How to get top N results for seq2seq?

[lingz]

In the example translate seq2seq model, mentioned in the tutorial and in the code here (reproduced below):

      # Get a 1-element batch to feed the sentence to the model.
      encoder_inputs, decoder_inputs, target_weights = model.get_batch(
          {bucket_id: [(token_ids, [])]}, bucket_id)
      # Get output logits for the sentence.
      _, _, output_logits = model.step(sess, encoder_inputs, decoder_inputs,
                                       target_weights, bucket_id, True)
      # This is a greedy decoder - outputs are just argmaxes of output_logits.
      outputs = [int(np.argmax(logit, axis=1)) for logit in output_logits]

It uses only a greedy decoder, and uses argmax to find the best match. I'm wondering if there's a way to get the top N results instead of just doing it greedily. I've tried argsort, but everything apart from the 0th index is just garbage results. I've also looked into tf.nn.top_k(), but because this is batched, I get the error "List of Tensors when single Tensor expected" and I'm not sure how to unroll the list within TF.

[lukaszkaiser]

To get top-N results in a sequential network you need to run a beam search. While it's not implemented in the tutorial, there were already some suggestions and code on github -- please see #654 .