Different model output on validation during training versus inference with checkpoint
I train a model with pytorch lightning and save the checkpoint (ckpt) file with parameters for later use. It works well except for one new model.
Difference between accuracy during training and accuracy during testing
In the model below the accuracy reported at the end of the final validation stage is 0.46, but when reported during the manual testing the value is 0.53. What can account for this discrepancy?