Automated Label Correction with Existing Models for Supervised Deep Learning

Overview

The repo is divided into the following folders:

Transimprove

This is the core implementation of our project. See the detailed readme for how it works.

Experiments

In this folder there are two files:

• SklearnExperimentDigits.py: With this file we tested our pipeline on the smaller sklearn MNIST dataset containing only 1000 images.
• DeepDivaMnistExperiment.py: With this file we tested our pipeline on the full MNIST dataset containing 60000 training images.

The subfolder helper contains multiple helper functions used during our pipeline.

Testing

In this folder we tested different functionalities in early stages. The proof_of_concept.py shows how to use the pipeline. To run it follow the environment setup.

Environment setup

This project is managed with conda. Use the environment.yaml to automatically install all dependencies:

conda env create -f environment.yaml

To use this environment in pycharm select the created conda environment under project interpreter.

SklearnExperimentDigits.py

This implementation uses sacred to save the parameters in a database and these parameters can be viewed with omniboard in a GUI. See the docs at https://sacred.readthedocs.io for more info about sacred. Before starting this experiment, make sure to start the omniboard/mongodb docker containers using:

➜ docker-compose up -d

To see the results in omniboard go to: http://localhost:9000

DeepDivaMnistExperiment.py

Make sure that deepdiva is installed according to https://diva-dia.github.io/DeepDIVAweb/getting-started.html It is preferred to have GPU support enabled.

To start this experiment in detached mode you can use:

cd [projectDirectory]
source activate deepdiva
nohup python Experiments/DeepDivaMnistExperiment.py &