Performance issues in the definition of _inference, MemN2N-split-memory/memn2n/memn2n_dialog.py(P1)

[DLPerf]

Hello, I found a performance issue in the definition of _inference, MemN2N-split-memory/memn2n/memn2n_dialog.py, tf.nn.embedding_lookup(self.A, stories) will be created repeatedly during program execution, resulting in reduced efficiency. I think it should be created before the loop.

The same issues exist in:

• tf.nn.embedding_lookup(self.A, profile) in line 158;
• tf.reduce_sum(m_emb, 2) in line 160 and 161;
• tf.transpose(m, [0, 2, 1]) in line 183;
• tf.transpose(m_profile, [0, 2, 1]) in line 184;
• tf.nn.embedding_lookup(self.A, stories) in line 154;
• tf.reduce_sum(m_emb, 2) in line 155;
• tf.transpose(m, [0, 2, 1]);

Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.