This repository contains the work done for the Human Data Analytics course project of the Master's Degree in Data Science at the University of Padua.
The work done is reported in the submitted paper; two new bottlenecks based on the MobileNetV3 and Fused inverted bottlenecks are introduced, and the best fully-convolutional model found exploiting them achieves 0.488 mAP on FSD50K and 0.967 accuracy on ESC-50 outperforming the models from the EfficientNet series with a similar GPU utilization, and without employing special training techniques.
Pretrained weights for the official PyTorch implementation are available in the releases.