62d2ea56ae4a090aa68baf133137982a836700bd,thumt/utils/search.py,,create_inference_graph,#Any#Any#Any#,430

Before Change

            return ids[:, :top_beams, 1:]
    else:
        if top_beams == 1:
            return ids[:, 0, 1:], scores[:, 0]
        else:
            return ids[:, :top_beams, 1:], scores[:, :top_beams]

After Change

                              lp_constant=params.decode_constant)

    mask = tf.not_equal(ids, 0)
    output_length = tf.reduce_sum(tf.to_float(mask), axis=-1)

    // shape: [batch, beam_size]
    normalized_scores = scores / output_length

    if params.decode_normalize:
        scores, indices = tf.nn.top_k(normalized_scores, k=top_beams)
        // shape of ids: [batch, beam_size, max_length]
        // shape of coordinates: [batch, beam_size, 2]
        batch_pos = compute_batch_indices(batch_size, beam_size)
        coordinates = tf.stack([batch_pos, indices], axis=2)

Instances