Image Segmentation models don't converge

Hello everyone,
I’m running an experiment of image segmentation, I have used multiple algorithms such as Unet, Unet++, MAnet, DeepLabV3 but none of them converge, the error is fluctuating but staying at the same high level after many epochs (I tried different number of epoch, as well as different batch sizes and learning rates) but I can’t seem to find an answer as to why it’s not working. Here are snippets of my code if anybody could help me out I would be very greatfull.

Here’s my model

and here’s how I created the dataset before putting it through the data loaders, my inputs are images and their masks

As decoder I used efficient net, resnet, VGG but none of them work as well !