File Descriptions

• README.md: This file.
• weathernet.py: base WeatherNet model {fog, glare, weather, timeofday}
• weathernetplusplus.py: WeatherNet++ model with all seven pipelines
• mtl_weathernet.py: multi-task learning model
• weathernet_transformer.py: transformer model
• train.ipynb: Notebook for model training. Optionally saves model weights.
• evaluate.ipynb: Notebook for model evaluation. Can load checkpointed model weights
• data/: directory containing images, labels, and helper files
• data/BDD100K_plus.py Our custom dataset. It's a subclass of VisionDataset, and loosely follows CIFAR10's format.
• utils/set_seed.py: Sets the seed for reproducibility.

Setup Instructions

Extract the images.tar archive into the data subdirectory, so that it looks like weathernet_plus_plus/data/images/train/0a0a0b1a-7c39d841.jpg etc.

Extract the checkpoints.tar archive into the base project directory, so that it looks like weathernet_plus_plus/checkpoints/mtl_0.pth etc.

Run conda env create -f environment.yml to create a conda environment with the required packages/libraries.

Then run conda activate tensorboard-gpu to activate the conda environment

Directory Structure

weathernet_plus_plus/
├── checkpoints/            # Model weights
├── data/                   # Images, labels, helper files
│   ├── images/
│   │   ├── test/
│   │   ├── train/
│   │   └── val/
│   ├── labels/
│   └── BDD100K_plus.py     # Custom dataset
├── runs/                   # TensorBoard logging
├── utils/                  # Helper functions
├── environment.yml         # Set up conda environment
├── evaluate.ipynb          # Evaluate model
├── README.md               # This file
└── train.ipynb             # Train model

Pipelines and Labels

Pipeline	Label 0	Label 1	Label 2	Label 3	Label 4	Label 5
fog	no fog = 0	fog = 1
glare	no glare = 0	glare = 1
road	dry road = 0	wet road = 1	snowy road = 2
traffic	no traffic = 0	low/moderate traffic = 1	high traffic = 2
weather	clear => 0	partly cloudy => 1	overcast => 2	rainy => 3	snowy => 4	(all else) => 5
scene	residential => 0	highway => 1	city street => 2	(all else) => 3
timeofday	dawn/dusk => 0	daytime => 1	night => 2	(all else) => 3