C_API for C++ iterator only supports one data and one label for one sample

[squidszyd]

I'm currently implementing a C++ iterator for performance purpose.
My iterator (name it MyIterator) provides one data and multiple labels of different shapes.
These data and labels are stored in the data attribute of DataBatch object. The arrangement of data and labels may look like the following:

databatch.data[0] = data  // in this case only one data is provided (image data)
databatch.data[1] = label1 // the 1st label, of shape shape<dim>(?)
databatch.data[2] = label2 // the 2nd label, of shape shape<dim>(?)
...

However, as I inspecting into the C API code, I found that this API implementation does not take into account of multiple-data and multiple-label situation like the case above. Instead, only the 0-th data (Line767) of DataBatch is taken as data and only the 1-st data (Line745) of DataBatch is taken as label. All the remainders that MyIterator provides just go to some null space. For instance, at the python end, as I call next, an incomplete batch is returned:

my_iter = mx.io.MyIterator(...)
databatch = my_iter.next()
print len(databatch.data)
>>> 1 // the data is preseverd since only one data is provided
print len(databatch.label)
>>> 1 // only one label is preserved !

Should this issue be placed on some "feature request" ? Maybe the API should consider a more extensible implementation by using key-value pairs to store multiple data and labels. For example:

struct DataBatchEx {
    std::vector<std::pair<std::string, NDArray>> data;  // Instead of vector<NDArray>, store data as kv pairs
    std::vector<std::pair<std::string, NDArray>> labels;  // API will explicitly use this attribute to construct label
    std::vector<uint64_t> index;
    std::string extra_data; // may be discarded
    int num_batch_padd;
};

(Nevertheless, I could go hacking by concating all my labels into one and decoded at python end for this moment.)

[lanking520]

@squidszyd thanks for your discovery. The C++ API is incomplete comparing to other languages support. Please feel free to discover a good way to find a workaround and maintain the C++ package with this feature.

Let's put a feature request here:
@mxnet-label-bot please add [feature, c++] here.

[leleamol]

Hi @squidszyd,
Based on the descriptions, it seems you are referring to implementation of iterators in C API.
The C_API are internal and they are different from C++ API that are available in cpp-package. The later is meant for external consumption.

Also, will it be possible for you to share your use case? Are you planning to use these APIs for training or inference?

The feature request is more appropriate for C_API and not cpp-package.
@mxnet-label-bot can we replace C++ tag with C_API.

[harshp8l]

@mxnet-label-bot update [feature request, c api]

[squidszyd]

Hi @leleamol,
Sorry for the late reply.
Basically in my use case, I want to implement a C++ data loader and use it with Python interface:

// I have my C++ data parser here ...
class MyDataParser {
  // parse data from MXNet's record file using InputSplit class 
};

// and my DataLoader
class MyDataLoader : public IIterator<DataInst> {
  // ...  
  // Init function inherited from IIterator class
  virtual void Init(const std::vector<std::pair<std::string, std::string>>& kwargs) {
    param_.InitAllowUnknown(kwargs);
    parser_.Init(kwargs); // initialize MyDataParser
    iter_.set_max_capacity(4);
    // Init() takes two arguments, the first is a function object describes how single data instance are parsed
    // the second argument is a function object describes how to restart
    iter_.Init(
      [this](std::vector<MyDataInstVector<DType>>** dptr) {
        if(*dptr == nullptr) {
          *dptr = new std::vector<MyDataInstVector<DType>>();
        }
        return parser_.ParseNext(*dptr);
    },
    [this]() {parser_.BeforeFirst(); });
  }
};

You should have noticed that I've defined my own class for a collection of data instances - MyDataInstVector. The reason is that the provided type DataInst that comes with original MxNet does not meet my needs. It can only store one label for one data.

The above actually has nothing to do with MxNet's C_API.
But the point is that if I want to use Python Interface to use my custom data loader, then I have to use C_API, which does not support multi-labels use case.