This small (nano) library is used as a sandbox for training and testing models, such as neural networks and convolution networks, on various image classification and object detection problems.

The library is built around several key concepts mapped to C++ object interfaces. Each object instantation is registered with an ID and thus it can be selected from command line arguments. Also new objects can be easily registered and then they are automatically visibile across the library and its associated programs.

A task describes a classification or regression problem consisting of separate training and test image patches with associated target outputs if any. The library has built-in support for various standard benchmark datasets like: MNIST, CIFAR-10, CIFAR-100, STL-10, SVHN, NORB. These datasets are loaded directly from the original (compressed) files.

A model predicts the correct output for a given image patch, either its label (if a classification task) or a score (if a regression task). The library implements various networks with any user-selectable combination of layers: convolution, activation (hyperbolic tangent, unit, signed normalization), linear and pooling.

A loss function assigns a scalar score to the prediction of a model by comparing it with the ground truth target (if provided): the lower the score, the better the prediction.

A trainer optimizes the parameters of a given model to produce the correct outputs for a given task using the cumulated values of a given loss over the training samples as a numerical optimization criteria. Implemented instances: batch (using L-BFGS, conjugate gradient descent - CGD or gradient descent - GD), minibatch and stochastic (using Nesterov's accelerated gradient - AG, adaptive gradient - ADAGRAD, adaptive delta gradient - ADADELTA, stochastic gradient - SG, stochastic iterative averaging - SIA or stochastic gradient averaging - SGA).

Use a C++14 compiler (gcc 4.9+, clang) and install Boost, Eigen3, LibArchive and DevIL.

NanoCV is tested on ArchLinux (gcc 4.9+, CMake 3.1+, Ninja) and OSX (clang, homebrew, CMake 3.1+, Ninja). The code is written to be cross-platform, so it may work (with minor fixes) on other plaforms (e.g. Windows/MSVC).

The easiest way to compile (and install) is to run the build_release.sh bash script. The test programs and utilities will be found in the build-release directory.

The build_debug.sh bash script will build the debugging version with and without address, leak and thread gcc-based sanitizers (if available).

The library provides various command line programs and utilities. Each program displays its possible arguments with short explanations by running it with --help.

• ncv_info - prints all registered objects with their IDs and short descriptions.

• ncv_info_task - loads a task and prints its detailed description.

• ncv_trainer - train a model on a given task.

• ncv_tester - test a model on a given task.

• ncv_generator - creates input image patches that maximally activate an output unit (e.g. associated to a class label).

The scripts directory contains examples on how to train various models on different tasks.