ignite.metrics

Metrics provide a way to compute various quantities of interest in an online fashion without having to store the entire output history of a model.

In practice a user needs to attach the metric instance to an engine. The metric value is then computed using the output of the engine’s process_function:

def process_function(engine, batch):
    # ...
    return y_pred, y

engine = Engine(process_function)
metric = Accuracy()
metric.attach(engine, "accuracy")

If the engine’s output is not in the format (y_pred, y) or {'y_pred': y_pred, 'y': y, ...}, the user can use the output_transform argument to transform it:

def process_function(engine, batch):
    # ...
    return {'y_pred': y_pred, 'y_true': y, ...}

engine = Engine(process_function)

def output_transform(output):
    # `output` variable is returned by above `process_function`
    y_pred = output['y_pred']
    y = output['y_true']
    return y_pred, y  # output format is according to `Accuracy` docs

metric = Accuracy(output_transform=output_transform)
metric.attach(engine, "accuracy")

Note

Most of implemented metrics are adapted to distributed computations and reduce their internal states across supported devices before computing metric value. This can be helpful to run the evaluation on multiple nodes/GPU instances/TPUs with a distributed data sampler. Following code snippet shows in detail how to use metrics:

device = "cuda:{}".format(local_rank)
model = torch.nn.parallel.DistributedDataParallel(model,
                                                  device_ids=[local_rank, ],
                                                  output_device=local_rank)
test_sampler = DistributedSampler(test_dataset)
test_loader = DataLoader(
    test_dataset,
    batch_size=batch_size,
    sampler=test_sampler,
    num_workers=num_workers,
    pin_memory=True
)

evaluator = create_supervised_evaluator(model, metrics={'accuracy': Accuracy()}, device=device)

Note

Metrics cannot be serialized using pickle module because the implementation is based on lambda functions. Therefore, use the third party library dill to overcome the limitation of pickle.

Metric arithmetics

Metrics could be combined together to form new metrics. This could be done through arithmetics, such as metric1 + metric2, use PyTorch operators, such as (metric1 + metric2).pow(2).mean(), or use a lambda function, such as MetricsLambda(lambda a, b: torch.mean(a + b), metric1, metric2).

For example:

precision = Precision(average=False)
recall = Recall(average=False)
F1 = (precision * recall * 2 / (precision + recall)).mean()

Note

This example computes the mean of F1 across classes. To combine precision and recall to get F1 or other F metrics, we have to be careful that average=False, i.e. to use the unaveraged precision and recall, otherwise we will not be computing F-beta metrics.

Metrics also support indexing operation (if metric’s result is a vector/matrix/tensor). For example, this can be useful to compute mean metric (e.g. precision, recall or IoU) ignoring the background:

cm = ConfusionMatrix(num_classes=10)
iou_metric = IoU(cm)
iou_no_bg_metric = iou_metric[:9]  # We assume that the background index is 9
mean_iou_no_bg_metric = iou_no_bg_metric.mean()
# mean_iou_no_bg_metric.compute() -> tensor(0.12345)

How to create a custom metric

To create a custom metric one needs to create a new class inheriting from Metric and override three methods :

• reset() : resets internal variables and accumulators

• update() : updates internal variables and accumulators with provided batch output (y_pred, y)

• compute() : computes custom metric and return the result

For example, we would like to implement for illustration purposes a multi-class accuracy metric with some specific condition (e.g. ignore user-defined classes):

from ignite.metrics import Metric
from ignite.exceptions import NotComputableError

# These decorators helps with distributed settings
from ignite.metrics.metric import sync_all_reduce, reinit__is_reduced


class CustomAccuracy(Metric):

    def __init__(self, ignored_class, output_transform=lambda x: x):
        self.ignored_class = ignored_class
        self._num_correct = None
        self._num_examples = None
        super(CustomAccuracy, self).__init__(output_transform=output_transform)

    @reinit__is_reduced
    def reset(self):
        self._num_correct = 0
        self._num_examples = 0
        super(CustomAccuracy, self).reset()

