Arcana Losses

Submodules

arcana.losses.loss module

Custom loss module.

class arcana.losses.loss.CombinedHPLoss(*args: Any, **kwargs: Any)

Bases: Module

Combined huberized pinball loss for quantile regression.

forward(predicted, target, mask=None)

Compute the combined huberized quantile loss.

Parameters:

• predicted (torch.Tensor or np.ndarray) – predicted values

• target (torch.Tensor or np.ndarray) – target values

• mask (torch.Tensor or np.ndarray) – mask to apply to the loss (optional. Default: None)

Returns:

torch.Tensor or np.ndarray – mean loss

class arcana.losses.loss.CombinedLPLoss(*args: Any, **kwargs: Any)

Bases: Module

Combined logcosh pinball loss for quantile regression.

forward(predicted, target)

Compute the combined logcosh quantile loss.

Parameters:

• predicted (torch.Tensor or np.ndarray) – predicted values

• target (torch.Tensor or np.ndarray) – target values

Returns:

torch.Tensor or np.ndarray – mean loss

class arcana.losses.loss.HuberLoss(delta=1.0)

Bases: object

Huber loss for quantile regression.

class arcana.losses.loss.LogCoshLoss(*args: Any, **kwargs: Any)

Bases: Module

Log-cosh loss for quantile regression.

forward(predicted, target)

Compute the log-cosh loss.

Parameters:

• predicted (torch.Tensor or np.ndarray) – predicted values

• target (torch.Tensor or np.ndarray) – target values

Returns:

torch.Tensor or np.ndarray – mean loss

class arcana.losses.loss.LossFactory

Bases: object

Factory class for losses.

static create_loss(config)

Create a loss.

Parameters:

config (ModelConfig) – model config

Returns:

torch.nn.Module – loss function

class arcana.losses.loss.MSELoss

Bases: object

Mean squared error loss.

class arcana.losses.loss.PinballLoss(*args: Any, **kwargs: Any)

Bases: Module

Pinball loss for quantile regression.

forward(predicted, target, mask=None)

Compute the pinball loss.

Parameters:

• predicted (torch.Tensor or np.ndarray) – predicted values

• target (torch.Tensor or np.ndarray) – target values

• mask (torch.Tensor or np.ndarray) – mask to apply to the loss (optional. Default: None)

Returns:

torch.Tensor or np.ndarray – mean loss

pinball_loss_numpy(predicted, target, quantiles=None)

Compute the pinball loss.

class arcana.losses.loss.QuantileLoss(*args: Any, **kwargs: Any)

Bases: Module

Quantile loss for quantile regression.

forward(predicted, target)

Compute the quantile loss.

Parameters:

• predicted (torch.Tensor or np.ndarray) – predicted values

• target (torch.Tensor or np.ndarray) – target values

Returns:

torch.Tensor or np.ndarray – mean loss

arcana.losses.loss.is_numpy(x_vector)

Check if x is a numpy array.

Parameters:

x_vector (torch.Tensor or np.ndarray) – input vector

Returns:

bool – True if x is a numpy array, False otherwise

arcana.losses.loss.is_tensor(x_vector)

Check if x is a tensor.

Parameters:

x_vector (torch.Tensor or np.ndarray) – input vector

Returns:

bool – True if x is a tensor, False otherwise

arcana.losses.loss.to_numpy(x_vector)

Convert x to a numpy array. If x is a tensor, detach it from the graph and move it to the cpu.

Parameters:

x_vector (torch.Tensor or np.ndarray) – input vector

Returns:

np.ndarray – numpy array

arcana.losses.loss.to_tensor(x_vector, device=None)

Convert x to a tensor.

Parameters:

• x_vector (torch.Tensor or np.ndarray) – input vector

• device (torch.device) – device to move the tensor to

Returns:

torch.Tensor – tensor

arcana.losses.loss_trace module

Class to trace the loss

class arcana.losses.loss_trace.LossTrace(output_dim)

Bases: object

Class to trace the loss of the model during training and validation

calculate_batch_loss_train(train_loss)

Calculate the loss for each batch and add it to the loss_trace

Parameters:

train_loss (torch.Tensor) – loss for each batch

calculate_batch_loss_validation(validation_loss)

Calculate the validation loss for each batch and add it to the loss_trace

Parameters:

validation_loss (torch.Tensor) – validation loss for each batch

calculate_epoch_loss(train_loader, val_loader)

Calculate the average loss for each epoch and add it to the loss_trace

Parameters:

• train_loader (torch.utils.data.DataLoader) – training data loader

• val_loader (torch.utils.data.DataLoader) – validation data loader

reset_temp_loss_trace()

Reset the temporary loss trace

Module contents