CCNN

Machine learning models

For now, only the LeNet model is available. Since CCNN doesn't have training support (it only does inference), the model has been trained in TensorFlow/Keras. The weights are stored in the src/models/lenet/data directory. The input image is a grayscale handwritten number "6" from the MNIST dataset (used to train our LeNet model).

Installation

This C++ project uses CMake 3.10. It uses the PAPI library to provide useful profiling utility functions.

Quick install

CMake will automatically detect if you have the PAPI library installed on your computer and build with or without profiling support depending on that.

mkdir build
cd build
cmake ..
cmake --build .

Profiling support

You can manually set the ENABLE_PROFILING flag to enable or disable profiling support. Setting it to 0 will disable profiling support, while setting it to anything else will enable profiling support.

The following commands will disable profiling suport:

mkdir build
cd build
cmake -DENABLE_PROFILING=0 ..
cmake --build .

The following commands will enable profiling suport:

mkdir build
cd build
cmake -DENABLE_PROFILING=1 ..
cmake --build .

Execution

To run the executable in the build directory, run the following command:

./ccnn

Additionally, some CLI arguments are available:

--debug flag (or -d): print every step on your terminal (intermediate feature maps, filters, biases and the result).
--batchsize=<int> option: specify a batch size (defaults to 1). Since there is only one input image, this will duplicate the input n times.
--profile=<list int> option: specify the IDs of the layers you want to profile.

An example of running CCNN with a batchsize of 256 and profiling enabled for layers 1 and 3:

./ccnn --batchsize=256 --profile=1,3