CS-433-Project2

This project is about trying to extract roads from satellite images, as part of the EPFL course CS-433. We implement various Transformers and Convolutional Neural Networks.

Setup

The code requires some external libraries, mainly for algorithmic implementations. All the detail can be found in requirement.txt and installed using the command $ pip install -r requirements.txt

Data

Get data from AICrowd page. Organize the files as:

.
├── CNN
│   ├── FullCNN.py 
├── LICENSE
├── README.md
├── data_loader.py
├── dataset
│   ├── test_set_images
│   │   ├── test_1
│   │   │   └── test_1.png
│   │   ├── …
│   ├── test_set_images_short
│   │   ├── test_1
│   │   │   └── test_1.png
│   │   ├── …
│   ├── training
│   │   ├── groundtruth
│   │   │   ├── satImage_001.png
│   │   │   ├── …
│   │   └── images
│   │       ├── satImage_001.png
│   │       ├── …
│   └── training_short
│       ├── groundtruth
│       │   ├── satImage_001.png
│       │   ├── …
│       └── images
│           ├── satImage_001.png
│           ├── …
├── graphs
│   ├── …
├── logs
│   ├── …
├── model_weights
│   ├── …
├── predictions
│   ├── test_1.png
│   ├── …
├── run.py
├── submission.py
├── submissions
│   ├── submission.csv
├── transfer
│   └── transfer_learning.py
├── utils
│   ├── helpers.py
│   ├── mixup.py
│   ├── trainer.py
│   ├── transforms.py
│   └── variables
│       ├── constants.py
│       ├── env_variables.py
│       ├── path_variables.py
│       ├── transformers_variables.py
│       └── variables.py
└── vision_transformer
    ├── ViT_models.py
    ├── ViT_train.py

Usage

The code can be used for different tasks; and only modules run.py and variables.py need to be modified.

Model training

To train a model:

1. Go to variables, then:

   • Set TRAIN_MODELS = True in variables.py
   • Set the variable of the model of your choice to True (FULL_CNN, UNET, TRANSFORMER or TRANSFER), and the other to False
   • Specifiy the training parameters of your choice (for data augmentation, batch size, number of epochs).

2. Go to run.py:

   • In the script, there is one data loader call for each model. Change the desired mixup and short parameter if needed.
   • Run this module and get the weights in the model_weights folder, and training logs in logs

Note: it is possible to use short=True in all prepare_data function, but it is required to first download the training_short.zipfile from the Google Drive and add it in the dataset folder.

Model submission

You can download some of the pretrained weights from the Google Drive and save them on your disk. Then, set SUBMISSION=True in variables.py, then specify the path in run.py.

To create a submission file, as for the training, set to True the same variables corresponding to each model, then specify the name of the model weights (only the name.pth). The submission file is created as submission.csv in submissions.

Acknowledgements

We warmly thank all the CS-433 education team and EPFL for the infrastructure.

Contact

Higlhy Optimized Trio (HOT), Luna, Agustina, Gustave.