Training loss remains stable!


I am working on the MS-COCO dataset for the multi-label classification task. While training my neural network the model does not learn anything. I mean training loss values are nearly similar and training loss remains stable (e.g. loss value min. ~ 0.68, max. ~0.71 ) after many iterations. I use the ResNet18 model with BCELogitLoss and AdamW optimizer, and no data augmentation. I tried many things such as change learning rate, weight_decay, batch size, data augmentation, different model architecture, and different loss, etc. But there are not any changes at all. Do you have any ideas or suggestions?

Thank you very much in advance!