TorchMetrics in PyTorch Lightning

TorchMetrics was originally created as part of PyTorch Lightning, a powerful deep learning research framework designed for scaling models without boilerplate.

Note

TorchMetrics always offers compatibility with the last 2 major PyTorch Lightning versions, but we recommend to always keep both frameworks up-to-date for the best experience.

While TorchMetrics was built to be used with native PyTorch, using TorchMetrics with Lightning offers additional benefits:

  • Modular metrics are automatically placed on the correct device when properly defined inside a LightningModule. This means that your data will always be placed on the same device as your metrics. No need to call .to(device) anymore!

  • Native support for logging metrics in Lightning using self.log inside your LightningModule.

  • The .reset() method of the metric will automatically be called at the end of an epoch.

The example below shows how to use a metric in your LightningModule:

class MyModel(LightningModule):

    def __init__(self, num_classes):
        ...
        self.accuracy = torchmetrics.classification.Accuracy(task="multiclass", num_classes=num_classes)

    def training_step(self, batch, batch_idx):
        x, y = batch
        preds = self(x)
        ...
        # log step metric
        self.accuracy(preds, y)
        self.log('train_acc_step', self.accuracy)
        ...

    def on_train_epoch_end(self):
        # log epoch metric
        self.log('train_acc_epoch', self.accuracy)

Metric logging in Lightning happens through the self.log or self.log_dict method. Both methods only support the logging of scalar-tensors. While the vast majority of metrics in torchmetrics returns a scalar tensor, some metrics such as ConfusionMatrix, ROC, MeanAveragePrecision, ROUGEScore return outputs that are non-scalar tensors (often dicts or list of tensors) and should therefore be dealt with separately. For info about the return type and shape please look at the documentation for the compute method for each metric you want to log.

Logging TorchMetrics

Logging metrics can be done in two ways: either logging the metric object directly or the computed metric values. When Metric objects, which return a scalar tensor are logged directly in Lightning using the LightningModule self.log method, Lightning will log the metric based on on_step and on_epoch flags present in self.log(...). If on_epoch is True, the logger automatically logs the end of epoch metric value by calling .compute().

Note

sync_dist, sync_dist_group and reduce_fx flags from self.log(...) don’t affect the metric logging in any manner. The metric class contains its own distributed synchronization logic.

This however is only true for metrics that inherit the base class Metric, and thus the functional metric API provides no support for in-built distributed synchronization or reduction functions.

class MyModule(LightningModule):

    def __init__(self, num_classes):
        ...
        self.train_acc = torchmetrics.classification.Accuracy(task="multiclass", num_classes=num_classes)
        self.valid_acc = torchmetrics.classification.Accuracy(task="multiclass", num_classes=num_classes)

    def training_step(self, batch, batch_idx):
        x, y = batch
        preds = self(x)
        ...
        self.train_acc(preds, y)
        self.log('train_acc', self.train_acc, on_step=True, on_epoch=False)

    def validation_step(self, batch, batch_idx):
        logits = self(x)
        ...
        self.valid_acc(logits, y)
        self.log('valid_acc', self.valid_acc, on_step=True, on_epoch=True)

As an alternative to logging the metric object and letting Lightning take care of when to reset the metric etc. you can also manually log the output of the metrics.

class MyModule(LightningModule):

    def __init__(self):
        ...
        self.train_acc = torchmetrics.classification.Accuracy(task="multiclass", num_classes=num_classes)
        self.valid_acc = torchmetrics.classification.Accuracy(task="multiclass", num_classes=num_classes)

    def training_step(self, batch, batch_idx):
        x, y = batch
        preds = self(x)
        ...
        batch_value = self.train_acc(preds, y)
        self.log('train_acc_step', batch_value)

    def on_train_epoch_end(self):
        self.train_acc.reset()

    def validation_step(self, batch, batch_idx):
        logits = self(x)
        ...
        self.valid_acc.update(logits, y)

    def on_validation_epoch_end(self, outputs):
        self.log('valid_acc_epoch', self.valid_acc.compute())
        self.valid_acc.reset()

Note that logging metrics this way will require you to manually reset the metrics at the end of the epoch yourself. In general, we recommend logging the metric object to make sure that metrics are correctly computed and reset. Additionally, we highly recommend that the two ways of logging are not mixed as it can lead to wrong results.

