road_classifier_satellite

This is the repository for project 2 from the EPFL CS-433 ML course for which we use different Deep Learning and data engineering techniques for data segmentation of roads, using satellite images for Google Maps.

Dependencies setup

To get a working development environment, you can:

go to the utilitary folder and run this line in anaconda prompt:
```
  conda env create -f environment.yml
```

use your own environment and install PyTorch with cuda using:

  conda install -c pytorch torchvision cudatoolkit=10.1 pytorch

Dataset setup

Training data

Create a folder named training with the following structure:

training
├── images
│   └── default
│   └── expanded
└── groundtruth
    └── default
    └── expanded

Copy all training images in the training/images/default folder and all groundtruth images in the training/groundtruth/default folder. Then run the data_augmentation.py script at the root of the project, this will create the augmented dataset in the folders mentionned above.

Test data

Create a folder named test. Take all test images out of their subfolders and copy them in the test folder.

Folders and files

models folder will contain all the trained models for the different architectures trained (unet, vgg, cnn ...).
test contains all the test images (ie. the ones on which we will use our models to estimate the groundtruth mask).
results where the masks of the test images predictions will be stored. This file will be created by running the test.pyscripts.
training which contains the images from AICrowd as well as the data after running data_augmentation.py as described above.
utilitary contains different useful util functions used in this project.
data_augmentation is used to generate the augmented training dataset.
submission_to_mask.py is useful to display the predictions with 16x16 patches as explained in the report.
test.py runs the model on the test data-set and generates the mask predictions in the results folder.
train.py allows to train the different model as explained below.

Running the code

To get the full testing dataset, run:

python split_test.py

Every command below needs to be run from the root of the project and with the name of the model as argument:

unet
resnet
cnn2
cnn4
cnn8
cnn16

To train a model, run:

python train.py <model_name>

To test a model, run:

python test.py <model_name>

To get the full size result images, run:

python unite_results.py <model_name>

To generate the submission file, run:

python mask_to_submission.py <model_name>