Testing accuracy significantly greater than Training and Validation accuracies

I’m training a CNN using Transfer Learning, resnet34 to be precise, the accuracies of training and validation set are not improving. They are stuck at 72% to 18% respectively. I trained the model for 25 epochs, and it didn’t show any signs of improvement at all. But when I tested the model, the accuracy was nowhere near the training and validation set, it was 86%. Can anyone explain why this is happening?