Hi @ybendou ,
Thank you very much for open-sourcing your code and for the insightful work presented in your paper. I’m currently using your method as a reference baseline in my research on few-shot image classification, and I truly appreciate the clarity and reproducibility efforts in your repository.
I'm reaching out to kindly ask if you could share the numerical results of your method on ImageNet under various few-shot settings (e.g., 1-shot, 2-shot, 4-shot, 8-shot, and 16-shot). I’ve been trying to reproduce the results following your instructions, but due to GPU memory limitations (32GB VRAM, e.g., A100), I can only set augment-epoch to 1 — increasing it further leads to out-of-memory errors.
With augment-epoch=1, I observe performance that is approximately 0.5% lower than the numbers reported in the paper across several shot settings. I suspect this gap may be attributed to the reduced augmentation schedule, but I’d like to confirm this by comparing against your original results under the same evaluation protocol.
If possible, could you please provide:
The exact accuracy numbers for each shot setting (1/2/4/8/16) on ImageNet from your experiments?
Or, any suggestions on how to reduce memory usage while preserving performance?
This would greatly help ensure fair comparisons in my study and deepen my understanding of your approach.
Thank you again for your time and for contributing such valuable work to the community. I’m happy to provide more details about my setup if needed.
Hi @ybendou ,
Thank you very much for open-sourcing your code and for the insightful work presented in your paper. I’m currently using your method as a reference baseline in my research on few-shot image classification, and I truly appreciate the clarity and reproducibility efforts in your repository.
I'm reaching out to kindly ask if you could share the numerical results of your method on ImageNet under various few-shot settings (e.g., 1-shot, 2-shot, 4-shot, 8-shot, and 16-shot). I’ve been trying to reproduce the results following your instructions, but due to GPU memory limitations (32GB VRAM, e.g., A100), I can only set augment-epoch to 1 — increasing it further leads to out-of-memory errors.
With augment-epoch=1, I observe performance that is approximately 0.5% lower than the numbers reported in the paper across several shot settings. I suspect this gap may be attributed to the reduced augmentation schedule, but I’d like to confirm this by comparing against your original results under the same evaluation protocol.
If possible, could you please provide:
The exact accuracy numbers for each shot setting (1/2/4/8/16) on ImageNet from your experiments?
Or, any suggestions on how to reduce memory usage while preserving performance?
This would greatly help ensure fair comparisons in my study and deepen my understanding of your approach.
Thank you again for your time and for contributing such valuable work to the community. I’m happy to provide more details about my setup if needed.