Same architecture of ViT trained on CIFAR10 and MNIST in both Pytorch and Tensorflow adapted from kentaroy47/vision-transformers-cifar10 and sneakatyou/ViT-Tensorflow-2.0.
Even both implementations were trained with the identical architecture and hyper-parameters, validation accuracy 80.30 of the Pytorch implementation is not consistent with that of Tensorflow couterpart which is 70.15.
(coming soon)
tensorflow
tensorflow_addons
(for GELU activation)tensorflow_datasets
(for loading datasets)pytorch
torchvision
einops
wandb
(for logging data with minimal effort)