Rename LMUFFT to LMUFeedforward

CHANGES.rst

@@ -46,6 +46,7 @@ Release history
 - The ``A`` and ``B`` matrices are now stored as constants instead of non-trainable
   variables. This can improve the training/inference speed, but it means that saved
   weights from previous versions will be incompatible. (`#41`_)
+- Renamed ``keras_lmu.LMUFFT`` to ``keras_lmu.LMUFeedforward``. (`#42`_)
 
 .. _#40: https://github.com/nengo/keras-lmu/pull/40
 .. _#41: https://github.com/nengo/keras-lmu/pull/41

keras_lmu/__init__.py

@@ -1,6 +1,6 @@
 """KerasLMU provides a package for deep learning with Legendre Memory Units."""
 
-from .layers import LMU, LMUFFT, LMUCell
+from .layers import LMU, LMUCell, LMUFeedforward
 from .version import version as __version__
 
 __copyright__ = "2019-2021, Applied Brain Research"

keras_lmu/layers.py

@@ -516,7 +516,7 @@ def theta(self):
         if self.built:
             return (
                 self.layer.theta
-                if isinstance(self.layer, LMUFFT)
+                if isinstance(self.layer, LMUFeedforward)
                 else self.layer.cell.theta
             )
 
@@ -541,7 +541,7 @@ def build(self, input_shapes):
             and input_shapes[1] is not None
             and not self.trainable_theta
         ):
-            self.layer = LMUFFT(
+            self.layer = LMUFeedforward(
                 memory_d=self.memory_d,
                 order=self.order,
                 theta=self._init_theta,
@@ -620,15 +620,14 @@ def from_config(cls, config):
         return super().from_config(config)
 
 
-class LMUFFT(tf.keras.layers.Layer):
+class LMUFeedforward(tf.keras.layers.Layer):
     """
-    Layer class for the FFT variant of the LMU.
+    Layer class for the feedforward variant of the LMU.
 
     This class assumes no recurrent connections are desired in the memory component.
 
     Produces the output of the delay system by evaluating the convolution of the input
-    sequence with the impulse response from the LMU cell. The convolution operation is
-    calculated using the fast Fourier transform (FFT).
+    sequence with the impulse response from the LMU cell.
 
     Parameters
     ----------
@@ -747,8 +746,8 @@ def build(self, input_shape):
             # TODO: we could dynamically run the impulse response for longer if
             #  needed using stateful=True
             raise ValueError(
-                f"LMUFFT requires that the input shape's temporal axis be fully "
-                f"specified (got {seq_len})"
+                f"LMUFeedforward requires that the input shape's temporal axis be "
+                f"fully specified (got {seq_len})"
             )
 
         impulse = tf.reshape(tf.eye(seq_len, 1), (1, -1, 1))

keras_lmu/tests/test_benchmarks.py

@@ -50,7 +50,9 @@ def test_performance(mode, capsys):
             return_sequences=False,
         )
     elif mode in ("fft", "raw", "raw_nchw"):
-        lmu_layer = layers.LMUFFT(return_sequences=False, conv_mode=mode, **kwargs)
+        lmu_layer = layers.LMUFeedforward(
+            return_sequences=False, conv_mode=mode, **kwargs
+        )
 
     inputs = tf.keras.layers.Input((seq_len, dims), batch_size=batch_size)
     lmu = lmu_layer(inputs)