4287aef6a3a82436b4e3e156b22ede235eb4e6ba,texar/modules/decoders/transformer_decoders.py,TransformerDecoder,_build,#TransformerDecoder#Any#Any#,71

Before Change

                        dropout_rate = self._hparams.dropout,
                        causality = False,
                        scope = "vanilla_attention")
                dec = layers.poswise_feedforward(dec,
                        num_units=[4*self._hparams.embedding.dim, self._hparams.embedding.dim])
        self.dec = dec
        //even if don"t use tf.variable_scope(self.variable_scope),
        // the variables will appear in this scope automatically
        self.logits = tf.layers.dense(dec, self._vocab_size)

After Change

                        dropout_rate = self._hparams.dropout,
                        causality = False,
                        scope = "vanilla_attention")
                poswise_network = FeedForwardNetwork(
                        hparams=self._hparams["poswise_feedforward"])
                with tf.variable_scope(poswise_network.variable_scope, reuse=True):
                    // why there could be "multihead_attention_1" scope
                    dec += poswise_network(dec)
                    dec = layers.normalize(dec)
        self.dec = dec
        self.logits = tf.layers.dense(dec, self._vocab_size)
        self.preds = tf.to_int64(tf.argmax(self.logits, axis=-1))

Instances