LeViT training and bench on GTSRB dataset about levit HOT 3 CLOSED

Is the problem that the training accuracy goes up to 100%, but that it does not generalize to the test set?
LeViT was created with ImageNet in mind, and the statistics of ImageNet images are quite different to GTSRB images.
Also, the data augmentation (no crop augmentation, etc) might be ok for ConvNets, but suboptimal for ViT models.

from levit.

Hi,

Thank you for your reply, I have tried to adjust the training hyperparameters and data augmentation. The LeVit_128S now can achieve 30% accuracy - 346FPS, and LeViT_192 60% accuracy - FPS 277. While the MobileNetV3 Large can achieve 98.37%-355FPS, GhostNet 1.0 97.26%-275FPS. I use the same training and testing method for 3 models but the LeVit models do not fit my method and I really want to improve. That's why I contact you, I would love to see your model working as expected. :( The benchmark illustrates your LeViT_128s have higher accuracy and speed than both MobileNet and GhostNet. I'm looking forward to your advice.
I wonder whether my training method is correct: I download the Pytorch model of MobileNetV3, GhostNet, and your LeVit then load it with config (large for MobileNetV3, width = 1.0 for GhostNet, default for LeVit_128S/192), and num_classes = 43 for all 3 models. Then I use the training source as above. After that, I save the model and then load it in another script for testing.

(https://colab.research.google.com/drive/1yPpAtsibYw8jkzOarLBhdXnWvJR0Uqlo?usp=sharing#scrollTo=r377ZaaKKImT)

from levit.

Sorry, but I don't know enough about GTSRB to help.

from levit.