This repository contains the code and report produced by our team for the project 2 of the Machine Learning course at EPFL in 2018. It uses a modified UNet to do a road segmentation.

Our best result on crowd.ai is done with the username Mare, submission id: 25240

• Team Members
• Environment
• Libraries
• Technical problems
• Structure

• Marijana Peti
• Loïc Wisniewski
• Christophe Minutolo

As training a deep learning model takes a long time and needs powerful computers, we decided to use Google Colab to run our algorithms. We used the GPU runtime to accelerate the learning process. Moreover all useful libraries for deep learning are already available on Google Colab.

To reproduce results of this repository, you will have to put all the repository in Google Drive then run in Google Colab this:

from google.colab import drive
drive.mount('/content/drive')
%cd "your_path" # Path to your project folder in Google Drive
%pwd

Do not forget to set the accelerator on GPU in the runtime configuration of Google Colab. You will have to put in "your_path"/data/test_set_images all images from the test set (they should not have to be in sub-folders). You also have to put the saved weights in "your_path"/model_weights.hdf5. This weights are available here.

Once done, you can run in Google Colab this code:

from keras_prototyping import *
load_model_and_create_submission_file("./model_weights.hdf5", "./results.csv")

Or if you want to run it on your computer, you can just launch run.py in Google Colab with the command !python run.py

It will output predictions in "your_path"/results.csv.

If you want to recreate the model from scratch, you will have to download the training files available on crowd.ai. Put images from 001 to 090 in the "your_path"/data/training/train/images and respective groundtruth in folder "your_path"/data/training/train/groundtruth. Put the rest in "your_path"/data/training/validation/images and "your_path"/data/training/validation/groundtruth. Then you will be able to proceed to data augmentation and run the model generation:

from keras_prototyping import *
import datatools
datatools.generate_files()
usage_example()
load_model_and_create_submission_file("./model_weights.hdf5", "./results.csv")

It will output the weight in "your_path"/model_weights.hdf5

This project use the following libraries:

• Keras
• TensorFlow
• matplotlib
• numpy
• PIL
• skimage

On Google Cloud as memory is limited, we ofter had many OOM errors. To reduce memory consumption, we decided to do all the data-augmentation before starting the training, and save all augmentation in the training folder. Once done we were able to write a custom data generator that gives to Keras data incrementally to avoid to fill the memory that are not useful for now. You can check in the code in datatools.py the function custom_image_generator.

• keras_prototyping.py contains all the main function to launch models easily.
• plotting.py contains plot functions, transformations from mask to images and from CSV to mask.
• models.py contains the UNet models and some try to implement transfer learning.
• losses.py contains all loss function (Dice, BCE + Loss) and monitor functions (Accuracy, F1 Score).
• datatools.py contains function to do data augmentation.
• submission.py contains functions to create CSV.