keras-team/keras · error · ValueError

Threshold values must be in [0, 1]. Received: {invalid_thres

Error message

Threshold values must be in [0, 1]. Received: {invalid_thresholds}

What it means

Keras validates that every threshold used by thresholded metrics (Precision, Recall, AUC with explicit thresholds) lies in [0,1]. parse_init_thresholds() calls assert_thresholds_range(), which collects offending values and raises this ValueError. Thresholds below 0, above 1, or None entries inside the list trigger it.

Source

Thrown at keras/src/metrics/metrics_utils.py:20

from enum import Enum

import numpy as np

from keras.src import backend
from keras.src import ops
from keras.src.losses.loss import squeeze_or_expand_to_same_rank
from keras.src.utils.python_utils import to_list

NEG_INF = -1e10


def assert_thresholds_range(thresholds):
    if thresholds is not None:
        invalid_thresholds = [
            t for t in thresholds if t is None or t < 0 or t > 1
        ]
        if invalid_thresholds:
            raise ValueError(
                "Threshold values must be in [0, 1]. "
                f"Received: {invalid_thresholds}"
            )


def parse_init_thresholds(thresholds, default_threshold=0.5):
    if thresholds is not None:
        assert_thresholds_range(to_list(thresholds))
    thresholds = to_list(
        default_threshold if thresholds is None else thresholds
    )
    return thresholds


class ConfusionMatrix(Enum):
    TRUE_POSITIVES = "tp"
    FALSE_POSITIVES = "fp"
    TRUE_NEGATIVES = "tn"

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Clamp or filter thresholds to [0,1]: [t for t in thresholds if 0 <= t <= 1].
2. If you have logit-scale scores, convert to probabilities with a sigmoid before using them as thresholds.
3. Remove None entries from the thresholds list; pass thresholds=None to use the default 0.5.

Example fix

# before
metric = keras.metrics.Precision(thresholds=[0.5, 1.2])

# after
metric = keras.metrics.Precision(thresholds=[0.5, 0.9])

def check_thresholds(thresholds):
    if thresholds is not None:
        bad = [t for t in thresholds if t is None or t < 0 or t > 1]
        if bad:
            raise ValueError(f'thresholds outside [0,1]: {bad}')
    return thresholds

When it happens

Trigger: Constructing keras.metrics.Precision(thresholds=[0.2, 1.5]), Recall(thresholds=[-0.1]), or AUC(thresholds=[None, 0.5]) - any value <0, >1, or None inside the thresholds list.

Common situations: Reading thresholds from a config file or hyperparameter sweep where values escape [0,1]; mixing logit-scale values (e.g. 5.0) into probability thresholds.

Related errors

• Could not interpret activation function identifier: {identif
• If using `weights="imagenet"` as true, `classes` should be 1
• The number of repeats in `EfficientNet` must be > 0. Receive
• If using `weights="imagenet"` as true, `classes` should be 1
• The number of repeats in `EfficientNetV2` must be > 0. Recei

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/4531c0a9060d7a38. Report an issue: GitHub.