Transforms¶

Transforms are usefull for adjusting the target and predictions before the metrics are calculated. For example in the classification case if the output of a neural network is the class probabilities/logits i.e. tensor with shape (N,C), the ArgmaxTransform(dim=-1) can be used to transform into labels before any metrics are calculated.

To create your own transformation, you can simply inherent from the base class Transform and implement your own __init__ and forward methods. For an example, here is the implementation of the ArgmaxTransform:

class ArgmaxTransform(Transform):
    def __init__(self, dim=-1):
        self.dim = dim

    def forward(self, target, pred):
        return target, pred.argmax(dim=self.dim)

The forward method of a transform is alwarys expected to take in two arguments (target, pred) and return two variables, the transformed target and pred.

Transform¶

class pytorch_metrics.transforms.Transform¶

ComposeTransforms¶

class pytorch_metrics.transforms.ComposeTransforms(*transforms)¶

Compose multiple transforms into one single transform. The transforms: will be compose in a sequential matter.

Parameters: transforms – sequence of transforms to compose into one.

DefaultTransform¶

class pytorch_metrics.transforms.DefaultTransform¶: Default transform that does nothing to the target and predictions

ArgmaxTransform¶

class pytorch_metrics.transforms.ArgmaxTransform(dim=-1)¶

Applies the argmax transform to the predictions. Often used to transform probabilities/logits into labels in multiclass tasks.

Parameters: dim – int, which dimension to apply the argmax transform over

RoundTransform¶

class pytorch_metrics.transforms.RoundTransform(dim=-1)¶

Applies a round transform to the predictions. Often used to transform the output from sigmoid activations into labels in binary tasks

Parameters: dim – int, which dimension to apply the round transform over

ROCTransform¶

class pytorch_metrics.transforms.ROCTransform(line)¶