Metrics

Metric API

IMetric

class catalyst.metrics._metric.IMetric(compute_on_call: bool = True)

Bases: abc.ABC

Interface for all Metrics.

Parameters

compute_on_call – Computes and returns metric value during metric call. Used for per-batch logging. default: True

__init__(compute_on_call: bool = True)

Interface for all Metrics.

abstract compute() → Any

Computes the metric based on it’s accumulated state.

By default, this is called at the end of each loader (on_loader_end event).

Returns

computed value, # noqa: DAR202 it’s better to return key-value

Return type

Any

abstract reset() → None

Resets the metric to it’s initial state.

By default, this is called at the start of each loader (on_loader_start event).

abstract update(*args, **kwargs) → Any

Updates the metrics state using the passed data.

By default, this is called at the end of each batch (on_batch_end event).

Parameters

• *args – some args :)

• **kwargs – some kwargs ;)

ICallbackBatchMetric

class catalyst.metrics._metric.ICallbackBatchMetric(compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.IMetric

@TODO: docs here

__init__(compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

@TODO: docs here

abstract compute_key_value() → Dict[str, float]

Computes the metric based on it’s accumulated state.

By default, this is called at the end of each loader (on_loader_end event).

Returns

computed value in key-value format. # noqa: DAR202

Return type

Dict

abstract update_key_value(*args, **kwargs) → Dict[str, float]

@TODO: docs here

ICallbackLoaderMetric

class catalyst.metrics._metric.ICallbackLoaderMetric(compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.IMetric

Interface for all Metrics.

Parameters

• compute_on_call – @TODO: docs

• prefix – @TODO: docs

• suffix – @TODO: docs

AccumulationMetric

class catalyst.metrics._metric.AccumulationMetric(accumulative_fields: Optional[Iterable[str]] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackLoaderMetric

This metric accumulates all the input data along loader

Parameters

• accumulative_fields – list of keys to accumulate data from batch

• compute_on_call – if True, allows compute metric’s value on call

• prefix – metric prefix

• suffix – metric suffix

General Metrics

AdditiveValueMetric

class catalyst.metrics._additive.AdditiveValueMetric(compute_on_call: bool = True)

Bases: catalyst.metrics._metric.IMetric

This metric computes mean and std values of input data.

Parameters

compute_on_call – if True, computes and returns metric value during metric call

ConfusionMatrixMetric

class catalyst.metrics._confusion_matrix.ConfusionMatrixMetric(num_classes: int, normalized: bool = False, compute_on_call: bool = True)

Bases: catalyst.metrics._metric.IMetric

Constructs a confusion matrix for a multiclass classification problems.

Parameters

• num_classes – number of classes in the classification problem

• normalized – determines whether or not the confusion matrix is normalized or not

• compute_on_call – Boolean flag to computes and return confusion matrix during __call__. default: True

FunctionalBatchMetric

class catalyst.metrics._functional_metric.FunctionalBatchMetric(metric_fn: Callable, metric_key: str, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackBatchMetric

Class for custom metrics in a functional way.

Parameters

• metric_fn – metric function, that get outputs, targets and return score as torch.Tensor

• metric_key – metric name

• compute_on_call – Computes and returns metric value during metric call. Used for per-batch logging. default: True

• prefix – metric prefix

• suffix – metric suffix

Note

Loader metrics calculated as average over all batch metrics.

Runner Metrics

Accuracy - AccuracyMetric

class catalyst.metrics._accuracy.AccuracyMetric(topk_args: Optional[List[int]] = None, num_classes: Optional[int] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackBatchMetric

This metric computes accuracy for multiclass classification case. It computes mean value of accuracy and it’s approximate std value (note that it’s not a real accuracy std but std of accuracy over batch mean values).

Parameters

• topk_args – list of topk for accuracy@topk computing

• num_classes – number of classes

• compute_on_call – if True, computes and returns metric value during metric call

• prefix – metric prefix

• suffix – metric suffix

Accuracy - MultilabelAccuracyMetric

class catalyst.metrics._accuracy.MultilabelAccuracyMetric(threshold: Union[float, torch.Tensor] = 0.5, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._additive.AdditiveValueMetric, catalyst.metrics._metric.ICallbackBatchMetric

This metric computes accuracy for multilabel classification case. It computes mean value of accuracy and it’s approximate std value (note that it’s not a real accuracy std but std of accuracy over batch mean values).

Parameters

• compute_on_call – if True, computes and returns metric value during metric call

• prefix – metric prefix

• suffix – metric suffix

• threshold – thresholds for model scores

AUCMetric

class catalyst.metrics._auc.AUCMetric(compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackLoaderMetric

AUC metric,

Parameters

• compute_on_call – if True, computes and returns metric value during metric call

• prefix – metric prefix

• suffix – metric suffix

Warning

This metric is under API improvement.

Classification – BinaryPrecisionRecallF1Metric

class catalyst.metrics._classification.BinaryPrecisionRecallF1Metric(zero_division: int = 0, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._classification.StatisticsMetric

Precision, recall, f1_score and support metrics for binary classification.

Parameters

• zero_division – value to set in case of zero division during metrics (precision, recall) computation; should be one of 0 or 1

• compute_on_call – if True, allows compute metric’s value on call

• prefix – metric prefix

• suffix – metric suffix

Classification – MulticlassPrecisionRecallF1SupportMetric

class catalyst.metrics._classification.MulticlassPrecisionRecallF1SupportMetric(num_classes: Optional[int] = None, zero_division: int = 0, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._classification.PrecisionRecallF1SupportMetric

Precision, recall, f1_score and support metrics for multiclass classification. Counts metrics with macro, micro and weighted average.

Parameters

• num_classes – number of classes in loader’s dataset

• zero_division – value to set in case of zero division during metrics (precision, recall) computation; should be one of 0 or 1

• compute_on_call – if True, allows compute metric’s value on call

• prefix – metric prefix

• suffix – metric suffix

Classification – MultilabelPrecisionRecallF1SupportMetric

class catalyst.metrics._classification.MultilabelPrecisionRecallF1SupportMetric(num_classes: Optional[int] = None, zero_division: int = 0, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._classification.PrecisionRecallF1SupportMetric

Precision, recall, f1_score and support metrics for multilabel classification. Counts metrics with macro, micro and weighted average.

