search

TACJu / TransFG

This is the official PyTorch implementation of the paper "TransFG: A Transformer Architecture for Fine-grained Recognition" (Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, Changhu Wang, Alan Yuille).
MIT License
382 stars 88 forks source link

Accuracy on the CAR dataset #13

Open Thea1 opened 3 years ago

Thea1 commented 3 years ago

To the best of my own ability, I can only achieve up to 90% accuracy on the car dataset. Is there something wrong with me? I would like to ask if the parameters of the training car dataset are set the same as the cub dataset?

Markin-Wang commented 3 years ago

Also, I run the ViT on cars dataset with batchsize 8, 20000 steps on GPU cards. The accuracy is 90.86, much less than the 93.7 reported in the paper. Waiting for the author to release the code for training the cars datasets.

hyao1 commented 2 years ago

have you solved the problem？I also had this problem

wywd commented 2 years ago

Also, I run the ViT on cars dataset with batchsize 8, 20000 steps on GPU cards. The accuracy is 90.86, much less than the 93.7 reported in the paper. Waiting for the author to release the code for training the cars datasets.

Hi. Have you solved this problem? I have the same problem. It's very troublesome!!!

Markin-Wang commented 2 years ago

No, I finally cannot achieve the reported results on CUB, Dogs and Car datasets. More importantly, there are a large disparities. Therefore, I have to include both the reported results and my reimplemented results on my paper on dog dataset and abadon the Car dataset. This really imposed a negative influence on my work during the paper review, but fortunatelly my paper was accpeted finally. We are still waiting for the response from authors. But at this stage, maybe you can consider to change another backbone and report the reimplemented results of transFG.

hongbo-sun commented 2 years ago

is there anyone have achieved the reported results on car? can u share the training setting？