有关 3D Vision Transformer 训练的问题: 模型未收敛

Hi, this is my first time working on a Transformer model, in this case, a 3D vision Transformer model, I am working on a 3d medical image classification task, and the training set is around 300 3D images; here is what the image input looks like (1, 224, 224, 32); here, 1 is the number of channels, and 32 is the z dim size. I trained my data set on 3D efficientnet, and the accuracy was around 80%. I tried a 3D vision Transformer, but the model does not converge. Can you please review the code below? Why does the model not learn? Do you know if I am doing something wrong? Do you have any help or suggestions? Thank you in advance.