Parameters

• num_classes – number of classes in loader’s dataset

• zero_division – value to set in case of zero division during metrics (precision, recall) computation; should be one of 0 or 1

• compute_on_call – if True, allows compute metric’s value on call

• prefix – metric prefix

• suffix – metric suffix

CMCMetric

class catalyst.metrics._cmc_score.CMCMetric(embeddings_key: str, labels_key: str, is_query_key: str, topk_args: Optional[Iterable[int]] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.AccumulationMetric

Cumulative Matching Characteristics

Parameters

• embeddings_key – key of embedding tensor in batch

• labels_key – key of label tensor in batch

• is_query_key – key of query flag tensor in batch

• topk_args – list of k, specifies which cmc@k should be calculated

• compute_on_call – if True, allows compute metric’s value on call

• prefix – metric prefix

• suffix – metric suffix

Examples

>>> from collections import OrderedDict
>>> from torch.optim import Adam
>>> from torch.utils.data import DataLoader
>>> from catalyst.contrib import nn
>>> from catalyst.contrib.datasets import MnistMLDataset, MnistQGDataset
>>> from catalyst.data import BalanceBatchSampler, HardTripletsSampler
>>> from catalyst.dl import ControlFlowCallback, LoaderMetricCallback, SupervisedRunner
>>> from catalyst.metrics import CMCMetric
>>>
>>> dataset_root = "."
>>>
>>> # download dataset for train and val, create loaders
>>> dataset_train = MnistMLDataset(root=dataset_root, download=True, transform=None)
>>> sampler = BalanceBatchSampler(labels=dataset_train.get_labels(), p=5, k=10)
>>> train_loader = DataLoader(
>>>     dataset=dataset_train, sampler=sampler, batch_size=sampler.batch_size
>>> )
>>> dataset_valid = MnistQGDataset(root=dataset_root, transform=None, gallery_fraq=0.2)
>>> valid_loader = DataLoader(dataset=dataset_valid, batch_size=1024)
>>>
>>> # model, optimizer, criterion
>>> model = nn.Sequential(nn.Flatten(), nn.Linear(28 * 28, 100))
>>> optimizer = Adam(model.parameters())
>>> sampler_inbatch = HardTripletsSampler(norm_required=False)
>>> criterion = nn.TripletMarginLossWithSampler(
>>>     margin=0.5, sampler_inbatch=sampler_inbatch
>>> )
>>>
>>> # batch data processing
>>> class CustomRunner(SupervisedRunner):
>>>     def handle_batch(self, batch):
>>>         if self.is_train_loader:
>>>             images, targets = batch["features"].float(), batch["targets"].long()
>>>             features = model(images)
>>>             self.batch = {
>>>                 "embeddings": features,
>>>                 "targets": targets,
>>>             }
>>>         else:
>>>             images, targets, is_query = (
>>>                 batch["features"].float(),
>>>                 batch["targets"].long(),
>>>                 batch["is_query"].bool(),
>>>             )
>>>             features = model(images)
>>>             self.batch = {
>>>                 "embeddings": features,
>>>                 "targets": targets,
>>>                 "is_query": is_query,
>>>             }
>>>
>>> # training
>>> runner = CustomRunner(input_key="features", output_key="embeddings")
>>> runner.train(
>>>     model=model,
>>>     criterion=criterion,
>>>     optimizer=optimizer,
>>>     callbacks=OrderedDict(
>>>         {
>>>             "cmc": ControlFlowCallback(
>>>                 LoaderMetricCallback(
>>>                     CMCMetric(
>>>                         embeddings_key="embeddings",
>>>                         labels_key="targets",
>>>                         is_query_key="is_query",
>>>                         topk_args=(1, 3)
>>>                     ),
>>>                     input_key=["embeddings", "is_query"],
>>>                     target_key=["targets"]
>>>                 ),
>>>                 loaders="valid",
>>>             ),
>>>         }
>>>     ),
>>>     loaders=OrderedDict({"train": train_loader, "valid": valid_loader}),
>>>     valid_loader="valid",
>>>     valid_metric="cmc01",
>>>     minimize_valid_metric=False,
>>>     logdir="./logs",
>>>     verbose=True,
>>>     num_epochs=3,
>>> )

ReidCMCMetric

class catalyst.metrics._cmc_score.ReidCMCMetric(embeddings_key: str, pids_key: str, cids_key: str, is_query_key: str, topk_args: Optional[Iterable[int]] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.AccumulationMetric

Cumulative Matching Characteristics for Reid case

Parameters

• embeddings_key – key of embedding tensor in batch

• pids_key – key of pids tensor in batch

• cids_key – key of cids tensor in batch

• is_query_key – key of query flag tensor in batch

• topk_args – list of k, specifies which cmc@k should be calculated

• compute_on_call – if True, allows compute metric’s value on call

• prefix – metric prefix

• suffix – metric suffix

RecSys – HitrateMetric

class catalyst.metrics._hitrate.HitrateMetric(topk_args: Optional[List[int]] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackBatchMetric

Calculates the hitrate.

Parameters

• topk_args – list of topk for hitrate@topk computing

• compute_on_call – if True, computes and returns metric value during metric call

• prefix – metric prefix

• suffix – metric suffix

Compute mean value of hitrate and it’s approximate std value.

RecSys – MAPMetric

class catalyst.metrics._map.MAPMetric(topk_args: Optional[List[int]] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackBatchMetric

Calculates the Mean Average Precision (MAP) for RecSys. The precision metric summarizes the fraction of relevant items out of the whole the recommendation list.

Parameters

• topk_args – list of topk for map@topk computing

• compute_on_call – if True, computes and returns metric value during metric call

• prefix – metric prefix

• suffix – metric suffix

It computes mean value of map and it’s approximate std value

RecSys – MRRMetric

class catalyst.metrics._mrr.MRRMetric(topk_args: Optional[List[int]] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackBatchMetric

Calculate the Mean Reciprocal Rank (MRR) score given model outputs and targets The precision metric summarizes the fraction of relevant items

Parameters

• topk_args – list of topk for mrr@topk computing

• compute_on_call – if True, computes and returns metric value during metric call

• prefix – metric prefix

• suffix – metric suffix

Compute mean value of map and it’s approximate std value

RecSys – NDCGMetric

class catalyst.metrics._ndcg.NDCGMetric(topk_args: Optional[List[int]] = None, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackBatchMetric

Calculate the Normalized discounted cumulative gain (NDCG) score given model outputs and targets The precision metric summarizes the fraction of relevant items

Parameters

• topk_args – list of topk for ndcg@topk computing

• compute_on_call – if True, computes and returns metric value during metric call

• prefix – metric prefix

• suffix – metric suffix

Compute mean value of ndcg and it’s approximate std value

Segmentation – RegionBasedMetric

class catalyst.metrics._segmentation.RegionBasedMetric(metric_fn: Callable, metric_name: str, class_dim: int = 1, weights: Optional[List[float]] = None, class_names: Optional[List[str]] = None, threshold: Optional[float] = 0.5, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._metric.ICallbackBatchMetric

Logic class for all region based metrics, like IoU, Dice, Trevsky.

Parameters

• metric_fn – metric function, that get statistics and return score

• metric_name – name of the metric

• class_dim – indicates class dimension (K) for outputs and targets tensors (default = 1)

• weights – class weights

• class_names – class names

• threshold – threshold for outputs binarization

• compute_on_call – Computes and returns metric value during metric call. Used for per-batch logging. default: True

• prefix – metric prefix

• suffix – metric suffix

Segmentation – DiceMetric

class catalyst.metrics._segmentation.DiceMetric(class_dim: int = 1, weights: Optional[List[float]] = None, class_names: Optional[List[str]] = None, threshold: Optional[float] = None, eps: float = 1e-07, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._segmentation.RegionBasedMetric

Dice Metric, dice score = 2 * intersection / (intersection + union)) = 2 * tp / (2 * tp + fp + fn)

Parameters

• class_dim – indicates class dimention (K) for outputs and

• tensors (targets) –

• weights – class weights

• class_names – class names

• threshold – threshold for outputs binarization

• eps – epsilon to avoid zero division

• compute_on_call – Computes and returns metric value during metric call. Used for per-batch logging. default: True

• prefix – metric prefix

• suffix – metric suffix

Segmentation – IOUMetric

class catalyst.metrics._segmentation.IOUMetric(class_dim: int = 1, weights: Optional[List[float]] = None, class_names: Optional[List[str]] = None, threshold: Optional[float] = None, eps: float = 1e-07, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._segmentation.RegionBasedMetric

IoU Metric, iou score = intersection / union = tp / (tp + fp + fn).

