vasgaowei

Thanks for your response, it would be nice if you help refine this repository. Looking forward to your pull request.

Hi, thanks for your attention. Can you tell me the error or post a screenshot？

Hi, thanks for your attention. The config file has an error. You can try to update `conformer_ts_cam_small_patch16` to `conformer_tscam_small_patch16` in line 7 in `conformer_tscam_small_patch16_224.yaml`

If training model under the same setting as paper, the result should not differ a lot. Could you try running the code on one or two GPUs?

From your scripts, you have already trained **ts-cam** based on **deit-base** backbone, and saved corresponding model weights. You may try adding a line in **test_cam.py**: `from models.vgg import vgg16_cam` 

From your screenshot, the` top1` (`Cls@1`) and `top5` (`Cls@5`) classification accuracy are 68.7% and 88.9% respectively. And `top1` (`Loc@1`) and `top5` (`Loc@5`) localization accuracy are 47.4% and 59.4%, respectively. The...

Hi, thanks for your attention. The localization accuracies are a little lower than those in the paper. Maybe you can try to train the model yourself, seeing the final results.

Hi, thanks for your attention. I think the attention for the cls_token can reflect the relative importance of different patches for classification. The reason is that cls_token is updated in...

Hello, thanks for your attention. Which kind of transformer do you use as backbone, deit-tiny, deit-small or deit-large? If you choose deit-small, and the default setting, e.g, batch_size, learning-rate, `MODEL.CAM_THR`,...

Hi, thanks for your attention. Maybe the "model" in your script is an instance of the class **VisionTransformer**. And class **VisionTransformer** only needs two arguments: one is **self**, another is...