Jiaming Han

You can try with newer Pytorch. I tested the pretrained checkpoint on Pytorch 1.12 and it still works.

@SoonFa Just set some configs in `test.py` https://github.com/csuhan/VFA/blob/e35411eb22b4fc48b524debe58dc7c09be2bf9a6/test.py#L30-L37

Thanks! They are similar in `parameter freeze`, but different in `cosine classifier` and `last layer weight init`. Please check Tab.1 in our paper.

@hhxzzy Actually, we do not use `cosine classifier` and `last layer init` in the open-sourced version as it takes additional `checkpoint combination step`, similar to https://github.com/ucbdrive/few-shot-object-detection/blob/master/tools/ckpt_surgery.py If we want to...

@873552584 We do not use t-SNE in the paper. Do you mean Similarity matrix visualization? You can store the latent region features on disk and calculate their cosine similarities. Then...

@hhxzzy We do not report bAP here, but the full GFSOD results should contain AP, bAP and nAP.

Yes. But we only report nAP.

@hhxzzy Yes. This difference makes minor effect on nAP.

The current model is not trained on joint multimodal data, so it may not perform well at the test time.

Hi @Cece1031 , hope the script in https://github.com/csuhan/OneLLM/issues/29 can help you.