ef2e9106b17ea0d8bb56cee2ce7442ca4971afae,basic/model.py,Model,_build_forward,#Model#,50

Before Change

            qqc = tf.reshape(tf.reduce_max(tf.nn.relu(qqc), 2), [-1, JQ, d])

        with tf.variable_scope("word_emb"):
            if config.finetune:
                if config.mode == "train":
                    word_emb_mat = tf.get_variable("word_emb_mat", dtype="float", shape=[VW, config.word_emb_size], initializer=get_initializer(config.emb_mat))
                else:
                    word_emb_mat = tf.get_variable("word_emb_mat", shape=[VW, config.word_emb_size], dtype="float")
            else:
                word_emb_mat = config.emb_mat.astype("float32")
            Ax = tf.nn.embedding_lookup(word_emb_mat, self.x)  // [N, M, JX, d]
            Aq = tf.nn.embedding_lookup(word_emb_mat, self.q)  // [N, JQ, d]
            Ax = linear([Ax], d, False, scope="Ax_reshape", wd=config.wd, input_keep_prob=config.input_keep_prob,
                        is_train=self.is_train)

After Change

        with tf.variable_scope("main"):
            u = tf.tile(tf.expand_dims(tf.expand_dims(u, 1), 1), [1, M, JX, 1])
            g0 = tf.concat(3, [h, u, h*u, tf.abs(h-u)])
            (fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(cell, cell, g0, x_len, dtype="float", scope="h1")  // [N, M, JX, 2d]
            g1 = tf.concat(3, [fw_g1, bw_g1])
            (fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(cell, cell, g1, x_len, dtype="float", scope="h2")  // [N, M, JX, 2d]
            g2 = tf.concat(3, [fw_g2, bw_g2])

Instances