How do I use an MXNet model in C++?

Question

After I have trained a model, how do I use it with C++?

I have tried MXNet incubator-mxnet/example/image-classification/predict-cpp/ and incubator-mxnet/cpp-package/example/.

Answer 1

As part of training you should periodically evaluate your model against a validation set, at the end of each epoch for example. You should then have a good idea of the expected accuracy of the model when using the model to score new data, to determine if the model is really performing worse than expected at inference time.

If the validation accuracy of the model while training the model is no better than random (i.e. 1/number of classes), there could be many reasons for this including; poor model selection, incorrect loss calculation, wrong optimization technique and hyperparameters (e.g. learning rate).

If the test accuracy of the model on unseen data is poor, you might be trying to apply the model to a different domain to which it was trained. You can't use a model trained on handwritten characters (e.g. MNIST) to classify real world objects (e.g. ImageNet).

If you need a C++ example of model training, take a look at this tutorial.