poplar.nn.training

Training routines and associated helper functions.

Functions

train(model, data, n_epochs, n_batches, loss_function)	Train/test loop for an instance of LinearModel. This function allows for some basic monitoring of the training process, including regular loss curve plots
train_test_split(data, ratio[, device, dtype])	Splits data into two instances of torch.tensor with sizes of ratio ratio along their first axis. Also supports device

Module Contents

poplar.nn.training.train(model: poplar.nn.networks.LinearModel, data: list, n_epochs: int, n_batches: int, loss_function: Any, optimiser=None, verbose=False, plot=True, update_every=1, n_test_batches=None, save_best=False, scheduler=None, outdir='models')

Train/test loop for an instance of LinearModel. This function allows for some basic monitoring of the training process, including regular loss curve plots and a command line output indicating current progress.

If you need more complex functionality, it is advisable to write your own training function using this as a starting point.

Parameters:

modelLinearModel

Instance of LinearModel to be trained.

datalist

List of torch.Tensors for the training data, training labels, testing data and testing labels respectively.

n_epochsint

Number of epochs to train for.

n_batchesint

Number of batches per epoch.

loss_functionAny

Loss function to use. It is recommended to use one of the pytorch loss functions (https://pytorch.org/docs/stable/nn.html#loss-functions)

optimiserAny, optional

The pytorch optimiser to use, by default the Adam optimiser with a learning rate of 1e-3 is used. This should be instantiated before passing to this function.

verbosebool, optional

If True, also displays training progress on the command line. By default False

plotbool, optional

If True, loss curves are regularly produced (with interval update_every) and saved in the model directory, by default True

update_everyint, optional

Number of epochs between updating the saved model and plotting diagnostic data, by default 1

n_test_batchesint, optional

Number of batches to run the testing data in, by default n_batches

save_bestbool, optional

If True, saves the network that achieved the lowest validation losses, by default False

schedulerAny, optional

pytorch scheduler to use for learning rate adjustment during training, by default None

outdirstr, optional

Output directory for the trained model directory, by default ‘models’

poplar.nn.training.train_test_split(data: torch.tensor | numpy.ndarray | list, ratio: int, device=None, dtype=None)

Splits data into two instances of torch.tensor with sizes of ratio ratio along their first axis. Also supports device switching and dtype casting.

Parameters:

datatorch.tensor, numpy.ndarray or list

The tensors/ndarrays (or list of tensors/ndarrays) to be split.

ratioint

The ratio between the sizes of the two output tensors along their first axis.

devicestr, optional

device to move tensors to, by default None (maintains device of inputs)

dtypeoptional

data type of output tensors, by default None (the same dtype as the input is returned)

Returns:

list of tensors

A list of the two split tensors.