Is there a way to load a pre-trained model.
I have tried the load("model.joblib") and save("model.joblib", model) functions but the loaded model only ha about 10% accuracy on the validation data, and successfully generates about 10% successful adversarial examples. The model accuracy was about 99.3% before saving, and about 87% of adversarial examples generated were successful. If I train the loaded the loaded model for the same number of epochs as the original then I get the expected accuracy and adversarial example generation rate.
Is there a way to so save the model so that it can be loaded without needing to be retrained?
After calling load(filename) you need to call model.get_logits(x) and CrossEntropy again to update the values used for calculating loss and accuracy.