Accuracy drop after loading model

hi i’m training a model for semantic segmentation . in each epoch model and two optimizer are saved and we have all the weights . i’m using resnet50 but each time i stop and continue training accuracy drops . i also don’t change the dataset or batch size and batches are normalized before training. when i continuously trained model , got 85% accuracy but after stop and resuming training accuracy dropped to 70% .

How are you saving the model?
Note that the optimizers have an state too (which have to be saved)
Otherwise, when you resume the statistics are reset (typically happens with Adam).

Screenshot from 2022-08-02 18-31-11

this is how . can it be because of the file format ? i saw some where that its better to use .pt format instead of .pth for making checkpoints