torchtraining.functional¶

Submodules

Module providing functions (like metrics, losses) directly usable.

Note

IMPORTANT: This module should be rarely used and non-functional counterparts should be preferred.

For example this:

class TrainStep(tt.steps.Train):
    def forward(self, module, sample):
        images, labels = sample
        # Assume predictions are obtained
        ...
        return loss, predictions, labels


train_step = TrainStep(criterion, device)


iteration = (
    tt.iterations.Train(train_step, ...)
    ** tt.Select(predictions=1, labels=2)
    ** tt.metrics.classification.binary.Accuracy()
    ** tt.accumulators.Mean()
    ** tt.callbacks.Log("Accuracy")
)

Should be preferred instead of this (notice accuracy calculation in step):

class TrainStep(tt.steps.Train):
    def forward(self, module, sample):
        images, labels = sample
        # Assume predictions are obtained
        ...
        accuracy = tt.functional.metrics.classification.binary(predictions, labels)
        return loss, accuracy


train_step = TrainStep(criterion, device)

iteration = (
    tt.iterations.Train(train_step, ...)
    ** tt.Select(accuracy=1)
    ** tt.accumulators.Mean()
    ** tt.callbacks.Log("Accuracy")
)

Second approach has the following shortcomings:

calculation of metrics is mixed with what your network actually does with inputs and what it produces

Step’s’ forward function has more limited usage. If user wants to calculate other metrics they have to change step manually instead of simply adding another ** pipe

Note

IMPORTANT: Only reasonable modules to use (currently) in functional manner (ha, irony), should be torchtraining.functional.inputs and torchtraining.functional.data.