    @reinit__is_reduced
    def update(self, output):
        y_pred, y = output

        indices = torch.argmax(y_pred, dim=1)

        mask = (y != self.ignored_class)
        mask &= (indices != self.ignored_class)
        y = y[mask]
        indices = indices[mask]
        correct = torch.eq(indices, y).view(-1)

        self._num_correct += torch.sum(correct).item()
        self._num_examples += correct.shape[0]

    @sync_all_reduce("_num_examples", "_num_correct")
    def compute(self):
        if self._num_examples == 0:
            raise NotComputableError('CustomAccuracy must have at least one example before it can be computed.')
        return self._num_correct / self._num_examples

We imported necessary classes as Metric, NotComputableError and decorators to adapt the metric for distributed setting. In reset method, we reset internal variables _num_correct and _num_examples which are used to compute the custom metric. In updated method we define how to update the internal variables. And finally in compute method, we compute metric value.

We can check this implementation in a simple case:

import torch
torch.manual_seed(8)

m = CustomAccuracy(ignored_class=3)

batch_size = 4
num_classes = 5

y_pred = torch.rand(batch_size, num_classes)
y = torch.randint(0, num_classes, size=(batch_size, ))

m.update((y_pred, y))
res = m.compute()

print(y, torch.argmax(y_pred, dim=1))
# Out: tensor([2, 2, 2, 3]) tensor([2, 1, 0, 0])

print(m._num_correct, m._num_examples, res)
# Out: 1 3 0.3333333333333333

Metrics and its usages

By default, Metrics are epoch-wise, it means

• reset() is triggered every EPOCH_STARTED (See Events).

• update() is triggered every ITERATION_COMPLETED.

• compute() is triggered every EPOCH_COMPLETED.

Usages can be user defined by creating a class inheriting for MetricUsage. See the list below of usages.

Complete list of usages

• MetricUsage

• EpochWise

• BatchWise

• BatchFiltered

Metrics and distributed computations

In the above example, CustomAccuracy has reset, update, compute methods decorated with reinit__is_reduced, sync_all_reduce. The purpose of these features is to adapt metrics in distributed computations on supported backend and devices (ignite.distributed). More precisely, in the above example we added @sync_all_reduce("_num_examples", "_num_correct") over compute method. This means that when compute method is called, metric’s interal variables self._num_examples and self._num_correct are summed up over all participating devices. Therefore, once collected, these internal variables can be used to compute the final metric value.

Complete list of metrics

• Accuracy

• Average

• ConfusionMatrix

• DiceCoefficient()

• EpochMetric

• Fbeta()

• GeometricAverage

• IoU()

• mIoU()

• Loss

• MeanAbsoluteError

• MeanPairwiseDistance

• MeanSquaredError

• Metric

• MetricsLambda

• Precision

• Recall

• RootMeanSquaredError

• RunningAverage

• TopKCategoricalAccuracy

• VariableAccumulation

class ignite.metrics.Accuracy(output_transform=<function Accuracy.<lambda>>, is_multilabel=False, device=None)

Calculates the accuracy for binary, multiclass and multilabel data.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

• y_pred must be in the following shape (batch_size, num_categories, …) or (batch_size, …).

• y must be in the following shape (batch_size, …).

• y and y_pred must be in the following shape of (batch_size, num_categories, …) and num_categories must be greater than 1 for multilabel cases.

In binary and multilabel cases, the elements of y and y_pred should have 0 or 1 values. Thresholding of predictions can be done as below:

def thresholded_output_transform(output):
    y_pred, y = output
    y_pred = torch.round(y_pred)
    return y_pred, y

binary_accuracy = Accuracy(thresholded_output_transform)

Parameters

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• is_multilabel (bool, optional) – flag to use in multilabel case. By default, False.

• device (str of torch.device, optional) – unused argument.

class ignite.metrics.Average(output_transform=<function Average.<lambda>>, device=None)

Helper class to compute arithmetic average of a single variable.

• update must receive output of the form x.

• x can be a number or torch.Tensor.

Note

Number of samples is updated following the rule:

• +1 if input is a number

• +1 if input is a 1D torch.Tensor

• +batch_size if input is an ND torch.Tensor. Batch size is the first dimension (shape[0]).

