MobileNetV3-Segmentation-Keras
This repository implements the semantic segmentation version of the MobileNetV3 architecture (source), which is inspired by the DeeplabV3 architecture.
The model is implemented using Keras and TensorFlow 2.x. The MobileNetV3 backbone is based on the official model from the keras_applications package.
This repository was created as part of a master thesis, which includes more details on design choices and the limits of the implementation. It can be found under this link.
Cite:
Bernhard Haas, Alexander Wendt, Axel Jantsch and Matthias Wess: Neural Network Compression Through Shunt Connections and Knowledge Distillation for Semantic Segmentation Problems, in Proceedings of Artificial Intelligence Applications and Innovations, 17th IFIP WG 12.5 International Conference, Greece (online), pp. 349-361, 2021, doi: https://doi.org/10.1007/978-3-030-79150-6
and
Bernhard Haas, Compressing MobileNet With Shunt Connections for NVIDIA Hardware, Master Thesis, TU Wien, 2021, url: https://publik.tuwien.ac.at/files/publik_295948.pdf
Loading pre-trained weights from the official repository
The original implementation of the model was built using TensorFlow 1 and TF-Slim. Pre-trained weights for this model can be found here.
Using these weights in Keras can be done by converting the saved tensors to numpy arrays and saving them as .npy files. Those files can later be loaded as Keras weights, as long as the correct naming is used.
The script utils/convert_checkpoint_to_npy.py extracts the tensors from a checkpoint file and saves them as numpy arrays with the correct names for this Keras implementation.
Public release from XXXX-XX-XX. Contact Mission Embedded for release date.