How to make top-k predict

[shuokay]

I have made top-1 predict using ARGMAX layer, but now I want to make a top-k predict. I have found #531 ,but I cant find how to make predict using accuracy layer with the top-k field.
secondly, I found this code https://gist.github.com/onauparc/dd80907401b26b602885 mentioned in #499 , however, as I used a IMAGE_DATA layer, I cannot get right result using that code.
here is my code:

  Net<float> caffe_test_net(argv[1]);
  caffe_test_net.CopyTrainedLayersFrom(argv[2]);
  float loss;
  float right=0;
  const vector<Blob<float>*>& result =  caffe_test_net.ForwardPrefilled(&loss);
  for (int j=0; j<result[1]->num(); ++j)
  {
    float max = 0;
    float max_i = 0;
    for (int i = 0; i < 3904; ++i) {
        float value = result[1]->cpu_data()[j*3904+i];
        if (max < value){
        max = value;
        max_i = i;
        }
    }
  }

here is just the top-1 code, I will change it to top-k code later.
where 3904 is the number of categories, because I dont know how to get them from the blobs, I just write them in code.
Will someone kindly tell me the last layer is SOFTMAX or ACCURACY if i want to make predict in C++ without an ARGMAX layer ? I think it is SOFTMAX , however because I want to get the labels, if I make the last layer SOFTMAX, more errors come out that ACCURACY.
the last two layers of my proto file:

layers {
  name: "prob"
  type: SOFTMAX
  bottom: "ip2"
  top: "prob"
}
layers{
  name:"accuracy"
  type: ACCURACY
  bottom: "prob"
  bottom: "label"
  top: "accuracy"
}

updated:
I have solved most problems of this code, the answers are followed:

1. I have updated my code, in the code, the only problem is how to get the numbers of categories(3904) from the blobs instead of just fix it in code.
2. the last layer should beACCURACY layer.

I have not really understand the difference between SOFTMAX layer and ACCURACY layer, and I will go on to read the caffe source code.

[kloudkl]

Would you mind surrounding the codes with the GitHub markdown? for (int j=0; j { and for (int i = 0; i { are not complete. What ranges do they loop over?
The ArgMaxLayer must be extended to support top-k prediction.

layers {
  name:"argmax"
  type: ARGMAX
  bottom: "prob"
  top: "argmax"
  argmax_param {
    top_k: 10
  }
}

layers {
  name:"accuracy"
  type: ACCURACY
  bottom: "prob"
  bottom: "label"
  top: "accuracy"
  accuracy_param {
    top_k: 10
  }
}

[shuokay]

@kloudkl thank you very much for the suggest of surrounding method and your answer about top-k prediction. I have updated my issue.
But I have new questions from your answer.

1. Was ACCURACY still needed to make top-k prediction?
2. How can I control it is muliple labels or single label if I use a top-K ARGMAX layer, or is there relationship between top-k prediction and multiple labels inputs?
3. Would you kindly tell me the the role that ACCURACY layer played in caffe. I am reading caffe source code, but have not read the ACCURACY layer

[kloudkl]

The top k ArgMaxLayer is in #615 and the multi-label AccuracyLayer is in #523.