I’m having an issue reloading a saved model that I can’t figure out.
I train a AWD-LSTM language model to a final loss of 0.2. When I generate sequences from the model, they are generally high quality. When I reload the model, it’s suddenly terrible.
For a specific example input, the loss of the original model is 0.14, while the loss of the re-loaded model on the same input is 2.5.
Does anyone have an idea as to why this might be happening?