Reproducibility of MoCoV3, cifar10 with ResNet18

vturrisi / solo-learn

solo-learn: a library of self-supervised methods for visual representation learning powered by Pytorch Lightning

MIT License

1.42k stars 185 forks source link

Reproducibility of MoCoV3, cifar10 with ResNet18 #385

Open tsirif opened 6 months ago

tsirif commented 6 months ago

Hi there! Thank you for your work! I have tried to reproduce the reported results (~93% top-1) on CIFAR10 with ResNet18 using MoCoV3, however running the provided script with default parameter only gives about 90% top-1 in the online classifier.

Do you know what I am missing? Are the reported results with an offline trained linear-probe classifier? What are the hyperparameters needed in that case?

vturrisi commented 5 months ago

Hey. I'm pretty sure the results that we have for cifar are just the online eval. The parameters should be the same as in the script (I can check later to confirm) but it can be an issue with library versions. These results are pretty old, so maybe a new pytorch/lightning version changed some stuff.