Fix TTQ layer tests: Account for softplus transformation

TTQ parameters (wp, wn, delta) are stored in inverse softplus space
to ensure they stay positive during training. Tests need to apply
softplus before comparing to expected initialization values.

Author: dariocazzani

tests/test_ttq_layers.py

@@ -36,11 +36,17 @@ def test_parameters_initialized_properly(self) -> None:
         """TTQ parameters should be initialized per paper (Eq. 2)."""
         layer = TTQLinear(64, 32)
         weight_mean_abs = layer.weight.data.abs().mean()
+
+        # wp and wn are stored in inverse softplus space, apply softplus to compare
+        wp_actual = torch.nn.functional.softplus(layer.wp)
+        wn_actual = torch.nn.functional.softplus(layer.wn)
+        delta_actual = torch.nn.functional.softplus(layer.delta)
+
         # wp and wn should be initialized to E[|W|]
-        assert torch.allclose(layer.wp, weight_mean_abs, rtol=1e-5)
-        assert torch.allclose(layer.wn, weight_mean_abs, rtol=1e-5)
+        assert torch.allclose(wp_actual, weight_mean_abs, rtol=1e-5)
+        assert torch.allclose(wn_actual, weight_mean_abs, rtol=1e-5)
         # delta should be initialized to 0.7 * E[|W|]
-        assert torch.allclose(layer.delta, 0.7 * weight_mean_abs, rtol=1e-5)
+        assert torch.allclose(delta_actual, 0.7 * weight_mean_abs, rtol=1e-5)
 
     def test_numerical_stability_during_training(self) -> None:
         """Training should not produce NaN losses."""
@@ -98,11 +104,17 @@ def test_parameters_initialized_properly(self) -> None:
         """TTQ parameters should be initialized per paper (Eq. 2)."""
         layer = TTQConv2d(3, 16, kernel_size=3)
         weight_mean_abs = layer.weight.data.abs().mean()
+
+        # wp and wn are stored in inverse softplus space, apply softplus to compare
+        wp_actual = torch.nn.functional.softplus(layer.wp)
+        wn_actual = torch.nn.functional.softplus(layer.wn)
+        delta_actual = torch.nn.functional.softplus(layer.delta)
+
         # wp and wn should be initialized to E[|W|]
-        assert torch.allclose(layer.wp, weight_mean_abs, rtol=1e-5)
-        assert torch.allclose(layer.wn, weight_mean_abs, rtol=1e-5)
+        assert torch.allclose(wp_actual, weight_mean_abs, rtol=1e-5)
+        assert torch.allclose(wn_actual, weight_mean_abs, rtol=1e-5)
         # delta should be initialized to 0.7 * E[|W|]
-        assert torch.allclose(layer.delta, 0.7 * weight_mean_abs, rtol=1e-5)
+        assert torch.allclose(delta_actual, 0.7 * weight_mean_abs, rtol=1e-5)
 
     def test_numerical_stability_during_training(self) -> None:
         """Training should not produce NaN losses."""