Accuracy and Confusion Matrix calculations

I am training a resnet34 model using model.fine_tune() and get a train_loss of 1.128555, valid_loss of 1.281499 and an error_rate of 0.53333 after the final epoch (I am using LabelSmoothingCrossEntropy as the loss function).

However when I am plotting the confusion matrix using ClassificationInterpreter, I get the matrix as follows:

           Class 1    Class 2    Class 3

Class 1 53 8 13
Class 2 19 107 8
Class 3 25 10 42

which should yield a lower error_rate? Does it matter if I saved and reloaded the model as below?

data = ImageDataLoaders.from_path_func(path, files, label_func, num_workers=0)
model = cnn_learner(data, resnet34, loss_func=LabelSmoothingCrossEntropy(), metrics=error_rate)
model.fine_tune(1)
lr_min, lr_steep = model.lr_find()
valid_loss, error_rate = model.validate()
model.save('model')

data = ImageDataLoaders.from_path_func(path, files, label_func, num_workers=0)
model = cnn_learner(data, resnet34, loss_func=LabelSmoothingCrossEntropy(), metrics=error_rate)
model.load('model')
interp = ClassificationInterpretation.from_learner(model)
interp.plot_confusion_matrix()