For input x being an ND torch.Tensor with N > 1, the first dimension is seen as the number of samples and is summed up and added to the accumulator: accumulator += x.sum(dim=0)

Examples:

evaluator = ...

custom_var_mean = Average(output_transform=lambda output: output['custom_var'])
custom_var_mean.attach(evaluator, 'mean_custom_var')

state = evaluator.run(dataset)
# state.metrics['mean_custom_var'] -> average of output['custom_var']

Parameters

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• device (str of torch.device, optional) – optional device specification for internal storage.

class ignite.metrics.ConfusionMatrix(num_classes, average=None, output_transform=<function ConfusionMatrix.<lambda>>, device=None)

Calculates confusion matrix for multi-class data.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

• y_pred must contain logits and has the following shape (batch_size, num_categories, …)

• y should have the following shape (batch_size, …) and contains ground-truth class indices

  with or without the background class. During the computation, argmax of y_pred is taken to determine predicted classes.

Parameters

• num_classes (int) – number of classes. See notes for more details.

• average (str, optional) – confusion matrix values averaging schema: None, “samples”, “recall”, “precision”. Default is None. If average=”samples” then confusion matrix values are normalized by the number of seen samples. If average=”recall” then confusion matrix values are normalized such that diagonal values represent class recalls. If average=”precision” then confusion matrix values are normalized such that diagonal values represent class precisions.

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• device (str of torch.device, optional) – optional device specification for internal storage.

Note

In case of the targets y in (batch_size, …) format, target indices between 0 and num_classes only contribute to the confusion matrix and others are neglected. For example, if num_classes=20 and target index equal 255 is encountered, then it is filtered out.

ignite.metrics.DiceCoefficient(cm, ignore_index=None)

Calculates Dice Coefficient for a given ConfusionMatrix metric.

Parameters

• cm (ConfusionMatrix) – instance of confusion matrix metric

• ignore_index (int, optional) – index to ignore, e.g. background index

Return type

MetricsLambda

class ignite.metrics.EpochMetric(compute_fn, output_transform=<function EpochMetric.<lambda>>, check_compute_fn=True)

Class for metrics that should be computed on the entire output history of a model. Model’s output and targets are restricted to be of shape (batch_size, n_classes). Output datatype should be float32. Target datatype should be long.

Warning

Current implementation stores all input data (output and target) in as tensors before computing a metric. This can potentially lead to a memory error if the input data is larger than available RAM.

Warning

Current implementation does not work with distributed computations. Results are not gather across all devices and computed results are valid for a single device only.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

If target shape is (batch_size, n_classes) and n_classes > 1 than it should be binary: e.g. [[0, 1, 0, 1], ].

Parameters

• compute_fn (callable) – a callable with the signature (torch.tensor, torch.tensor) takes as the input predictions and targets and returns a scalar.

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• check_compute_fn (bool) – if True, compute_fn is run on the first batch of data to ensure there are no issues. If issues exist, user is warned that there might be an issue with the compute_fn.

Warning

EpochMetricWarning: User is warned that there are issues with compute_fn on a batch of data processed.

ignite.metrics.Fbeta(beta, average=True, precision=None, recall=None, output_transform=None, device=None)

Calculates F-beta score

Parameters

• beta (float) – weight of precision in harmonic mean

• average (bool, optional) – if True, F-beta score is computed as the unweighted average (across all classes in multiclass case), otherwise, returns a tensor with F-beta score for each class in multiclass case.

• precision (Precision, optional) – precision object metric with average=False to compute F-beta score

• recall (Precision, optional) – recall object metric with average=False to compute F-beta score

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. It is used only if precision or recall are not provided.

• device (str of torch.device, optional) – optional device specification for internal storage.

Returns

MetricsLambda, F-beta metric

Return type

MetricsLambda

class ignite.metrics.GeometricAverage(output_transform=<function GeometricAverage.<lambda>>, device=None)

Helper class to compute geometric average of a single variable.

• update must receive output of the form x.

• x can be a positive number or a positive torch.Tensor, such that torch.log(x) is not nan.

