Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add new computational path when return_sequences=False #53

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add new computational path when return_sequences=False #53

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
50 changes: 40 additions & 10 deletions keras_lmu/layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -877,7 +877,12 @@ def build(self, input_shape):
self.delay_layer(impulse, training=False), axis=0
)

if self.conv_mode == "fft":
if not self.return_sequences and not hasattr(self.hidden_cell, "state_size"):
# when return_sequences=False and the hidden cell is not an RNN,
# it is computationally more efficient to compute the output as
# described in equation (25) in https://arxiv.org/pdf/2102.11417.pdf.
self.conv_mode = "matmul"
elif self.conv_mode == "fft":
self.impulse_response_fft = (
None
if seq_len is None
Expand Down Expand Up @@ -961,20 +966,27 @@ def call(self, inputs, training=None):
if self.bias is not None:
u = u + self.bias

if self.conv_mode == "fft":
if self.conv_mode == "matmul":
m = self._matmul(u)
elif self.conv_mode == "fft":
m = self._fft_convolution(u)
elif self.conv_mode == "raw":
m = self._raw_convolution(u)

# apply hidden cell
h_in = (
tf.concat((m, inputs), axis=-1) # pylint: disable=no-value-for-parameter
if self.input_to_hidden
else m
)
if self.conv_mode == "matmul" and self.input_to_hidden:
h_in = tf.concat(
(m, inputs[:, -1]), axis=-1
) # pylint: disable=no-value-for-parameter
elif self.input_to_hidden:
h_in = tf.concat(
(m, inputs), axis=-1
) # pylint: disable=no-value-for-parameter
else:
h_in = m

if self.hidden_cell is None:
h = h_in if self.return_sequences else h_in[:, -1]
h = h_in
elif hasattr(self.hidden_cell, "state_size"):
h = tf.keras.layers.RNN(
self.hidden_cell,
Expand All @@ -983,15 +995,33 @@ def call(self, inputs, training=None):
)(h_in, training=training)
else:
if not self.return_sequences:
# no point applying the hidden cell to the whole sequence
h = self.hidden_cell(h_in[:, -1], training=training)
h = self.hidden_cell(h_in, training=training)
else:
h = tf.keras.layers.TimeDistributed(self.hidden_cell)(
h_in, training=training
)

return h

def _matmul(self, u):
seq_len = tf.shape(u)[1]
impulse_len = self.impulse_response.shape[0]

if seq_len >= impulse_len:
impulse_response = tf.pad(
self.impulse_response,
paddings=[
[0, seq_len - impulse_len],
[0, 0],
],
)
else:
impulse_response = impulse_response[:seq_len, :]

u = tf.transpose(tf.reverse(u, axis=[1]), perm=[0, 2, 1])
m = tf.reshape(tf.matmul(u, impulse_response), (-1, self.memory_d * self.order))
return m

def _fft_convolution(self, u):
seq_len = tf.shape(u)[1]

Expand Down