From 310ad3701c74abb29e8cba78de25678580cc8b31 Mon Sep 17 00:00:00 2001
From: ChrisW09 <50968720+ChrisW09@users.noreply.github.com>
Date: Wed, 21 Oct 2026 20:40:21 +0200
Subject: [PATCH] fix(distributions): keep zero targets exact in the Tweedie
 deviance

evaluate_nll() clamped y to 1e-8 before y ** (2 - p), so every zero
target added 2 * 1e-8 ** (2 - p) / ((1 - p)(2 - p)) to the deviance
(about -3.5 per zero target for p = 1.9). Since 2 - p > 0, 0 ** (2 - p)
is already 0; clamp at 0 only to keep negative values out.

Fixes #525
---
 deeptab/distributions/tweedie.py          |  2 +-
 tests/distributions/test_distributions.py | 14 ++++++++++++++
 2 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/deeptab/distributions/tweedie.py b/deeptab/distributions/tweedie.py
index bd15dec0..efba250c 100644
--- a/deeptab/distributions/tweedie.py
+++ b/deeptab/distributions/tweedie.py
@@ -77,7 +77,7 @@ def evaluate_nll(self, y_true, y_pred):
         # Tweedie deviance: D(y, mu) = 2 * [y^(2-p)/((1-p)(2-p)) - y*mu^(1-p)/(1-p) + mu^(2-p)/(2-p)]
         p = self.p
         d = 2.0 * (
-            torch.pow(y_true_tensor.clamp(min=1e-8), 2.0 - p) / ((1.0 - p) * (2.0 - p))
+            torch.pow(y_true_tensor.clamp(min=0.0), 2.0 - p) / ((1.0 - p) * (2.0 - p))
             - y_true_tensor * torch.pow(mu, 1.0 - p) / (1.0 - p)
             + torch.pow(mu, 2.0 - p) / (2.0 - p)
         )
diff --git a/tests/distributions/test_distributions.py b/tests/distributions/test_distributions.py
index 22d3cc6f..1eb727f2 100644
--- a/tests/distributions/test_distributions.py
+++ b/tests/distributions/test_distributions.py
@@ -288,6 +288,20 @@ def test_evaluate_nll_keys(self):
         for key in ("NLL", "mse", "mae", "rmse", "tweedie_deviance"):
             assert key in metrics
 
+    @pytest.mark.parametrize("p", [1.5, 1.9])
+    def test_evaluate_nll_deviance_matches_sklearn_with_zero_targets(self, p):
+        import numpy as np
+        from sklearn.metrics import mean_tweedie_deviance
+
+        from deeptab.distributions import TweedieDistribution
+
+        family = TweedieDistribution(p=p)
+        raw = self.torch.tensor([[-0.5], [0.3], [1.8], [-0.1]])
+        targets = np.array([0.0, 0.0, 3.0, 1.0])
+        means = family(raw)[:, 0].detach().numpy()
+        deviance = float(family.evaluate_nll(targets, raw.numpy())["tweedie_deviance"])
+        assert deviance == pytest.approx(mean_tweedie_deviance(targets, means, power=p), rel=1e-5)
+
     @pytest.mark.parametrize("p", [1.1, 1.5, 1.9])
     def test_various_p_values(self, p):
         from deeptab.distributions import TweedieDistribution