Note

Number of samples is updated following the rule:

• +1 if input is a number

• +1 if input is a 1D torch.Tensor

• +batch_size if input is a ND torch.Tensor. Batch size is the first dimension (shape[0]).

For input x being an ND torch.Tensor with N > 1, the first dimension is seen as the number of samples and is aggregated and added to the accumulator: accumulator *= prod(x, dim=0)

Parameters

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• device (str of torch.device, optional) – optional device specification for internal storage.

ignite.metrics.IoU(cm, ignore_index=None)

Calculates Intersection over Union using ConfusionMatrix metric.

Parameters

• cm (ConfusionMatrix) – instance of confusion matrix metric

• ignore_index (int, optional) – index to ignore, e.g. background index

Returns

MetricsLambda

Return type

MetricsLambda

Examples:

train_evaluator = ...

cm = ConfusionMatrix(num_classes=num_classes)
IoU(cm, ignore_index=0).attach(train_evaluator, 'IoU')

state = train_evaluator.run(train_dataset)
# state.metrics['IoU'] -> tensor of shape (num_classes - 1, )

ignite.metrics.mIoU(cm, ignore_index=None)

Calculates mean Intersection over Union using ConfusionMatrix metric.

Parameters

• cm (ConfusionMatrix) – instance of confusion matrix metric

• ignore_index (int, optional) – index to ignore, e.g. background index

Returns

MetricsLambda

Return type

MetricsLambda

Examples:

train_evaluator = ...

cm = ConfusionMatrix(num_classes=num_classes)
mIoU(cm, ignore_index=0).attach(train_evaluator, 'mean IoU')

state = train_evaluator.run(train_dataset)
# state.metrics['mean IoU'] -> scalar

class ignite.metrics.Loss(loss_fn, output_transform=<function Loss.<lambda>>, batch_size=<function Loss.<lambda>>, device=None)

Calculates the average loss according to the passed loss_fn.

Parameters

• loss_fn (callable) – a callable taking a prediction tensor, a target tensor, optionally other arguments, and returns the average loss over all observations in the batch.

• output_transform (callable) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs. The output is expected to be a tuple (prediction, target) or (prediction, target, kwargs) where kwargs is a dictionary of extra keywords arguments. If extra keywords arguments are provided they are passed to loss_fn.

• batch_size (callable) – a callable taking a target tensor that returns the first dimension size (usually the batch size).

• device (str of torch.device, optional) – unused argument.

class ignite.metrics.MeanAbsoluteError(output_transform=<function Metric.<lambda>>, device=None)

Calculates the mean absolute error.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

Parameters

• output_transform (Callable) –

• device (Optional[Union[device, str]]) –

class ignite.metrics.MeanPairwiseDistance(p=2, eps=1e-06, output_transform=<function MeanPairwiseDistance.<lambda>>, device=None)

Calculates the mean pairwise distance: average of pairwise distances computed on provided batches.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

Parameters

• p (int) –

• eps (float) –

• output_transform (Callable) –

• device (Optional[Union[device, str]]) –

class ignite.metrics.MeanSquaredError(output_transform=<function Metric.<lambda>>, device=None)

Calculates the mean squared error.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

Parameters

• output_transform (Callable) –

• device (Optional[Union[device, str]]) –

class ignite.metrics.Metric(output_transform=<function Metric.<lambda>>, device=None)

Base class for all Metrics.

Parameters

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs. By default, metrics require the output as (y_pred, y) or {'y_pred': y_pred, 'y': y}.

• device (str of torch.device, optional) – optional device specification for internal storage.

attach(engine, name, usage=<ignite.metrics.metric.EpochWise object>)

Attaches current metric to provided engine. On the end of engine’s run, engine.state.metrics dictionary will contain computed metric’s value under provided name.

Parameters

• engine (Engine) – the engine to which the metric must be attached

• name (str) – the name of the metric to attach

• usage (str or MetricUsage, optional) – the usage of the metric. Valid string values should be ‘EpochWise.usage_name’ (default) or ‘BatchWise.usage_name’.

