tessl/pypi-torchmetrics
PyTorch native metrics library providing 400+ rigorously tested metrics across classification, regression, audio, image, text, and other ML domains
regression.mddocs/
Regression Metrics
Metrics for evaluating regression tasks including error measurements, correlation coefficients, and explained variance measures for continuous target prediction evaluation.
Capabilities
Error Metrics
Measures various types of prediction errors for regression models.
class MeanSquaredError(Metric):
def __init__(
self,
squared: bool = True,
num_outputs: int = 1,
**kwargs
): ...
class MeanAbsoluteError(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...
class MeanSquaredLogError(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...
class LogCoshError(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...Percentage Error Metrics
Error metrics expressed as percentages of the true values.
class MeanAbsolutePercentageError(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...
class SymmetricMeanAbsolutePercentageError(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...
class WeightedMeanAbsolutePercentageError(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...Normalized Error Metrics
Error metrics normalized by baseline measures.
class NormalizedRootMeanSquaredError(Metric):
def __init__(
self,
normalization: Union[str, Tensor] = "rmse",
num_outputs: int = 1,
**kwargs
): ...
class RelativeSquaredError(Metric):
def __init__(
self,
num_outputs: int = 1,
multioutput: str = "uniform_average",
**kwargs
): ...Correlation Coefficients
Measures of linear and monotonic relationships between predictions and targets.
class PearsonCorrCoef(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...
class SpearmanCorrCoef(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...
class KendallRankCorrCoef(Metric):
def __init__(
self,
variant: str = "b",
num_outputs: int = 1,
**kwargs
): ...
class ConcordanceCorrCoef(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...Coefficient of Determination
Measures the proportion of variance in the target variable explained by the model.
class R2Score(Metric):
def __init__(
self,
num_outputs: int = 1,
multioutput: str = "uniform_average",
adjusted: int = 0,
**kwargs
): ...
class ExplainedVariance(Metric):
def __init__(
self,
num_outputs: int = 1,
multioutput: str = "uniform_average",
**kwargs
): ...Distance and Similarity Metrics
Measures of distance and similarity between prediction and target vectors.
class CosineSimilarity(Metric):
def __init__(
self,
reduction: str = "sum",
**kwargs
): ...
class MinkowskiDistance(Metric):
def __init__(
self,
p: float = 2.0,
**kwargs
): ...Divergence Metrics
Information-theoretic measures of difference between prediction and target distributions.
class KLDivergence(Metric):
def __init__(
self,
reduction: str = "mean",
log_prob: bool = False,
**kwargs
): ...
class JensenShannonDivergence(Metric):
def __init__(
self,
reduction: str = "mean",
log_prob: bool = False,
**kwargs
): ...Specialized Regression Metrics
Domain-specific regression metrics for particular use cases.
class TweedieDevianceScore(Metric):
def __init__(
self,
power: float = 0.0,
**kwargs
): ...
class CriticalSuccessIndex(Metric):
def __init__(
self,
threshold: float = 0.0,
**kwargs
): ...
class ContinuousRankedProbabilityScore(Metric):
def __init__(
self,
num_outputs: int = 1,
**kwargs
): ...Usage Examples
Basic Error Metrics
import torch
from torchmetrics import MeanSquaredError, MeanAbsoluteError, R2Score
# Initialize metrics
mse = MeanSquaredError()
mae = MeanAbsoluteError()
r2 = R2Score()
# Sample predictions and targets
preds = torch.randn(10, 1)
target = torch.randn(10, 1)
# Compute metrics
mse_score = mse(preds, target)
mae_score = mae(preds, target)
r2_score = r2(preds, target)
print(f"MSE: {mse_score:.4f}")
print(f"MAE: {mae_score:.4f}")
print(f"R²: {r2_score:.4f}")Correlation Analysis
from torchmetrics import PearsonCorrCoef, SpearmanCorrCoef
# Initialize correlation metrics
pearson = PearsonCorrCoef()
spearman = SpearmanCorrCoef()
# Sample data
preds = torch.randn(100)
target = 2 * preds + torch.randn(100) * 0.1 # Linear relationship with noise
# Compute correlations
pearson_corr = pearson(preds, target)
spearman_corr = spearman(preds, target)
print(f"Pearson correlation: {pearson_corr:.4f}")
print(f"Spearman correlation: {spearman_corr:.4f}")Multi-output Regression
from torchmetrics import MeanSquaredError, R2Score
# Multi-output regression
mse_multi = MeanSquaredError(num_outputs=3)
r2_multi = R2Score(num_outputs=3, multioutput="raw_values")
# Sample multi-output data
preds = torch.randn(50, 3)
target = torch.randn(50, 3)
# Compute metrics
mse_scores = mse_multi(preds, target)
r2_scores = r2_multi(preds, target)
print(f"MSE (overall): {mse_scores:.4f}")
print(f"R² per output: {r2_scores}")Percentage Error Metrics
from torchmetrics import MeanAbsolutePercentageError, SymmetricMeanAbsolutePercentageError
# Initialize percentage error metrics
mape = MeanAbsolutePercentageError()
smape = SymmetricMeanAbsolutePercentageError()
# Sample data (avoid zeros in targets for MAPE)
preds = torch.randn(100) + 10
target = torch.randn(100) + 10
# Compute percentage errors
mape_score = mape(preds, target)
smape_score = smape(preds, target)
print(f"MAPE: {mape_score:.2f}%")
print(f"SMAPE: {smape_score:.2f}%")Probabilistic Regression
from torchmetrics import ContinuousRankedProbabilityScore
# CRPS for probabilistic forecasts
crps = ContinuousRankedProbabilityScore()
# Sample probabilistic predictions (ensemble of forecasts)
num_samples, num_forecasts = 100, 50
ensemble_preds = torch.randn(num_samples, num_forecasts)
target = torch.randn(num_samples)
# Compute CRPS
crps_score = crps(ensemble_preds, target)
print(f"CRPS: {crps_score:.4f}")Types
MultioutputType = Union["raw_values", "uniform_average", "variance_weighted"]
ReductionType = Union["mean", "sum", "none"]Install with Tessl CLI
npx tessl i tessl/pypi-torchmetrics