In CIFAR10, why the fine-tuning training accuracy (<90%) is much lower than test accuracy (96.6%)? about vision_transformer HOT 8 OPEN

This is probably due to the data augmentation. Check out the "augmented" training data:

https://colab.sandbox.google.com/github/google-research/vision_transformer/blob/master/vit_jax.ipynb#scrollTo=jFqi3h7yMEsB

You can see that those images actually are a lot harder to recognize and thus the training accuracy is lower. Training on these still helps the network to score better on the testset though. You can modify the Colab to train without data augmentation, and you should see how the train accuracy is better than the test accuracy.

from vision_transformer.

@andsteing but why there isn't such a phenomenon when training with ResNet/MobileNet which also uses data augmentation?

from vision_transformer.

I found when I use 0.5 mean and std as follows, the test accuracy can be higher (96%) and the training accuracy is super low (82%), but when I change it to the real mean and std, the test accuracy drops to 93% but the training accuracy is good at 95%.

# CIFAR_MEAN = [0.49139968, 0.48215827, 0.44653124]
# CIFAR_STD = [0.24703233, 0.24348505, 0.26158768]

CIFAR_MEAN = [0.5, 0.5, 0.5]
CIFAR_STD = [0.5, 0.5, 0.5]

from vision_transformer.

Did you use ResNet/MobileNet on the same input pipeline? (Using data augmentation on different datasets with different parameters will produce different results)

I just re-ran the Colab with data augmentation disabled and got 0.97265625 on test set and 0.979321 on train set (i.e. slightly higher on train set, as expected).

Where exactly did you modify CIFAR_MEAN and CIFAR_STD ?

from vision_transformer.

my code is based on my own PyTorch Implementation. I fixed the issue of low training accuracy by changing the transforms from

transform_train = transforms.Compose([
    transforms.RandomSizedCrop((args.img_size, args.img_size), scale=(0.05, 1.0)),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize(CIFAR_MEAN, CIFAR_STD),
])

to
transform_train = transforms.Compose([
transforms.Resize(args.img_size),
transforms.RandomCrop(args.img_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(CIFAR_MEAN, CIFAR_STD),
])

This confirms that the issue belongs to "RandomSizedCrop".

Also the 0.5 normalization can increase the accuracy around 3%.

But now I still only get 96.5% test accuracy, cannot replicate the 98% test accuracy. May I know what other data augmentation or tricks should I use?

from vision_transformer.

Indeed, I think in your snippet the observed difference has more to due with RandomSizedCrop() than with Normalize().

It might be interesting to see if this observed difference is larger for ViT than for ResNet/MobileNet, but I would expect for ResNet/MobileNet also to have a deterioration in training accuracy if you apply RandomSizedCrop() with the same parameters.

from vision_transformer.

Could you help to check my code? It is a single file that expresses the training and tricks.
https://github.com/FedML-AI/FedML/blob/fed-transformer/fedml_experiments/centralized/fed_transformer/main_vit.py

I am not sure whether my reimplementation is correct or not.

from vision_transformer.

Indeed, I think in your snippet the observed difference has more to due with RandomSizedCrop() than with Normalize().

It might be interesting to see if this observed difference is larger for ViT than for ResNet/MobileNet, but I would expect for ResNet/MobileNet also to have a deterioration in training accuracy if you apply RandomSizedCrop() with the same parameters.

I never experienced such a phenomenon before. I don't think ResNet has such an issue when using RandomSizedCrop()

from vision_transformer.