Return type

None

Example:

metric = ...
metric.attach(engine, "mymetric")

assert "mymetric" in engine.run(data).metrics

assert metric.is_attached(engine)

Example with usage:

metric = ...
metric.attach(engine, "mymetric", usage=BatchWise.usage_name)

assert "mymetric" in engine.run(data).metrics

assert metric.is_attached(engine, usage=BatchWise.usage_name)

completed(engine, name)

Helper method to compute metric’s value and put into the engine. It is automatically attached to the engine with attach().

Parameters

• engine (Engine) – the engine to which the metric must be attached

• name (str) –

Return type

None

abstract compute()

Computes the metric based on it’s accumulated state.

By default, this is called at the end of each epoch.

Returns

the actual quantity of interest. However, if a Mapping is returned, it will be (shallow) flattened into engine.state.metrics when completed() is called.

Return type

Any

Raises

NotComputableError – raised when the metric cannot be computed.

detach(engine, usage=<ignite.metrics.metric.EpochWise object>)

Detaches current metric from the engine and no metric’s computation is done during the run. This method in conjunction with attach() can be useful if several metrics need to be computed with different periods. For example, one metric is computed every training epoch and another metric (e.g. more expensive one) is done every n-th training epoch.

Parameters

• engine (Engine) – the engine from which the metric must be detached

• usage (str or MetricUsage, optional) – the usage of the metric. Valid string values should be ‘epoch_wise’ (default) or ‘batch_wise’.

Return type

None

Example:

metric = ...
engine = ...
metric.detach(engine)

assert "mymetric" not in engine.run(data).metrics

assert not metric.is_attached(engine)

Example with usage:

metric = ...
engine = ...
metric.detach(engine, usage="batch_wise")

assert "mymetric" not in engine.run(data).metrics

assert not metric.is_attached(engine, usage="batch_wise")

is_attached(engine, usage=<ignite.metrics.metric.EpochWise object>)

Checks if current metric is attached to provided engine. If attached, metric’s computed value is written to engine.state.metrics dictionary.

Parameters

• engine (Engine) – the engine checked from which the metric should be attached

• usage (str or MetricUsage, optional) – the usage of the metric. Valid string values should be ‘epoch_wise’ (default) or ‘batch_wise’.

Return type

bool

iteration_completed(engine)

Helper method to update metric’s computation. It is automatically attached to the engine with attach().

Parameters

engine (Engine) – the engine to which the metric must be attached

Return type

None

abstract reset()

Resets the metric to it’s initial state.

By default, this is called at the start of each epoch.

Return type

None

started(engine)

Helper method to start data gathering for metric’s computation. It is automatically attached to the engine with attach().

Parameters

engine (Engine) – the engine to which the metric must be attached

Return type

None

abstract update(output)

Updates the metric’s state using the passed batch output.

By default, this is called once for each batch.

Parameters

output – the is the output from the engine’s process function.

Return type

None

class ignite.metrics.MetricsLambda(f, *args, **kwargs)

Apply a function to other metrics to obtain a new metric. The result of the new metric is defined to be the result of applying the function to the result of argument metrics.

When update, this metric does not recursively update the metrics it depends on. When reset, all its dependency metrics would be resetted. When attach, all its dependency metrics would be attached automatically (but partially, e.g is_attached() will return False).

Parameters

• f (callable) – the function that defines the computation

• args (sequence) – Sequence of other metrics or something else that will be fed to f as arguments.

Example:

precision = Precision(average=False)
recall = Recall(average=False)

def Fbeta(r, p, beta):
    return torch.mean((1 + beta ** 2) * p * r / (beta ** 2 * p + r + 1e-20)).item()

F1 = MetricsLambda(Fbeta, recall, precision, 1)
F2 = MetricsLambda(Fbeta, recall, precision, 2)
F3 = MetricsLambda(Fbeta, recall, precision, 3)
F4 = MetricsLambda(Fbeta, recall, precision, 4)

When check if the metric is attached, if one of its dependency metrics is detached, the metric is considered detached too.

engine = ...
precision = Precision(average=False)

aP = precision.mean()

aP.attach(engine, "aP")

assert aP.is_attached(engine)
# partially attached
assert not precision.is_attached(engine)

precision.detach(engine)

assert not aP.is_attached(engine)
# fully attached
assert not precision.is_attached(engine)

class ignite.metrics.Precision(output_transform=<function Precision.<lambda>>, average=False, is_multilabel=False, device=None)

Calculates precision for binary and multiclass data.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

• y_pred must be in the following shape (batch_size, num_categories, …) or (batch_size, …).