Parameters

• class_dim – indicates class dimension (K) for outputs and targets tensors (default = 1)

• weights – class weights

• class_names – class names

• threshold – threshold for outputs binarization

• eps – epsilon to avoid zero division

• compute_on_call – Computes and returns metric value during metric call. Used for per-batch logging. default: True

• prefix – metric prefix

• suffix – metric suffix

Segmentation – TrevskyMetric

class catalyst.metrics._segmentation.TrevskyMetric(alpha: float, beta: Optional[float] = None, class_dim: int = 1, weights: Optional[List[float]] = None, class_names: Optional[List[str]] = None, threshold: Optional[float] = None, eps: float = 1e-07, compute_on_call: bool = True, prefix: Optional[str] = None, suffix: Optional[str] = None)

Bases: catalyst.metrics._segmentation.RegionBasedMetric

Trevsky Metric, trevsky score = tp / (tp + fp * beta + fn * alpha)

Parameters

• alpha – false negative coefficient, bigger alpha bigger penalty for false negative. if beta is None, alpha must be in (0, 1)

• beta – false positive coefficient, bigger alpha bigger penalty for false positive. Must be in (0, 1), if None beta = (1 - alpha)

• class_dim – indicates class dimension (K) for outputs and targets tensors (default = 1)

• weights – class weights

• class_names – class names

• threshold – threshold for outputs binarization

• eps – epsilon to avoid zero division

• compute_on_call – Computes and returns metric value during metric call. Used for per-batch logging. default: True

• prefix – metric prefix

• suffix – metric suffix

Functional API

Accuracy

catalyst.metrics.functional._accuracy.accuracy(outputs: torch.Tensor, targets: torch.Tensor, topk: Sequence[int] = (1)) → Sequence[torch.Tensor]

Computes multiclass accuracy@topk for the specified values of topk.

Parameters

• outputs – model outputs, logits with shape [bs; num_classes]

• targets – ground truth, labels with shape [bs; 1]

• topk – topk for accuracy@topk computing

Returns

list with computed accuracy@topk

Example

>>> accuracy(
>>>     outputs=torch.tensor([
>>>         [1, 0, 0],
>>>         [0, 1, 0],
>>>         [0, 0, 1],
>>>     ]),
>>>     targets=torch.tensor([0, 1, 2]),
>>> )
[tensor([1.])]
>>> accuracy(
>>>     outputs=torch.tensor([
>>>         [1, 0, 0],
>>>         [0, 1, 0],
>>>         [0, 1, 0],
>>>     ]),
>>>     targets=torch.tensor([0, 1, 2]),
>>> )
[tensor([0.6667])]
>>> accuracy(
>>>     outputs=torch.tensor([
>>>         [1, 0, 0],
>>>         [0, 1, 0],
>>>         [0, 0, 1],
>>>     ]),
>>>     targets=torch.tensor([0, 1, 2]),
>>>     topk=[1, 3],
>>> )
[tensor([1.]), tensor([1.])]
>>> accuracy(
>>>     outputs=torch.tensor([
>>>         [1, 0, 0],
>>>         [0, 1, 0],
>>>         [0, 1, 0],
>>>     ]),
>>>     targets=torch.tensor([0, 1, 2]),
>>>     topk=[1, 3],
>>> )
[tensor([0.6667]), tensor([1.])]

catalyst.metrics.functional._accuracy.multilabel_accuracy(outputs: torch.Tensor, targets: torch.Tensor, threshold: Union[float, torch.Tensor]) → torch.Tensor

Computes multilabel accuracy for the specified activation and threshold.

Parameters

• outputs – NxK tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary NxK tensort that encodes which of the K classes are associated with the N-th input (eg: a row [0, 1, 0, 1] indicates that the example is associated with classes 2 and 4)

• threshold – threshold for for model output

Returns

computed multilabel accuracy

Example

>>> multilabel_accuracy(
>>>     outputs=torch.tensor([
>>>         [1, 0],
>>>         [0, 1],
>>>     ]),
>>>     targets=torch.tensor([
>>>         [1, 0],
>>>         [0, 1],
>>>     ]),
>>>     threshold=0.5,
>>> )
tensor([1.])
>>> multilabel_accuracy(
>>>     outputs=torch.tensor([
>>>         [1.0, 0.0],
>>>         [0.6, 1.0],
>>>     ]),
>>>     targets=torch.tensor([
>>>         [1, 0],
>>>         [0, 1],
>>>     ]),
>>>     threshold=0.5,
>>> )
tensor(0.7500)
>>> multilabel_accuracy(
>>>     outputs=torch.tensor([
>>>         [1.0, 0.0],
>>>         [0.4, 1.0],
>>>     ]),
>>>     targets=torch.tensor([
>>>         [1, 0],
>>>         [0, 1],
>>>     ]),
>>>     threshold=0.5,
>>> )
tensor(1.0)

AUC

catalyst.metrics.functional._auc.auc(scores: torch.Tensor, targets: torch.Tensor) → torch.Tensor

Computes ROC-AUC.

Parameters

• scores – NxK tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary NxK tensort that encodes which of the K classes are associated with the N-th input (eg: a row [0, 1, 0, 1] indicates that the example is associated with classes 2 and 4)

Returns

Tensor with [num_classes] shape of per-class-aucs

Return type

torch.Tensor

Example

>>> auc(
>>>     scores=torch.tensor([
>>>         [0.9, 0.1],
>>>         [0.1, 0.9],
>>>     ]),
>>>     targets=torch.tensor([
>>>         [1, 0],
>>>         [0, 1],
>>>     ]),
>>> )
tensor([1., 1.])
>>> auc(
>>>     scores=torch.tensor([
>>>         [0.9],
>>>         [0.8],
>>>         [0.7],
>>>         [0.6],
>>>         [0.5],
>>>         [0.4],
>>>         [0.3],
>>>         [0.2],
>>>         [0.1],
>>>         [0.0],
>>>     ]),
>>>     targets=torch.tensor([
>>>         [0],
>>>         [1],
>>>         [1],
>>>         [1],
>>>         [1],
>>>         [1],
>>>         [1],
>>>         [0],
>>>         [0],
>>>         [0],
>>>     ]),
>>> )
tensor([0.7500])

Warning

This metric is under API improvement.

catalyst.metrics.functional._auc.binary_auc(scores: torch.Tensor, targets: torch.Tensor) → Tuple[float, numpy.ndarray, numpy.ndarray]

Binary AUC computation.

Parameters

• scores – estimated scores from a model.

• targets – ground truth (correct) target values.

Returns

measured roc-auc, true positive rate, false positive rate

Return type

Tuple[float, np.ndarray, np.ndarray]

Warning

This metric is under API improvement.

Average Precision

catalyst.metrics.functional._average_precision.average_precision(outputs: torch.Tensor, targets: torch.Tensor, k: int) → torch.Tensor

Calculate the Average Precision for RecSys. The precision metric summarizes the fraction of relevant items out of the whole the recommendation list.

To compute the precision at k set the threshold rank k, compute the percentage of relevant items in topK, ignoring the documents ranked lower than k.

