2d325991589692bb25d398209022ae03a65074d3,examples/transformer/transformer.py,,,#,27

Before Change

                swords = [ " ".join([vocab._id_to_token_map_py[i] for i in sent]) for sent in source ]
                dwords = [ " ".join([vocab._id_to_token_map_py[i] for i in sent]) for sent in dec_in ]
                twords = [ " ".join([vocab._id_to_token_map_py[i] for i in sent]) for sent in target ]
                if step < 1000:
                    logging.info("source:{}".format(source))
                    logging.info("swords:{}".format(swords))
                    logging.info("dec_in:{}".format(dec_in))
                    logging.info("dwords:{}".format(dwords))
                    logging.info("target:{}".format(target))
                    logging.info("twords:{}".format(twords))
                writer.add_summary(mgd, global_step=step)
                if step % 1000 == 0:
                    print("step:{} loss:{}".format(step, loss))
                    saver.save(sess, logdir+"my-model", global_step=step)

After Change

    else:
        learning_rate = 2 * tf.minimum(1.0, (fstep / opt_hparams["warmup_steps"])) \
            * tf.rsqrt(tf.maximum(fstep, opt_hparams["warmup_steps"])) \
            * encoder_hparams["embedding"]["dim"]**-0.5
    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
            beta1=0.9, beta2=0.98, epsilon=1e-6)
    train_op = optimizer.minimize(mle_loss, global_step)

Instances