• y must be in the following shape (batch_size, …).

In binary and multilabel cases, the elements of y and y_pred should have 0 or 1 values. Thresholding of predictions can be done as below:

def thresholded_output_transform(output):
    y_pred, y = output
    y_pred = torch.round(y_pred)
    return y_pred, y

precision = Precision(output_transform=thresholded_output_transform)

In multilabel cases, average parameter should be True. However, if user would like to compute F1 metric, for example, average parameter should be False. This can be done as shown below:

precision = Precision(average=False, is_multilabel=True)
recall = Recall(average=False, is_multilabel=True)
F1 = precision * recall * 2 / (precision + recall + 1e-20)
F1 = MetricsLambda(lambda t: torch.mean(t).item(), F1)

Warning

In multilabel cases, if average is False, current implementation stores all input data (output and target) in as tensors before computing a metric. This can potentially lead to a memory error if the input data is larger than available RAM.

Warning

In multilabel cases, if average is False, current implementation does not work with distributed computations. Results are not reduced across the GPUs. Computed result corresponds to the local rank’s (single GPU) result.

Parameters

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• average (bool, optional) – if True, precision is computed as the unweighted average (across all classes in multiclass case), otherwise, returns a tensor with the precision (for each class in multiclass case).

• is_multilabel (bool, optional) – parameter should be True and the average is computed across samples, instead of classes.

• device (str of torch.device, optional) – unused argument.

class ignite.metrics.Recall(output_transform=<function Recall.<lambda>>, average=False, is_multilabel=False, device=None)

Calculates recall for binary and multiclass data.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

• y_pred must be in the following shape (batch_size, num_categories, …) or (batch_size, …).

• y must be in the following shape (batch_size, …).

In binary and multilabel cases, the elements of y and y_pred should have 0 or 1 values. Thresholding of predictions can be done as below:

def thresholded_output_transform(output):
    y_pred, y = output
    y_pred = torch.round(y_pred)
    return y_pred, y

recall = Recall(output_transform=thresholded_output_transform)

In multilabel cases, average parameter should be True. However, if user would like to compute F1 metric, for example, average parameter should be False. This can be done as shown below:

precision = Precision(average=False, is_multilabel=True)
recall = Recall(average=False, is_multilabel=True)
F1 = precision * recall * 2 / (precision + recall + 1e-20)
F1 = MetricsLambda(lambda t: torch.mean(t).item(), F1)

Warning

In multilabel cases, if average is False, current implementation stores all input data (output and target) in as tensors before computing a metric. This can potentially lead to a memory error if the input data is larger than available RAM.

Warning

In multilabel cases, if average is False, current implementation does not work with distributed computations. Results are not reduced across the GPUs. Computed result corresponds to the local rank’s (single GPU) result.

Parameters

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• average (bool, optional) – if True, precision is computed as the unweighted average (across all classes in multiclass case), otherwise, returns a tensor with the precision (for each class in multiclass case).

• is_multilabel (bool, optional) – parameter should be True and the average is computed across samples, instead of classes.

• device (str of torch.device, optional) – unused argument.

class ignite.metrics.RootMeanSquaredError(output_transform=<function Metric.<lambda>>, device=None)

Calculates the root mean squared error.

• update must receive output of the form (y_pred, y) or {‘y_pred’: y_pred, ‘y’: y}.

Parameters

• output_transform (Callable) –

• device (Optional[Union[device, str]]) –

class ignite.metrics.RunningAverage(src=None, alpha=0.98, output_transform=None, epoch_bound=True, device=None)

Compute running average of a metric or the output of process function.

Parameters

• src (Metric or None) – input source: an instance of Metric or None. The latter corresponds to engine.state.output which holds the output of process function.

• alpha (float, optional) – running average decay factor, default 0.98

• output_transform (callable, optional) – a function to use to transform the output if src is None and corresponds the output of process function. Otherwise it should be None.