The average precision at k (AP at k) summarizes the average precision for relevant items up to the k-th one. Wikipedia entry for the Average precision

<https://en.wikipedia.org/w/index.php?title=Information_retrieval& oldid=793358396#Average_precision>

If a relevant document never gets retrieved, we assume the precision corresponding to that relevant doc to be zero

Parameters

• outputs (torch.Tensor) – Tensor with predicted score size: [batch_size, slate_length] model outputs, logits

• targets (torch.Tensor) – Binary tensor with ground truth. 1 means the item is relevant and 0 not relevant size: [batch_szie, slate_length] ground truth, labels

• k – Parameter for evaluation on top-k items

Returns

The map score for each batch. size: [batch_size, 1]

Return type

ap_score (torch.Tensor)

Examples

>>> average_precision(
>>>   outputs=torch.tensor([
>>>     [9, 8, 7, 6, 5, 4, 3, 2, 1, 0],
>>>     [9, 8, 7, 6, 5, 4, 3, 2, 1, 0],
>>>   ]),
>>>   targets=torch.tensor([
>>>     [1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0, 1.0],
>>>     [0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 0.0, 0.0, 0.0],
>>>   ]),
>>> )
tensor([0.6222, 0.4429])

catalyst.metrics.functional._average_precision.binary_average_precision(outputs: torch.Tensor, targets: torch.Tensor, weights: Optional[torch.Tensor] = None) → torch.Tensor

Computes the average precision.

Parameters

• outputs – NxK tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary NxK tensort that encodes which of the K classes are associated with the N-th input (eg: a row [0, 1, 0, 1] indicates that the example is associated with classes 2 and 4)

• weights – importance for each sample

Returns

tensor of [K; ] shape, with average precision for K classes

Return type

torch.Tensor

Examples

>>> binary_average_precision(
>>>     outputs=torch.Tensor([0.1, 0.4, 0.35, 0.8]),
>>>     targets=torch.Tensor([0, 0, 1, 1]),
>>> )
tensor([0.8333])

catalyst.metrics.functional._average_precision.mean_average_precision(outputs: torch.Tensor, targets: torch.Tensor, topk: List[int]) → List[torch.Tensor]

Calculate the mean average precision (MAP) for RecSys. The metrics calculate the mean of the AP across all batches

MAP amplifies the interest in finding many relevant items for each query

Parameters

• outputs (torch.Tensor) – Tensor with predicted score size: [batch_size, slate_length] model outputs, logits

• targets (torch.Tensor) – Binary tensor with ground truth. 1 means the item is relevant and 0 not relevant size: [batch_szie, slate_length] ground truth, labels

• topk (List[int]) – List of parameter for evaluation topK items

Returns

The map score for every k. size: len(top_k)

Return type

map_at_k (Tuple[float])

Examples

>>> mean_average_precision(
>>>   outputs=torch.tensor([
>>>     [9, 8, 7, 6, 5, 4, 3, 2, 1, 0],
>>>     [9, 8, 7, 6, 5, 4, 3, 2, 1, 0],
>>>   ]),
>>>   targets=torch.tensor([
>>>     [1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0, 1.0],
>>>     [0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 0.0, 0.0, 0.0],
>>>   ]),
>>>   topk=[10],
>>> )
[tensor(0.5325)]

Classification

catalyst.metrics.functional._classification.f1score(precision_value, recall_value, eps=1e-05)

Calculating F1-score from precision and recall to reduce computation redundancy.

Parameters

• precision_value – precision (0-1)

• recall_value – recall (0-1)

• eps – epsilon to use

Returns

F1 score (0-1)

catalyst.metrics.functional._classification.get_aggregated_metrics(tp: numpy.array, fp: numpy.array, fn: numpy.array, support: numpy.array, zero_division: int = 0) → Tuple[numpy.array, numpy.array, numpy.array, numpy.array]

Count precision, recall, f1 scores per-class and with macro, weighted and micro average with statistics.

Parameters

• tp – array of shape (num_classes, ) of true positive statistics per class

• fp – array of shape (num_classes, ) of false positive statistics per class

• fn – array of shape (num_classes, ) of false negative statistics per class

• support – array of shape (num_classes, ) of samples count per class

• zero_division – int value, should be one of 0 or 1; used for precision and recall computation

Returns

per-class, micro, macro, weighted averaging

Return type

arrays of metrics

catalyst.metrics.functional._classification.get_binary_metrics(tp: int, fp: int, fn: int, zero_division: int) → Tuple[float, float, float]

Get precision, recall, f1 score metrics from true positive, false positive,

false negative statistics for binary classification

Parameters

• tp – true positive

• fp – false positive

• fn – false negative

• zero_division – int value, should be 0 or 1

Returns

precision, recall, f1 scores

catalyst.metrics.functional._classification.precision(tp: int, fp: int, zero_division: int = 0) → float

Calculates precision (a.k.a. positive predictive value) for binary classification and segmentation.

Parameters

• tp – number of true positives

• fp – number of false positives

• zero_division – int value, should be one of 0 or 1; if both tp==0 and fp==0 return this value as s result

Returns

precision value (0-1)

catalyst.metrics.functional._classification.precision_recall_fbeta_support(outputs: torch.Tensor, targets: torch.Tensor, beta: float = 1, eps: float = 1e-06, argmax_dim: int = - 1, num_classes: Optional[int] = None, zero_division: int = 0) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

Counts precision_val, recall, fbeta_score.

Parameters

• outputs – A list of predicted elements

• targets – A list of elements that are to be predicted

• beta – beta param for f_score

• eps – epsilon to avoid zero division

• argmax_dim – int, that specifies dimension for argmax transformation in case of scores/probabilities in outputs

• num_classes – int, that specifies number of classes if it known.

• zero_division – int value, should be one of 0 or 1; used for precision_val and recall computation

Returns

tuple of precision_val, recall, fbeta_score

Examples

>>> precision_recall_fbeta_support(
>>>     outputs=torch.tensor([
>>>         [1, 0, 0],
>>>         [0, 1, 0],
>>>         [0, 0, 1],
>>>     ]),
>>>     targets=torch.tensor([0, 1, 2]),
>>>     beta=1,
>>> )
(
    tensor([1., 1., 1.]),  # precision_val per class
    tensor([1., 1., 1.]),  # recall per class
    tensor([1., 1., 1.]),  # fbeta per class
    tensor([1., 1., 1.]),  # support per class
)
>>> precision_recall_fbeta_support(
>>>     outputs=torch.tensor([[0, 0, 1, 1, 0, 1, 0, 1]]),
>>>     targets=torch.tensor([[0, 1, 0, 1, 0, 0, 1, 1]]),
>>>     beta=1,
>>> )
(
    tensor([0.5000, 0.5000]),  # precision per class
    tensor([0.5000, 0.5000]),  # recall per class
    tensor([0.5000, 0.5000]),  # fbeta per class
    tensor([4., 4.]),          # support per class
)

catalyst.metrics.functional._classification.recall(tp: int, fn: int, zero_division: int = 0) → float

Calculates recall (a.k.a. true positive rate) for binary classification and segmentation.

