NeuralNetwork

c++/asm implementation of neural network.

The layers works on type Tensor which represents $n$-dimensional array and supports mathematical operations (addition, subtraction, dot product, tensor product, $\ldots$) and other not math realted (adding padding, shuffling, reshaping $\ldots$). Some of the operations are optimized using AVX128 instructions, which made them a lot faster.

Example uses of NeuralNetwork class can be found in:

• applications/mnist digit recognition,
• applications/mnist_vae digit generation.

At this moment there are implemented layers of type:

• dense,
• activation (sigmoid, ReLU, leakyReLU),
• conv2D (not optimized yet),
• pool2D (max or mean pooling),
• reshape,
• flatten,
• dropout,
• normal distribution layer (for VAE).

Build scripts:

• Build/build_debug.sh/Build/build_debug.bat - builds debug configuration,
• Build/build_release.sh/Build/build_release.bat - builds release configuration,
• Build/build_release_sse.sh/Build/build_release_sse.bat - builds release configuration using assembly optimizations,
• Build/clean.sh/Build/clean.bat - cleans all builds.

Other scripts:

• Build/generic_builder.sh/Build/generic_builder.bat - generic builder script used in all build scripts,
• Build/unit_tests_run.sh/Build/unit_tests_run.bat - runs unit tests,
• Build/performance_tests_run.sh - runs performance tests (available only on Linux),
• Build/generate_performance_report.py - runs performance tests, saves the results and saves them on plots ($y$ axis is the measured time and $x$ axis is commit hash). Results can be found here: Build/performance_report/repord.md.