Semantic segmentation with MobileNetV3

This repository contains the code for training of MobileNetV3 for segmentation as well as default model for classification. Every module here is subject for subsequent customizing.

Content

• Requirements

• Quick setup and start

• CNN architectures

• Loss functions

• Augmentations

• Training

• Trained model

Requirements

Machine with an NVIDIA GPU
NVIDIA driver >= 418
CUDA >= 10.1
Docker >= 19.03
NVIDIA Container Toolkit (https://github.com/NVIDIA/nvidia-docker)

Quick setup and start

Preparations

• Clone the repo, build a docker image using provided Makefile and Dockerfile.

  git clone 
  make build

• The final folder structure should be:

  airbus-ship-detection
  ├── data
  ├── notebooks
  ├── modules
  ├── train
  ├── Dockerfile
  ├── Makefile
  ├── requirements.txt
  ├── README.md

Run

• The container could be started by a Makefile command. Training and evaluation process was made in Jupyter Notebooks so Jupyter Notebook should be started.

  make run
  jupyter notebook --allow-root

CNN architectures

MobileNetV3 backnone with Lite-RASSP modules were implemented. Architecture may be found in modules/keras_models.py

Loss functions

F-beta and FbCombinedLoss (F-beta with Cross Entropy) losses were implemented. Loss functions may be found in modules/loss.py

Augmentations

There were implemented the following augmentations: Random rotation, random crop, scaling, horizontal flip, brightness, gamma and contrast augmentations, Gaussian blur and noise.

Details of every augmentation may be found in modules/segm_transforms.py

Training

Training process is implemented in notebooks/train_mobilenet.ipynb notebook.

Provided one has at least Pixart and Supervisely Person Dataset it is only needed to run every cell in the notebook subsequently.

Trained model

To successfully convert this version of MobileNetV3 model to TFLite optional argument "training" must be removed from every batchnorm layer in the model and after that pretrained weights may be loaded and notebook cells for automatic conversion may be executed.