this happened after finished the training the model for 50 epochs and resume the training later to be 70 which means add additional 20 epochs … Does that normal i mean resuming the training from the last epoch … Does the curve normal ?
Based on your description the jump in validation accuracy happens when you are resuming the training. If so, I would check for data leaks in your script as it seems the validation data might now contain some or all of the training data.
Yes, the splitting indices would be reasonable to check first. If you are storing the file names, you could also compare them making sure training samples are not used in the validation run.
3 images in total is quite a low number of samples. Image augmentation could help, but using augmentation should not cause any data leaks assuming the data splitting is still valid.