Assessment of Uncertainty of Neural Net Predictions in the Tasks of Classification, Detection and Segmentation

This repository provides the code used in my bachelor`s thesis "Assessment of Uncertainty of Neural Net Predictions in the Tasks of Classification, Detection and Segmentation".

You can watch the presentational video: https://youtu.be/rt9T6uYYrIQ

Installation

with conda:

conda env create --file environment.yml

with pip:

pip install -r requirements.txt

Repository structure

The training scripts and evaluating jupyter notebooks are located in the top level directory. They are named by the task (classification, segmentation, detection) and dataset used.

Code for evaluation, training and uncertainty estimation methods is located in the utils folder.

Code for the models is located in the models folder.

Code for data transformations and pytorch datasets is located in the datasets directory.

The figures folder contains all of the figures used in the paper.

The experiments folder contains exported results of evaluation on shifted datasets.

Pre-trained models

Download the pre-trained models from https://drive.google.com/drive/folders/1Uzw0pO-NPe6l5SGLFjZEqQRggZMfMFAl?usp=sharing and place them in the checkpoints folder.

Training

The training scripts are .py files named by the dataset used.

Example: training U-Net on the MedSeg Covid19 dataset:

python segmentation-covid19-training.py

You can change the model architecture used by uncommenting lines in the file. For example:

# standard model
model = models.unet_model.UNet(1, 4)

# dropout model
# model = models.unet_model.UNet_Dropout(1, 4, p=0.1)

Evaluation

To evaluate the models, use the jupyter notebooks.