Skip to content

loss

sorix.nn.loss

MSELoss

Computes the Mean Squared Error loss between the prediction and the target.

BCEWithLogitsLoss

This loss combines a Sigmoid layer and the BCELoss in one single class. More numerically stable than using a plain Sigmoid followed by a BCELoss.

CrossEntropyLoss 

CrossEntropyLoss(weight=None, one_hot=False)

Computes the cross entropy loss between input and target.

This criterion is useful when training a classification problem with C classes. If provided, the optional argument weight should be a 1D Tensor assigning weight to each of the classes. This is particularly useful for unbalanced training sets.

The input is expected to contain raw, unnormalized scores for each class. y_pred has to be a Tensor of size (minibatch, C).

The targets are expected to be class indices in the range [0, C-1] or one-hot encoded values.

The loss can be described as: L = - (1 / sum(w_yi)) * sum(w_yi * log(exp(x_i, yi) / sum(exp(x_i, j))))

Attributes:

  • weight (Optional[Tensor]) –

    A manual rescaling weight given to each class. If given, has to be a Tensor of size C.

  • one_hot (bool) –

    Whether the target labels are one-hot encoded.

Initializes the CrossEntropyLoss.

Parameters:

  • weight (Optional[Tensor], default: None ) –

    A manual rescaling weight given to each class. If given, has to be a Tensor of size C. Defaults to None.

  • one_hot (bool, default: False ) –

    Whether the target is one-hot encoded. Defaults to False.

Source code in sorix/nn/loss.py 
76
77
78
79
80
81
82
83
84
85
86
def __init__(self, weight: Optional[Tensor] = None, one_hot: bool = False) -> None:
    """Initializes the CrossEntropyLoss.

    Args:
        weight (Optional[Tensor], optional): A manual rescaling weight given to each class.
            If given, has to be a Tensor of size C. Defaults to None.
        one_hot (bool, optional): Whether the target is one-hot encoded. Defaults to False.
    """
    self.weight = weight
    self.one_hot = one_hot
    self.xp = np