High gap between losses

Hello;

I have an unsupervised network model to train with unlabeled datasets. While the model is training my learning curves are like the figure below. I use a similar model like ResNet18 with a smaller kernel size. I changed many hyperparameters like learning rate, weight_decay, batch size etc. Also, adding many augmentations to get more complex data is not working. Do you have any opinion to solve this problem?

Thank you!!!

train-val-fig-task4

Cheers,
G.