Note

When using any Modular metric, calling self.metric(...) or self.metric.forward(...) serves the dual purpose of calling self.metric.update() on its input and simultaneously returning the metric value over the provided input. So if you are logging a metric only on epoch-level (as in the example above), it is recommended to call self.metric.update() directly to avoid the extra computation.

class MyModule(LightningModule):

    def __init__(self, num_classes):
        ...
        self.valid_acc = torchmetrics.classification.Accuracy(task="multiclass", num_classes=num_classes)

    def validation_step(self, batch, batch_idx):
        logits = self(x)
        ...
        self.valid_acc.update(logits, y)
        self.log('valid_acc', self.valid_acc, on_step=True, on_epoch=True)

Common Pitfalls

The following contains a list of pitfalls to be aware of:

  • Modular metrics contain internal states that should belong to only one DataLoader. In case you are using multiple DataLoaders, it is recommended to initialize a separate modular metric instances for each DataLoader and use them separately. The same holds for using separate metrics for training, validation and testing.

class MyModule(LightningModule):

    def __init__(self, num_classes):
        ...
        self.val_acc = nn.ModuleList(
            [torchmetrics.classification.Accuracy(task="multiclass", num_classes=num_classes) for _ in range(2)]
        )

    def val_dataloader(self):
        return [DataLoader(...), DataLoader(...)]

    def validation_step(self, batch, batch_idx, dataloader_idx):
        x, y = batch
        preds = self(x)
        ...
        self.val_acc[dataloader_idx](preds, y)
        self.log('val_acc', self.val_acc[dataloader_idx])
  • Mixing the two logging methods by calling self.log("val", self.metric) in {training|validation|test}_step method and then calling self.log("val", self.metric.compute()) in the corresponding on_{train|validation|test}_epoch_end method. Because the object is logged in the first case, Lightning will reset the metric before calling the second line leading to errors or nonsense results.

  • If you decorate a lightning method with the rank_zero_only decorator with the goal of only calculating a particular

    metric on the main process, you need to disable the default behavior of the metric to synchronize the metric values across all processes. This can be done by setting the sync_on_compute flag to False when initializing the metric. Not doing so can lead to race conditions and processes hanging.

class MyModule(LightningModule):

    def __init__(self, num_classes):
        ...
        self.metric = torchmetrics.image.FrechetInceptionDistance(sync_on_compute=False)

    @rank_zero_only
    def validation_step(self, batch, batch_idx):
        image, target = batch
        generated_image = self(x)
        ...
        self.metric(image, real=True)
        self.metric(generated_image, real=False)
        val = self.metric.compute()  # this will only be called on the main process
        self.log('val_fid', val)
  • Calling self.log("val", self.metric(preds, target)) with the intention of logging the metric object. Because self.metric(preds, target) corresponds to calling the forward method, this will return a tensor and not the metric object. Such logging will be wrong in this case. Instead, it is essential to separate into several lines:

def training_step(self, batch, batch_idx):
    x, y = batch
    preds = self(x)
    ...
    # log step metric
    self.accuracy(preds, y)  # compute metrics
    self.log('train_acc_step', self.accuracy)  # log metric object
  • Using MetricTracker wrapper with Lightning is a special case, because the wrapper in itself is not a metric i.e. it does not inherit from the base Metric class but instead from ModuleList. Thus, to log the output of this metric one needs to manually log the returned values (not the object) using self.log and for epoch level logging this should be done in the appropriate on_{train|validation|test}_epoch_end method.