• epoch_bound (boolean, optional) – whether the running average should be reset after each epoch (defaults to True).

• device (str of torch.device, optional) – unused argument.

Examples:

alpha = 0.98
acc_metric = RunningAverage(Accuracy(output_transform=lambda x: [x[1], x[2]]), alpha=alpha)
acc_metric.attach(trainer, 'running_avg_accuracy')

avg_output = RunningAverage(output_transform=lambda x: x[0], alpha=alpha)
avg_output.attach(trainer, 'running_avg_loss')

@trainer.on(Events.ITERATION_COMPLETED)
def log_running_avg_metrics(engine):
    print("running avg accuracy:", engine.state.metrics['running_avg_accuracy'])
    print("running avg loss:", engine.state.metrics['running_avg_loss'])

class ignite.metrics.TopKCategoricalAccuracy(k=5, output_transform=<function TopKCategoricalAccuracy.<lambda>>, device=None)

Calculates the top-k categorical accuracy.

• update must receive output of the form (y_pred, y) or {'y_pred': y_pred, 'y': y}.

Parameters

• output_transform (Callable) –

• device (Optional[Union[device, str]]) –

class ignite.metrics.VariableAccumulation(op, output_transform=<function VariableAccumulation.<lambda>>, device=None)

Single variable accumulator helper to compute (arithmetic, geometric, harmonic) average of a single variable.

• update must receive output of the form x.

• x can be a number or torch.Tensor.

Note

The class stores input into two public variables: accumulator and num_examples. Number of samples is updated following the rule:

• +1 if input is a number

• +1 if input is a 1D torch.Tensor

• +batch_size if input is a ND torch.Tensor. Batch size is the first dimension (shape[0]).

Parameters

• op (callable) – a callable to update accumulator. Method’s signature is (accumulator, output). For example, to compute arithmetic mean value, op = lambda a, x: a + x.

• output_transform (callable, optional) – a callable that is used to transform the Engine’s process_function’s output into the form expected by the metric. This can be useful if, for example, you have a multi-output model and you want to compute the metric with respect to one of the outputs.

• device (str of torch.device, optional) – optional device specification for internal storage.

class ignite.metrics.MetricUsage(started, completed, iteration_completed)

Base class for all usages of metrics.

A usage of metric defines the events when a metric starts to compute, updates and completes. Valid events are from Events.

Parameters

• started – event when the metric starts to compute. This event will be associated to started().

• completed – event when the metric completes. This event will be associated to completed().

• iteration_completed – event when the metric updates. This event will be associated to iteration_completed().

class ignite.metrics.EpochWise

Epoch-wise usage of Metrics. It’s the default and most common usage of metrics.

Metric’s methods are triggered on the following engine events:

• started() on every EPOCH_STARTED (See Events).

• iteration_completed() on every ITERATION_COMPLETED.

• completed() on every EPOCH_COMPLETED.

class ignite.metrics.BatchWise

Batch-wise usage of Metrics.

Metric’s methods are triggered on the following engine events:

• started() on every ITERATION_STARTED (See Events).

• iteration_completed() on every ITERATION_COMPLETED.

• completed() on every ITERATION_COMPLETED.

class ignite.metrics.BatchFiltered(*args, **kwargs)

Batch filtered usage of Metrics. This usage is similar to epoch-wise but update event is filtered.

Metric’s methods are triggered on the following engine events:

• started() on every EPOCH_STARTED (See Events).

• iteration_completed() on filtered ITERATION_COMPLETED.

• completed() on every EPOCH_COMPLETED.

Parameters

args (sequence) – arguments for the setup of ITERATION_COMPLETED handled by iteration_completed().

ignite.metrics.metric.sync_all_reduce(*attrs)

Helper decorator for distributed configuration to collect instance attribute value across all participating processes.

See ignite.metrics on how to use it.

Parameters

*attrs – attribute names of decorated class

Return type

Callable

ignite.metrics.metric.reinit__is_reduced(func)

Helper decorator for distributed configuration.

See ignite.metrics on how to use it.

Parameters

func (Callable) –

Return type

Callable