Trainer

Trainer is an orchestrator class used to streamline, train, and manage the training process.

Parameters:

Name	Type	Description	Default
model	Module	Model to train.	required
optimizer	Optimizer	Optimizer used for parameter updates.	required
loss		Loss function used during training.	required

Example:

trainer = ntk.Trainer(
    model=my_model,
    optimizer=ntk.Adam(parameters=my_model.parameters()),
    loss=ntk.MeanSquaredError()
)

fit

Train the model on the provided data.

Training data may be provided as raw tensors, a Dataset, or a DataLoader. When using raw input tensors, labels must be provided through the y argument.

Parameters:

Name	Type	Description	Default
data		Training data.	required
y		Training labels when using raw input tensors.	None
epochs	int	Number of training epochs.	1
batch_size	int	Number of samples per batch.	None
shuffle	bool	Whether to shuffle the training data.	None
drop_remainder_batch	bool	Whether to drop incomplete batches.	None
validation_split	float	Fraction of training data reserved for validation.	None
validation_data		Separate validation dataset or dataloader.	None
verbose	bool	Whether to track progess in the terminal. Defaults to True	None

Returns:

Type	Description
History	After training has completed, a History object is returned that can be used to plot training metrics overtime.

Example

history = trainer.fit(dataloader, epochs=100)

history.plot("loss", "val_loss)

Implementatino

trainer = ntk.Trainer(
    model=my_model,
    optimizer=ntk.Adam(parameters=my_model.parameters()),
    loss=ntk.BinaryCrossEntropy()
)

history = trainer.fit(dataloader, epochs=300)
history.plot("loss")

History Plot Output

History Metrics

The History object tracks two training metrics (with more to come), loss and validation loss, which can be plotted like so:

history.plot("loss", "val_loss")