def extract_patches(inputs: tf.Tensor, size: int) -> tf.Tensor:
"""Extract overlapping patches from a batch of 1D tensors.
Args:
inputs: Tensor with dimensions [batch, x].
size: number of elements to include in each patch.
Returns:
Tensor with dimensions [batch, x, size].
"""
padded_inputs = layers.pad_periodic(inputs[..., tf.newaxis],
size - 1,
center=True)
extracted = tf.extract_image_patches(padded_inputs[..., tf.newaxis],
ksizes=[1, size, 1, 1],
strides=[1, 1, 1, 1],
rates=[1, 1, 1, 1],
padding='VALID')
return tf.squeeze(extracted, axis=2)
def pad_periodic_1d(inputs, padding, center=False): padded_inputs = inputs[tf.newaxis, :, tf.newaxis] padded_outputs = layers.pad_periodic(padded_inputs, padding, center) return tf.squeeze(padded_outputs, axis=(0, 2))