How to train a DeepLab model with a dataset that contains different size of images without unifying their sizes?

Question

I have a dataset that contains different size of images, and I want to train it on a DeepLab model with pytorch without unifying their sizes . However, I learned that when training with batch , there must be same size in the same batch , unless set batchsize as 1, I'm wondering if there any method to train it, in a custom batch, and not uniform picture size .