Open yjh576 opened 3 months ago
Can you share the results that you get for Imagenet-V?
Here are some of my reported results, and the model seems to neither converge nor perform effectively.
Re-updating prompt initializations to current prompts. => Using native Torch AMP. Training in mixed precision. number of test samples: 10000 Running for 64 Augmented views Running for 1 TTA steps Test: [ 199/10000] Time 0.879 ( 1.067) Acc@1 0.00 ( 1.00) Acc@5 0.00 ( 2.50) Test: [ 399/10000] Time 0.891 ( 0.977) Acc@1 0.00 ( 0.75) Acc@5 0.00 ( 2.00)
I have the same issue reproducing results for v2. Did you manage to come up with some solutions?
@jameelhassan I wonder if you would provide some insights on reproducing results for the v2 dataset? Thanks a lot!
Hi @yjh576 and @Bostoncake The issue is in the dataset class. Unfortunately, I do not have access to the computing resources till mid-July. Will check this and update the code soon.
hi @jameelhassan I was wondering if the issue with the ImageNet-v2 dataset not being reproducible has been resolved. Could you please let me know when you might be planning to update the code? Thank you very much for your help.
Despite successfully reproducing results on other datasets, I am unable to replicate the results of AlignPrompt on ImageNet-V using your code. I observe that the accuracy on ImageNet-V is almost zero, and it seems that the model is unable to converge. I confirm that I have not altered the code and have implemented it correctly. Could you kindly provide suggestions or assistance? Your assistance would be greatly appreciated.