Parameters

• tp – number of true positives

• fn – number of false negatives

• zero_division – int value, should be one of 0 or 1; if both tp==0 and fn==0 return this value as s result

Returns

recall value (0-1)

CMC Score

catalyst.metrics.functional._cmc_score.cmc_score(query_embeddings: torch.Tensor, gallery_embeddings: torch.Tensor, conformity_matrix: torch.Tensor, topk: int = 1) → float

Function to count CMC score from query and gallery embeddings.

Parameters

• query_embeddings – tensor shape of (n_embeddings, embedding_dim) embeddings of the objects in query

• gallery_embeddings – tensor shape of (n_embeddings, embedding_dim) embeddings of the objects in gallery

• conformity_matrix – binary matrix with 1 on same label pos and 0 otherwise

• topk – number of top examples for cumulative score counting

Returns

cmc score

catalyst.metrics.functional._cmc_score.cmc_score_count(distances: torch.Tensor, conformity_matrix: torch.Tensor, topk: int = 1) → float

Function to count CMC from distance matrix and conformity matrix.

Parameters

• distances – distance matrix shape of (n_embeddings_x, n_embeddings_y)

• conformity_matrix – binary matrix with 1 on same label pos and 0 otherwise

• topk – number of top examples for cumulative score counting

Returns

cmc score

catalyst.metrics.functional._cmc_score.masked_cmc_score(query_embeddings: torch.Tensor, gallery_embeddings: torch.Tensor, conformity_matrix: torch.Tensor, available_samples: torch.Tensor, topk: int = 1) → float

Parameters

• query_embeddings – tensor shape of (n_embeddings, embedding_dim) embeddings of the objects in query

• gallery_embeddings – tensor shape of (n_embeddings, embedding_dim) embeddings of the objects in gallery

• conformity_matrix – binary matrix with 1 on same label pos and 0 otherwise

• available_samples – tensor of shape (query_size, gallery_size), available_samples[i][j] == 1 means that j-th element of gallery should be used while scoring i-th query one

• topk – number of top examples for cumulative score counting

Returns

cmc score with mask

Raises

ValueError – if there are items that have different labels and are unavailable for each other according to availability matrix

F1 score

catalyst.metrics.functional._f1_score.f1_score(outputs: torch.Tensor, targets: torch.Tensor, eps: float = 1e-07, argmax_dim: int = - 1, num_classes: Optional[int] = None) → Union[float, torch.Tensor]

Fbeta_score with beta=1.

Parameters

• outputs – A list of predicted elements

• targets – A list of elements that are to be predicted

• eps – epsilon to avoid zero division

• argmax_dim – int, that specifies dimension for argmax transformation in case of scores/probabilities in outputs

• num_classes – int, that specifies number of classes if it known

Returns

F_1 score

Return type

float

catalyst.metrics.functional._f1_score.fbeta_score(outputs: torch.Tensor, targets: torch.Tensor, beta: float = 1.0, eps: float = 1e-07, argmax_dim: int = - 1, num_classes: Optional[int] = None) → Union[float, torch.Tensor]

Counts fbeta score for given outputs and targets.

Parameters

• outputs – A list of predicted elements

• targets – A list of elements that are to be predicted

• beta – beta param for f_score

• eps – epsilon to avoid zero division

• argmax_dim – int, that specifies dimension for argmax transformation in case of scores/probabilities in outputs

• num_classes – int, that specifies number of classes if it known

Raises

ValueError – If beta is a negative number.

Returns

F_1 score.

Return type

float

Focal

catalyst.metrics.functional._focal.reduced_focal_loss(outputs: torch.Tensor, targets: torch.Tensor, threshold: float = 0.5, gamma: float = 2.0, reduction='mean') → torch.Tensor

Compute reduced focal loss between target and output logits.

It has been proposed in Reduced Focal Loss: 1st Place Solution to xView object detection in Satellite Imagery paper.

Note

size_average and reduce params are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction.

Source: https://github.com/BloodAxe/pytorch-toolbelt

Parameters

• outputs – tensor of arbitrary shape

• targets – tensor of the same shape as input

• threshold – threshold for focal reduction

• gamma – gamma for focal reduction

• reduction – specifies the reduction to apply to the output: "none" | "mean" | "sum" | "batchwise_mean". "none": no reduction will be applied, "mean": the sum of the output will be divided by the number of elements in the output, "sum": the output will be summed. "batchwise_mean" computes mean loss per sample in batch. Default: “mean”

Returns: # noqa: DAR201

torch.Tensor: computed loss

catalyst.metrics.functional._focal.sigmoid_focal_loss(outputs: torch.Tensor, targets: torch.Tensor, gamma: float = 2.0, alpha: float = 0.25, reduction: str = 'mean')

Compute binary focal loss between target and output logits.

Parameters

• outputs – tensor of arbitrary shape

• targets – tensor of the same shape as input

• gamma – gamma for focal loss

• alpha – alpha for focal loss

• reduction (string, optional) – specifies the reduction to apply to the output: "none" | "mean" | "sum" | "batchwise_mean". "none": no reduction will be applied, "mean": the sum of the output will be divided by the number of elements in the output, "sum": the output will be summed.

Returns

computed loss

Source: https://github.com/BloodAxe/pytorch-toolbelt

Hitrate

catalyst.metrics.functional._hitrate.hitrate(outputs: torch.Tensor, targets: torch.Tensor, topk: List[int], zero_division: int = 0) → List[torch.Tensor]

Calculate the hit rate (aka recall) score given model outputs and targets. Hit-rate is a metric for evaluating ranking systems. Generate top-N recommendations and if one of the recommendation is actually what user has rated, you consider that a hit. By rate we mean any explicit form of user’s interactions. Add up all of the hits for all users and then divide by number of users

Compute top-N recommendation for each user in the training stage and intentionally remove one of this items fro the training data.

Parameters

• outputs (torch.Tensor) – Tensor with predicted score size: [batch_size, slate_length] model outputs, logits

• targets (torch.Tensor) – Binary tensor with ground truth. 1 means the item is relevant for the user and 0 not relevant size: [batch_size, slate_length] ground truth, labels

• topk (List[int]) – Parameter fro evaluation on top-k items

• zero_division (int) – value, returns in the case of the divison by zero should be one of 0 or 1

Returns

the hitrate score

Return type

hitrate_at_k (List[torch.Tensor])

MRR

catalyst.metrics.functional._mrr.mrr(outputs: torch.Tensor, targets: torch.Tensor, topk: List[int]) → List[torch.Tensor]

Calculate the Mean Reciprocal Rank (MRR) score given model outputs and targets Data aggregated in batches.

The MRR@k is the mean overall batch of the reciprocal rank, that is the rank of the highest ranked relevant item, if any in the top k, 0 otherwise. https://en.wikipedia.org/wiki/Mean_reciprocal_rank

Parameters

• outputs – Tensor weith predicted score size: [batch_size, slate_length] model outputs, logits

• targets – Binary tensor with ground truth. 1 means the item is relevant and 0 if it’s not relevant size: [batch_szie, slate_length] ground truth, labels

• topk – Parameter fro evaluation on top-k items

Returns

MRR score

Examples

>>> mrr(
>>>     outputs=torch.Tensor([
>>>         [4.0, 2.0, 3.0, 1.0],
>>>         [1.0, 2.0, 3.0, 4.0],
>>>     ]),
>>>     targets=torch.Tensor([
>>>         [0, 0, 1.0, 1.0],
>>>         [0, 0, 1.0, 1.0],
>>>     ]),
>>>     topk=[1, 3],
>>> )
[tensor(0.5000), tensor(0.7500)]

catalyst.metrics.functional._mrr.reciprocal_rank(outputs: torch.Tensor, targets: torch.Tensor, k: int) → torch.Tensor

Calculate the Reciprocal Rank (MRR) score given model outputs and targets Data aggregated in batches.

