Fix Linear .nam tap serialization order

nam/models/linear.py

@@ -32,6 +32,7 @@ def import_weights(self, weights):
         self._net.weight.data = (
             _torch.Tensor([w for w in weights[: self._net.weight.numel()]])
             .reshape(self._net.weight.shape)
+            .flip(-1)
             .to(self._net.weight.device)
         )
         if self._bias:
@@ -55,7 +56,7 @@ def _export_config(self):
         }
 
     def _export_weights(self) -> _np.ndarray:
-        params_list = [self._net.weight.flatten()]
+        params_list = [self._net.weight.flip(-1).flatten()]
         if self._bias:
             params_list.append(self._net.bias.flatten())
         params = _torch.cat(params_list).detach().cpu().numpy()

tests/test_nam/test_models/test_linear.py

@@ -3,6 +3,7 @@
 # Author: Steven Atkinson (steven@atkinson.mn)
 
 import pytest as _pytest
+import torch as _torch
 
 from nam.models import linear as _linear
 
@@ -16,3 +17,51 @@ def setup_class(cls):
         args = ()
         kwargs = {"receptive_field": 2, "sample_rate": 44100}
         super().setup_class(C, args, kwargs)
+
+    def test_export_weights_are_chronological(self):
+        model = _linear.Linear(receptive_field=3)
+        model._net.weight.data.copy_(_torch.tensor([[[0.5, -0.25, 0.125]]]))
+
+        exported_weights = model._export_weights()
+        impulse_response = model(_torch.tensor([1.0, 0.0, 0.0]))
+
+        _torch.testing.assert_close(
+            _torch.from_numpy(exported_weights), impulse_response
+        )
+        _torch.testing.assert_close(
+            _torch.from_numpy(exported_weights),
+            _torch.tensor([0.125, -0.25, 0.5]),
+        )
+
+    def test_import_weights_are_chronological(self):
+        weights = _torch.tensor([0.5, -0.25, 0.125])
+        model = _linear.Linear(receptive_field=3)
+
+        model.import_weights(weights)
+
+        _torch.testing.assert_close(
+            model._net.weight.data, _torch.tensor([[[0.125, -0.25, 0.5]]])
+        )
+        _torch.testing.assert_close(
+            model(_torch.tensor([1.0, 0.0, 0.0])), weights
+        )
+
+    def test_import_export_weights_round_trip_with_bias(self):
+        model = _linear.Linear(receptive_field=3, bias=True)
+        model._net.weight.data.copy_(_torch.tensor([[[0.5, -0.25, 0.125]]]))
+        model._net.bias.data.copy_(_torch.tensor([0.75]))
+        exported_weights = model._export_weights()
+        model2 = _linear.Linear(receptive_field=3, bias=True)
+
+        model2.import_weights(exported_weights)
+
+        _torch.testing.assert_close(
+            _torch.from_numpy(exported_weights),
+            _torch.tensor([0.125, -0.25, 0.5, 0.75]),
+        )
+        _torch.testing.assert_close(model2._net.weight, model._net.weight)
+        _torch.testing.assert_close(model2._net.bias, model._net.bias)
+        _torch.testing.assert_close(
+            _torch.from_numpy(model2._export_weights()),
+            _torch.from_numpy(exported_weights),
+        )

tests/test_nam/test_models/test_sequential.py

@@ -232,6 +232,18 @@ def test_batch_processing(self):
         assert y.shape[0] == 3  # Batch dimension preserved
         assert y.shape[1] == x.shape[1] - seq_model.receptive_field + 1
 
+    def test_export_weights_uses_chronological_linear_taps(self):
+        linear1 = _linear.Linear(receptive_field=3)
+        linear1._net.weight.data.copy_(_torch.tensor([[[1.0, 2.0, 3.0]]]))
+        linear2 = _linear.Linear(receptive_field=1)
+        linear2._net.weight.data.copy_(_torch.tensor([[[4.0]]]))
+        model = _sequential.Sequential(models=[linear1, linear2])
+
+        _torch.testing.assert_close(
+            _torch.from_numpy(model._export_weights()),
+            _torch.tensor([3.0, 2.0, 1.0, 4.0]),
+        )
+
 
 if __name__ == "__main__":
     _pytest.main()