Training with keras+tensorflow in python and predict in C/C++ - c

Question

It is possible to make a fast model with keras with tensor flow as backend and use it to predict in C or C++?
I need to do the prediction inside of a C++ program, but I feel much more comfortable doing the model and the training in keras.

Answer 1

In case you don't need to utilize a GPU in the environment you are deploying to, you could also use my library, called frugally-deep. It is available on GitHub and published under the MIT License: https://github.com/Dobiasd/frugally-deep
frugally-deep allows running forward passes on already-trained Keras models directly in C++ without the need to link against TensorFlow or any other backend.
It not only supports prediction with sequential models but also with more complex models build with the functional API.
Additionally to supporting many common layer types it can keep up with (and sometimes even beat) the performance of TensorFlow on a single CPU. You can find up-to-date benchmark results for some common model in the repo.
By automatic testing frugally-deep guarantees that the output of a model used with it in C++ is exactly the same as if run with Keras in Python.

Answer 2

Yes, it is possible. TensorFlow provides a stable C API as well as a C++ one.
For more details, you probably want to ask a more specific question.

Answer 3

You can use cv::dnn module in opencv3.2. See example in the opencv samples.

Answer 4

In simple steps you can do this -
Save Keras model into text files in python
In C++/C, read those text files, and run the model in a multithreading environment.
Above Steps in detail
For first step, call save_keras_model_as_text from keras2cpp.py on Github-
'''
save_keras_model_as_text(model, open('/content/modeltext.txt', 'w') )
save_keras_model_as_text(model, open('/content/modellayersonlytext.txt', 'w') , noweight=True)
'''
In c++/c projects, include keras.h /keras.cpp files.
convert input image into Keras format or read input image into Keras format from a file.
'''
//assume dst is pointer to an image object. This
//can be of CxImage / CImg /OpenCV or of any other
//library.
int h = dst->GetHeight();
int w= dst->GetWidth();;
int *img = new int[h*w*3];
int *img_r = img, *img_g =&img[h*w], *img_b=&img[2*h*w];
for (int y = 0; y < h; y++) {
for (int x = 0; x < w;x++) {
RGBQUAD rgb= dst->GetPixelColor(x,y);
*img_r= rgb.rgbRed; img_r++;
*img_g= rgb.rgbGreen; img_g++;
*img_b = rgb.rgbBlue; img_b++;
}
}
'''
Call ExecuteKerasSegmentation.
'''
//text files in which keras model is saved.
char *modelfile ="modellayersonlytext.txt";
char *weightfile="modeltext.txt";
int *result = ExecuteKerasSegmentation(img, h, w, 3, modelfile, weightfile);
'''
result contains the segmentation result. You may save it into pgm file or convert it into your image library object and use it further.
'''
save_image_pgm("segmentation_map.pgm",result,h,w,127); //127 is scaling factor for binary images.
'''