21403cafc0fed126c8c0582ea75cb85e641efd65,texar/modules/decoders/transformer_decoders.py,TransformerDecoder,_build,#TransformerDecoder#Any#Any#,71

Before Change

                    dec += poswise_network(dec)
                    dec = layers.layer_normalize(dec)
        self.dec = dec
        self.logits = tf.layers.dense(dec, self._vocab_size)
        self.preds = tf.to_int32(tf.argmax(self.logits, axis=-1))

        if not self._built:
            self._add_internal_trainable_variables()

After Change

        self.dec = dec

        batch_size, length= tf.shape(dec)[0], tf.shape(dec)[1]
        depth = dec.get_shape()[2]

        self.dec = tf.reshape(self.dec, [-1, depth])
        self.logits = tf.matmul(self.dec, tf.transpose(self._embedding))

Instances