def shuffle_features()

in tfops.py [0:0]

• 29 lines of code
• 8 McCabe index (conditional complexity)

def shuffle_features(name, h, indices=None, return_indices=False, reverse=False):
    with tf.variable_scope(name):

        rng = np.random.RandomState(
            (abs(hash(tf.get_variable_scope().name))) % 10000000)

        if indices == None:
            # Create numpy and tensorflow variables with indices
            n_channels = int(h.get_shape()[-1])
            indices = list(range(n_channels))
            rng.shuffle(indices)
            # Reverse it
            indices_inverse = [0]*n_channels
            for i in range(n_channels):
                indices_inverse[indices[i]] = i

        tf_indices = tf.get_variable("indices", dtype=tf.int32, initializer=np.asarray(
            indices, dtype='int32'), trainable=False)
        tf_indices_reverse = tf.get_variable("indices_inverse", dtype=tf.int32, initializer=np.asarray(
            indices_inverse, dtype='int32'), trainable=False)

        _indices = tf_indices
        if reverse:
            _indices = tf_indices_reverse

        if len(h.get_shape()) == 2:
            # Slice
            h = tf.transpose(h)
            h = tf.gather(h, _indices)
            h = tf.transpose(h)
        elif len(h.get_shape()) == 4:
            # Slice
            h = tf.transpose(h, [3, 1, 2, 0])
            h = tf.gather(h, _indices)
            h = tf.transpose(h, [3, 1, 2, 0])
        if return_indices:
            return h, indices
        return h