Parameters

• outputs – Tensor weith predicted score size: [batch_size, slate_length] model outputs, logits

• targets – Binary tensor with ground truth. 1 means the item is relevant and 0 if it’s not relevant size: [batch_size, slate_length] ground truth, labels

• k – Parameter for evaluation on top-k items

Returns

MRR score

Examples

>>> reciprocal_rank(
>>>     outputs=torch.Tensor([
>>>         [4.0, 2.0, 3.0, 1.0],
>>>         [1.0, 2.0, 3.0, 4.0],
>>>     ]),
>>>     targets=torch.Tensor([
>>>         [0, 0, 1.0, 1.0],
>>>         [0, 0, 1.0, 1.0],
>>>     ]),
>>>     k=1,
>>> )
tensor([[0.], [1.]])
>>> reciprocal_rank(
>>>     outputs=torch.Tensor([
>>>         [4.0, 2.0, 3.0, 1.0],
>>>         [1.0, 2.0, 3.0, 4.0],
>>>     ]),
>>>     targets=torch.Tensor([
>>>         [0, 0, 1.0, 1.0],
>>>         [0, 0, 1.0, 1.0],
>>>     ]),
>>>     k=3,
>>> )
tensor([[0.5000], [1.0000]])

NDCG

catalyst.metrics.functional._ndcg.dcg(outputs: torch.Tensor, targets: torch.Tensor, gain_function='exp_rank') → torch.Tensor

Computes Discounted cumulative gain (DCG) DCG@topk for the specified values of k. Graded relevance as a measure of usefulness, or gain, from examining a set of items. Gain may be reduced at lower ranks. Reference: https://en.wikipedia.org/wiki/Discounted_cumulative_gain

Parameters

• outputs – model outputs, logits with shape [batch_size; slate_length]

• targets – ground truth, labels with shape [batch_size; slate_length]

• gain_function – String indicates the gain function for the ground truth labels. Two options available: - exp_rank: torch.pow(2, x) - 1 - linear_rank: x On the default, exp_rank is used to emphasize on retrieving the relevant documents.

Returns

The discounted gains tensor

Return type

dcg_score (torch.Tensor)

Raises

ValueError – gain function can be either pow_rank or rank

Examples

>>> dcg(
>>>     outputs = torch.tensor([
>>>         [3, 2, 1, 0],
>>>     ]),
>>>     targets = torch.Tensor([
>>>         [2.0, 2.0, 1.0, 0.0],
>>>     ]),
>>>     gain_function="linear_rank",
>>> )
tensor([[2.0000, 2.0000, 0.6309, 0.0000]])
>>> dcg(
>>>     outputs = torch.tensor([
>>>         [3, 2, 1, 0],
>>>     ]),
>>>     targets = torch.Tensor([
>>>         [2.0, 2.0, 1.0, 0.0],
>>>     ]),
>>>     gain_function="linear_rank",
>>> ).sum()
tensor(4.6309)
>>> dcg(
>>>     outputs = torch.tensor([
>>>         [3, 2, 1, 0],
>>>     ]),
>>>     targets = torch.Tensor([
>>>         [2.0, 2.0, 1.0, 0.0],
>>>     ]),
>>>     gain_function="exp_rank",
>>> )
tensor([[3.0000, 1.8928, 0.5000, 0.0000]])
>>> dcg(
>>>     outputs = torch.tensor([
>>>         [3, 2, 1, 0],
>>>     ]),
>>>     targets = torch.Tensor([
>>>         [2.0, 2.0, 1.0, 0.0],
>>>     ]),
>>>     gain_function="exp_rank",
>>> ).sum()
tensor(5.3928)

catalyst.metrics.functional._ndcg.ndcg(outputs: torch.Tensor, targets: torch.Tensor, topk: List[int], gain_function='exp_rank') → List[torch.Tensor]

Computes nDCG@topk for the specified values of topk.

Parameters

• outputs (torch.Tensor) – model outputs, logits with shape [batch_size; slate_size]

• targets (torch.Tensor) – ground truth, labels with shape [batch_size; slate_size]

• gain_function – callable, gain function for the ground truth labels. Two options available: - exp_rank: torch.pow(2, x) - 1 - linear_rank: x On the default, exp_rank is used to emphasize on retrieving the relevant documents.

• topk (List[int]) – Parameter fro evaluation on top-k items

Returns

tuple with computed ndcg@topk

Return type

results (Tuple[float])

Examples

>>> ndcg(
>>>     outputs = torch.tensor([
>>>         [0.5, 0.2, 0.1],
>>>         [0.5, 0.2, 0.1],
>>>     ]),
>>>     targets = torch.Tensor([
>>>         [1.0, 0.0, 1.0],
>>>         [1.0, 0.0, 1.0],
>>>     ]),
>>>     topk=[2],
>>>     gain_function="exp_rank",
>>> )
[tensor(0.6131)]
>>> ndcg(
>>>     outputs = torch.tensor([
>>>         [0.5, 0.2, 0.1],
>>>         [0.5, 0.2, 0.1],
>>>     ]),
>>>     targets = torch.Tensor([
>>>         [1.0, 0.0, 1.0],
>>>         [1.0, 0.0, 1.0],
>>>     ]),
>>>     topk=[2],
>>>     gain_function="exp_rank",
>>> )
[tensor(0.5000)]

Precision

catalyst.metrics.functional._precision.precision(outputs: torch.Tensor, targets: torch.Tensor, argmax_dim: int = - 1, eps: float = 1e-07, num_classes: Optional[int] = None) → Union[float, torch.Tensor]

Multiclass precision score.

Parameters

• outputs – estimated targets as predicted by a model with shape [bs; …, (num_classes or 1)]

• targets – ground truth (correct) target values with shape [bs; …, 1]

• argmax_dim – int, that specifies dimension for argmax transformation in case of scores/probabilities in outputs

• eps – float. Epsilon to avoid zero division.

• num_classes – int, that specifies number of classes if it known

Returns

precision for every class

Return type

Tensor

Examples

>>> precision(
>>>     outputs=torch.tensor([
>>>         [1, 0, 0],
>>>         [0, 1, 0],
>>>         [0, 0, 1],
>>>     ]),
>>>     targets=torch.tensor([0, 1, 2]),
>>>     beta=1,
>>> )
tensor([1., 1., 1.])
>>> precision(
>>>     outputs=torch.tensor([[0, 0, 1, 1, 0, 1, 0, 1]]),
>>>     targets=torch.tensor([[0, 1, 0, 1, 0, 0, 1, 1]]),
>>> )
tensor([0.5000, 0.5000]

Recall

catalyst.metrics.functional._recall.recall(outputs: torch.Tensor, targets: torch.Tensor, argmax_dim: int = - 1, eps: float = 1e-07, num_classes: Optional[int] = None) → Union[float, torch.Tensor]

Multiclass recall score.

Parameters

• outputs – estimated targets as predicted by a model with shape [bs; …, (num_classes or 1)]

• targets – ground truth (correct) target values with shape [bs; …, 1]

• argmax_dim – int, that specifies dimension for argmax transformation in case of scores/probabilities in outputs

• eps – float. Epsilon to avoid zero division.

