training

Raul

Cross-platform on-device training library for mobile and IoT devices

Raul on-device training library is C++ based software designed to train complex neural network (NN) topologies using minimal external dependencies and minimize ROM/RAM footprint. The current implementation of Raul version is CPU-based with BLAS-compatible mathematical back-end used for intensive mathematical operations: matrix multiplication and element-wise vector operations.

Features

• CPU-based computations
• BLAS-compatible math back-end
• 90+ NN layers
• 9 NN optimization algorithms
  • lr schedulers
  • gradient clipping
  • regularization methods
• quantization-aware training
• 6 verified complex topologies (BERT, ResNet, Tacotron 2, NIN, MobileNet 2, MobileNet 3, Transformer)
• Memory efficiency strategies
• Gradient checkpoints
• Workflows aka dynamic allocations

Usage

cmake_minimum_required(VERSION 3.11)
project(sample)

add_subdirectory(raul)

add_executable(app main.cpp)
target_link_libraries(app PRIVATE Raul)

Build

Reqirements

• Build system
  • cmake 3.11+
• Compilers
  • Clang 11.0.0+
  • GCC 9.2.1+
  • Visual Studio 16 (2019)+
  • Android NDK r22+ (r21 without assets related tests)

Host build

cmake -B build -S raul
cmake --build build --target Raul --parallel

This short example shows how to configure, generate a project for default build system and buld using cmake. Here, raul is a directory with repository root and build is an output build directory. All required dependencies will be downloaded; a connection must be established.

Cross-platform Android build

Android NDK is required.

cmake -B build -S raul -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=%path to android.toolchain.cmake% -DRAUL_CONFIG_BLAS_VENDOR=Huawei -DRAUL_CONFIG_ENABLE_OPENMP=ON -DANDROID_ABI=arm64-v8a -DANDROID_NATIVE_API_LEVEL=19 -DANDROID_STL=c++_static
cmake --build build --target Raul --parallel

Read more information