CandiceUIC if you require I can send you an yml file where I was getting results close to them.
But can you please let me know if you also observe not a smooth accuracy curve. In the paper, the accuracy of both the models is very smooth but when I ran it it was wildly varying.
Check #3

[CandiceUIC]

@devraj89
I actually got almost similar curve as mentioned in the paper. You should check the number of epoch. Because when I changed epoch number, the curve became not smooth. I guess that is because they have some parameters which are very related to epoch number.
Also, could you please share your file with me? My email address is: [email protected]
Cheers