torchdms.loss

Loss functions and functions relevant to losses.

Functions

diff_penalty	Computes l1 norm of the difference between adjacent betas for each latent dimension.
l1	l1_loss, perhaps with loss decay or target exponentiation
mse	mse_loss, perhaps with loss decay or target exponentiation
product_penalty	Computes l1 norm of product of betas across latent dimensions.
rmse	Root mean square error, perhaps with loss decay or target exponentiation.
sitewise_group_lasso	The sum of the 2-norm across columns.
sum_diff_penalty	Computes l1 norm of the difference between aggregated betas at adjacent sites for each latent dimension.
weighted_loss	Generic loss function decorator with loss decay or target exponentiation.

torchdms.loss.weighted_loss(base_loss)

Generic loss function decorator with loss decay or target exponentiation.

torchdms.loss.l1(y_true, y_predicted, loss_decay=None, exp_target=None)

l1_loss, perhaps with loss decay or target exponentiation

torchdms.loss.mse(y_true, y_predicted, loss_decay=None, exp_target=None)

mse_loss, perhaps with loss decay or target exponentiation

torchdms.loss.rmse(*args, **kwargs)

Root mean square error, perhaps with loss decay or target exponentiation.

torchdms.loss.sitewise_group_lasso(matrix)

The sum of the 2-norm across columns.

We omit the square root of the group sizes, as they are all constant in our case.

torchdms.loss.product_penalty(betas)

Computes l1 norm of product of betas across latent dimensions.

torchdms.loss.diff_penalty(betas)

Computes l1 norm of the difference between adjacent betas for each latent dimension.

torchdms.loss.sum_diff_penalty(betas)

Computes l1 norm of the difference between aggregated betas at adjacent sites for each latent dimension.