Tensorflow-driven neural network quantization for resource-constrained devices

Design space exploration with the most promising discretization techniques offered by Tensorflow against some Neural Networks belonging to CNN, MLP and LSTM families. The quantization with different representational precision is performed using a Post-training approach with already trained Models, and with a Quantization-aware training approach with a comprehensive training session from scratch. The results collected demonstrate the effectiveness of these Techniques that can ensure minimal accuracy loss with interesting improvement in Model compression and, during the deploying on resource-constrained devices, hardware accelerations.