• num_classes – int, that specifies number of classes if it known

Returns

recall for every class

Return type

Tensor

Examples

>>> recall(
>>>     outputs=torch.tensor([
>>>         [1, 0, 0],
>>>         [0, 1, 0],
>>>         [0, 0, 1],
>>>     ]),
>>>     targets=torch.tensor([0, 1, 2]),
>>> )
tensor([1., 1., 1.])
>>> precision_recall_fbeta_support(
>>>     outputs=torch.tensor([[0, 0, 1, 1, 0, 1, 0, 1]]),
>>>     targets=torch.tensor([[0, 1, 0, 1, 0, 0, 1, 1]]),
>>> )
tensor([0.5000, 0.5000])

Segmentation

catalyst.metrics.functional._segmentation.dice(outputs: torch.Tensor, targets: torch.Tensor, class_dim: int = 1, threshold: Optional[float] = None, mode: str = 'per-class', weights: Optional[List[float]] = None, eps: float = 1e-07) → torch.Tensor

Computes the dice score, dice score = 2 * intersection / (intersection + union)) = = 2 * tp / (2 * tp + fp + fn)

Parameters

• outputs – [N; K; …] tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary [N; K; …] tensor that encodes which of the K classes are associated with the N-th input

• class_dim – indicates class dimention (K) for outputs and targets tensors (default = 1), if mode = “micro” means nothing

• threshold – threshold for outputs binarization

• mode – class summation strategy. Must be one of [‘micro’, ‘macro’, ‘weighted’, ‘per-class’]. If mode=’micro’, classes are ignored, and metric are calculated generally. If mode=’macro’, metric are calculated per-class and than are averaged over all classes. If mode=’weighted’, metric are calculated per-class and than summed over all classes with weights. If mode=’per-class’, metric are calculated separately for all classes

• weights – class weights(for mode=”weighted”)

• eps – epsilon to avoid zero division

Returns

Dice score for each class(if mode=’weighted’) or aggregated Dice

Example

>>> size = 4
>>> half_size = size // 2
>>> shape = (1, 1, size, size)
>>> empty = torch.zeros(shape)
>>> full = torch.ones(shape)
>>> left = torch.ones(shape)
>>> left[:, :, :, half_size:] = 0
>>> right = torch.ones(shape)
>>> right[:, :, :, :half_size] = 0
>>> top_left = torch.zeros(shape)
>>> top_left[:, :, :half_size, :half_size] = 1
>>> pred = torch.cat([empty, left, empty, full, left, top_left], dim=1)
>>> targets = torch.cat([full, right, empty, full, left, left], dim=1)
>>> dice(
>>>      outputs=pred,
>>>      targets=targets,
>>>      class_dim=1,
>>>      threshold=0.5,
>>>      mode="per-class"
>>> )
tensor([0.0000, 0.0000, 1.0000, 1.0000, 1.0000, 0.6667])

catalyst.metrics.functional._segmentation.get_segmentation_statistics(outputs: torch.Tensor, targets: torch.Tensor, class_dim: int = 1, threshold: Optional[float] = None) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor]

Computes true positive, false positive, false negative for a multilabel segmentation problem.

Parameters

• outputs – [N; K; …] tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary [N; K; …] tensor that encodes which of the K classes are associated with the N-th input

• class_dim – indicates class dimention (K) for outputs and targets tensors (default = 1)

• threshold – threshold for outputs binarization

Returns

Segmentation stats

Example

>>> size = 4
>>> half_size = size // 2
>>> shape = (1, 1, size, size)
>>> empty = torch.zeros(shape)
>>> full = torch.ones(shape)
>>> left = torch.ones(shape)
>>> left[:, :, :, half_size:] = 0
>>> right = torch.ones(shape)
>>> right[:, :, :, :half_size] = 0
>>> top_left = torch.zeros(shape)
>>> top_left[:, :, :half_size, :half_size] = 1
>>> pred = torch.cat([empty, left, empty, full, left, top_left], dim=1)
>>> targets = torch.cat([full, right, empty, full, left, left], dim=1)
>>> get_segmentation_statistics(
>>>     outputs=pred,
>>>     targets=targets,
>>>     class_dim=1,
>>>     threshold=0.5,
>>> )
(tensor([ 0.,  0.,  0., 16.,  8.,  4.]),
tensor([0., 8., 0., 0., 0., 0.]),
tensor([16.,  8.,  0.,  0.,  0.,  4.]))

catalyst.metrics.functional._segmentation.iou(outputs: torch.Tensor, targets: torch.Tensor, class_dim: int = 1, threshold: Optional[float] = None, mode: str = 'per-class', weights: Optional[List[float]] = None, eps: float = 1e-07) → torch.Tensor

Computes the iou/jaccard score, iou score = intersection / union = tp / (tp + fp + fn)

Parameters

• outputs – [N; K; …] tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary [N; K; …] tensor that encodes which of the K classes are associated with the N-th input

• class_dim – indicates class dimention (K) for outputs and targets tensors (default = 1), if mode = “micro” means nothing

• threshold – threshold for outputs binarization

• mode – class summation strategy. Must be one of [‘micro’, ‘macro’, ‘weighted’, ‘per-class’]. If mode=’micro’, classes are ignored, and metric are calculated generally. If mode=’macro’, metric are calculated per-class and than are averaged over all classes. If mode=’weighted’, metric are calculated per-class and than summed over all classes with weights. If mode=’per-class’, metric are calculated separately for all classes

• weights – class weights(for mode=”weighted”)

• eps – epsilon to avoid zero division

Returns

IoU (Jaccard) score for each class(if mode=’weighted’) or aggregated IOU

Example

>>> size = 4
>>> half_size = size // 2
>>> shape = (1, 1, size, size)
>>> empty = torch.zeros(shape)
>>> full = torch.ones(shape)
>>> left = torch.ones(shape)
>>> left[:, :, :, half_size:] = 0
>>> right = torch.ones(shape)
>>> right[:, :, :, :half_size] = 0
>>> top_left = torch.zeros(shape)
>>> top_left[:, :, :half_size, :half_size] = 1
>>> pred = torch.cat([empty, left, empty, full, left, top_left], dim=1)
>>> targets = torch.cat([full, right, empty, full, left, left], dim=1)
>>> iou(
>>>     outputs=pred,
>>>     targets=targets,
>>>     class_dim=1,
>>>     threshold=0.5,
>>>     mode="per-class"
>>> )
tensor([0.0000, 0.0000, 1.0000, 1.0000, 1.0000, 0.5])

catalyst.metrics.functional._segmentation.trevsky(outputs: torch.Tensor, targets: torch.Tensor, alpha: float, beta: Optional[float] = None, class_dim: int = 1, threshold: Optional[float] = None, mode: str = 'per-class', weights: Optional[List[float]] = None, eps: float = 1e-07) → torch.Tensor

Computes the trevsky score, trevsky score = tp / (tp + fp * beta + fn * alpha)

Parameters

• outputs – [N; K; …] tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary [N; K; …] tensor that encodes which of the K classes are associated with the N-th input

• alpha – false negative coefficient, bigger alpha bigger penalty for false negative. Must be in (0, 1)

