search

syp2ysy / SVF

[NeurIPS 2022] Singular Value Fine-tuning: Few-shot Segmentation requires Few-parameters Fine-tuning
MIT License
69 stars 6 forks source link

Reimplement the svf method based on the PFENet code and the result is inferior than original PFENet. #8

Closed lyphui closed 1 year ago

lyphui commented 2 years ago

Thanks for your nice work. I implement the svf method based on the PFENet code and the result is inferior than original PFENet. I followed the parameter setting in PFENet. And I used the same dataset and parameter setting in both original PFENet and PFENet with SVF. But the performance of PFENet with SVF performs worse than original PFENet. Could you please share your full-training code rather than svf module code? Thanks a lot!

YAOSL98 commented 2 years ago

I also met this question. I used SVF method in Meta-DETR (a few-shot detection method), and I decompose backbone layer3 and layer4. My results also decrease.

syp2ysy commented 2 years ago

@lyphui The reason for this is detailed in this issue https://github.com/syp2ysy/SVF/issues/3

syp2ysy commented 2 years ago

@YAOSL98 Sorry, we haven't tried it on the detection task yet. The exact reason may not be very clear. We will later verify the effectiveness of SVF in the detection task, and update the details on the homepage.