This repository contains scripts to train a small neural net classifying images of dogs detailed here:
Towards GPU-accelerated image classification on low-end hardware
The model of 216'152 trainable parameters achieves 72.28% top-1 single crop validation accuracy on 385*385 pixels images of dogs and cats sampled from a challenging 120-class subset of ILSVRC 2012 dataset.
The model architecture is designed to be deployable for inference using an OpenGL ES 2.0-conformant GLSL implementation. This enables inference on a large spectrum of GPUs, from low-end and mobile devices to powerful desktop GPUs. The use of group convolutions, feature maps shuffling and a quantization-friendly activation function effectively reduces the memory bandwidth requirement making the architecture mobile-friendly. Desktop and Android test apps are available in Beatmup.
TensorFlow/Keras is used for training.
-
Install cmake and gcc or clang
-
Install Beatmup Python package
-
Get the code, update submodules and build FastAugment and sigmoid-like activation function extension for TensorFlow:
git submodule update --init --recursive
cd fastaugment
mkdir -p build && cd build
cmake .. && make
cd ../../sigmoid_like_tf_op
mkdir -p build && cd build
cmake .. && make-
Get ILSVRC 2012 dataset
-
Run training/validation using training.ipynb
beatmup_export.ipynb is provided to convert a trained model into a model that can be deployed for inference with Beatmup.