fix: optimal threshold computation

CHANGELOG.md

@@ -4,6 +4,12 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 
+## [v.0.7.0.dev151]
+
+### Fixed
+
+- Use combined torch.nextafter and fixed epsilon to select optimal threshold when only one target is present.
+
 ## [v.0.7.0.dev150]
 
 ### Updated

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "anomalib-orobix"
-version = "0.7.0.dev150"
+version = "0.7.0.dev151"
 description = "Orobix anomalib fork"
 authors = [
     "Intel OpenVINO <[email protected]>",

src/anomalib/utils/metrics/optimal_f1.py

@@ -2,12 +2,14 @@
 # SPDX-License-Identifier: Apache-2.0
 
 """Implementation of Optimal F1 score based on TorchMetrics."""
-from typing import Optional
+
 import warnings
+from typing import Optional
 
 import torch
 from torch import Tensor
 from torchmetrics import Metric, PrecisionRecallCurve
+from zmq import device
 
 
 class OptimalF1(Metric):
@@ -54,12 +56,21 @@ def compute(self) -> Tensor:
 
         epsilon = 1e-3
         if len(current_targets.unique()) == 1:
+            # Use torch nextafter to ensure that the threshold is higher (or smaller)
+            # than the maximum (or minimum) score. This ensure correctness for lower precisions.
+            # Combined method is to avoid very small shifts around zero.
+            _inf = torch.tensor(torch.inf, dtype=current_preds.dtype, device=current_preds.device)
             optimal_f1_score = torch.tensor(1.0)
 
             if current_targets.max() == 0:
-                self.threshold = current_preds.max() + epsilon
+                max_score = current_preds.max()
+                self.threshold = torch.max(torch.nextafter(max_score, _inf), max_score + epsilon)
             else:
-                self.threshold = current_preds.min() - epsilon
+                min_score = current_preds.min()
+                self.threshold = torch.min(torch.nextafter(min_score, -_inf), min_score - epsilon)
+
+            if torch.isinf(self.threshold):
+                raise RuntimeError(f"Invalid value computed for the threshold: {self.threshold}.")
 
             return optimal_f1_score
         else: