Reusing your models

The training of neural networks can be a time-consuming process, so we need to track its progress as the process goes on. We need to learn to save and reload a trained model.

Every MXNet model consist of at least two files:

• model_name-symbol.json: Describes the architecture
• model_name-epoch.params: Describes the model weights

You can create them yourself or you can download them from the internet. We will cover the process of using pretrained networks in the next chapter.

Saving the model

Saving the model during the training process is done using the mx.do_checkpoint callback. A few important parameters are as follows:

• prefix: This defines the prefix of the filenames to save the model
• frequency: The frequency is measured in epochs to save checkpoints

Let's move back to the MNIST example we have been working on and adjust the mx.fit function to include mx.do_checkpoint:

mx.fit(nnet, mx.ADAM(), train_data_provider, eval_data = validation_data_provider, n_epoch = 50, callbacks = ...