• beta – false positive coefficient, bigger alpha bigger penalty for false positive. Must be in (0, 1), if None beta = (1 - alpha)

• class_dim – indicates class dimention (K) for outputs and targets tensors (default = 1)

• threshold – threshold for outputs binarization

• mode – class summation strategy. Must be one of [‘micro’, ‘macro’, ‘weighted’, ‘per-class’]. If mode=’micro’, classes are ignored, and metric are calculated generally. If mode=’macro’, metric are calculated per-class and than are averaged over all classes. If mode=’weighted’, metric are calculated per-class and than summed over all classes with weights. If mode=’per-class’, metric are calculated separately for all classes

• weights – class weights(for mode=”weighted”)

• eps – epsilon to avoid zero division

Returns

Trevsky score for each class(if mode=’weighted’) or aggregated score

Example

>>> size = 4
>>> half_size = size // 2
>>> shape = (1, 1, size, size)
>>> empty = torch.zeros(shape)
>>> full = torch.ones(shape)
>>> left = torch.ones(shape)
>>> left[:, :, :, half_size:] = 0
>>> right = torch.ones(shape)
>>> right[:, :, :, :half_size] = 0
>>> top_left = torch.zeros(shape)
>>> top_left[:, :, :half_size, :half_size] = 1
>>> pred = torch.cat([empty, left, empty, full, left, top_left], dim=1)
>>> targets = torch.cat([full, right, empty, full, left, left], dim=1)
>>> trevsky(
>>>     outputs=pred,
>>>     targets=targets,
>>>     alpha=0.2,
>>>     class_dim=1,
>>>     threshold=0.5,
>>>     mode="per-class"
>>> )
tensor([0.0000, 0.0000, 1.0000, 1.0000, 1.0000, 0.8333])

Misc

catalyst.metrics.functional._misc.check_consistent_length(*tensors)

Check that all arrays have consistent first dimensions. Checks whether all objects in arrays have the same shape or length.

Parameters

tensors – list or tensors of input objects. Objects that will be checked for consistent length.

Raises

ValueError – “Inconsistent numbers of samples”

catalyst.metrics.functional._misc.get_binary_statistics(outputs: torch.Tensor, targets: torch.Tensor, label: int = 1) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

Computes the number of true negative, false positive, false negative, true positive and support for a binary classification problem for a given label.

Parameters

• outputs – estimated targets as predicted by a model with shape [bs; …, 1]

• targets – ground truth (correct) target values with shape [bs; …, 1]

• label – integer, that specifies label of interest for statistics compute

Returns

stats

Return type

Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]

Example

>>> y_pred = torch.tensor([[0, 0, 1, 1, 0, 1, 0, 1]])
>>> y_true = torch.tensor([[0, 1, 0, 1, 0, 0, 1, 1]])
>>> tn, fp, fn, tp, support = get_binary_statistics(y_pred, y_true)
tensor(2) tensor(2) tensor(2) tensor(2) tensor(4)

catalyst.metrics.functional._misc.get_default_topk_args(num_classes: int) → Sequence[int]

Calculate list params for Accuracy@k and mAP@k.

Parameters

num_classes – number of classes

Returns

array of accuracy arguments

Return type

iterable

Examples

>>> get_default_topk_args(num_classes=4)
[1, 3]

>>> get_default_topk_args(num_classes=8)
[1, 3, 5]

catalyst.metrics.functional._misc.get_multiclass_statistics(outputs: torch.Tensor, targets: torch.Tensor, argmax_dim: int = - 1, num_classes: Optional[int] = None) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

Computes the number of true negative, false positive, false negative, true positive and support for a multiclass classification problem.

Parameters

• outputs – estimated targets as predicted by a model with shape [bs; …, (num_classes or 1)]

• targets – ground truth (correct) target values with shape [bs; …, 1]

• argmax_dim – int, that specifies dimension for argmax transformation in case of scores/probabilities in outputs

• num_classes – int, that specifies number of classes if it known

Returns

stats

Return type

Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]

Example

>>> y_pred = torch.tensor([1, 2, 3, 0])
>>> y_true = torch.tensor([1, 3, 4, 0])
>>> tn, fp, fn, tp, support = get_multiclass_statistics(y_pred, y_true)
tensor([3., 3., 3., 2., 3.]), tensor([0., 0., 1., 1., 0.]),
tensor([0., 0., 0., 1., 1.]), tensor([1., 1., 0., 0., 0.]),
tensor([1., 1., 0., 1., 1.])

catalyst.metrics.functional._misc.get_multilabel_statistics(outputs: torch.Tensor, targets: torch.Tensor) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

Computes the number of true negative, false positive, false negative, true positive and support for a multilabel classification problem.

Parameters

• outputs – estimated targets as predicted by a model with shape [bs; …, (num_classes or 1)]

• targets – ground truth (correct) target values with shape [bs; …, 1]

Returns

stats

Return type

Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]

Example

>>> y_pred = torch.tensor([[0, 0, 1, 1], [0, 1, 0, 1]])
>>> y_true = torch.tensor([[0, 1, 0, 1], [0, 0, 1, 1]])
>>> tn, fp, fn, tp, support = get_multilabel_statistics(y_pred, y_true)
tensor([2., 0., 0., 0.]) tensor([0., 1., 1., 0.]),
tensor([0., 1., 1., 0.]) tensor([0., 0., 0., 2.]),
tensor([0., 1., 1., 2.])

>>> y_pred = torch.tensor([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
>>> y_true = torch.tensor([0, 1, 2])
>>> tn, fp, fn, tp, support = get_multilabel_statistics(y_pred, y_true)
tensor([2., 2., 2.]) tensor([0., 0., 0.])
tensor([0., 0., 0.]) tensor([1., 1., 1.])
tensor([1., 1., 1.])

>>> y_pred = torch.tensor([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
>>> y_true = torch.nn.functional.one_hot(torch.tensor([0, 1, 2]))
>>> tn, fp, fn, tp, support = get_multilabel_statistics(y_pred, y_true)
tensor([2., 2., 2.]) tensor([0., 0., 0.])
tensor([0., 0., 0.]) tensor([1., 1., 1.])
tensor([1., 1., 1.])

catalyst.metrics.functional._misc.process_multilabel_components(outputs: torch.Tensor, targets: torch.Tensor, weights: Optional[torch.Tensor] = None) → Tuple[torch.Tensor, torch.Tensor, torch.Tensor]

General preprocessing for multilabel-based metrics.

Parameters

• outputs – NxK tensor that for each of the N examples indicates the probability of the example belonging to each of the K classes, according to the model.

• targets – binary NxK tensor that encodes which of the K classes are associated with the N-th input (eg: a row [0, 1, 0, 1] indicates that the example is associated with classes 2 and 4)

• weights – importance for each sample

Returns

processed outputs and targets with [batch_size; num_classes] shape

catalyst.metrics.functional._misc.process_recsys_components(outputs: torch.Tensor, targets: torch.Tensor) → torch.Tensor

General pre-processing for calculation recsys metrics

Parameters

• outputs (torch.Tensor) – Tensor with predicted scores size: [batch_size, slate_length] model outputs, logits

• targets (torch.Tensor) – Binary tensor with ground truth. 1 means the item is relevant for the user and 0 not relevant size: [batch_szie, slate_length] ground truth, labels

Returns

targets tensor sorted by outputs

Return type

targets_sorted_by_outputs (torch.Tensor)