1st Place Solution

This repository contains code for the 1st place solution of the BYU - Locating Bacterial Flagellar Motors 2025 Competition here.

Solution Summary: here

My solution uses a 3D U-Net trained with heavy augmentations and auxiliary loss functions. During inference, I rank each tomogram based on the max predicted pixel value and use quantile thresholding to determine if a motor is present. Please read the solution summary for more details.

Setup

This section covers how to reproduce model training. The hardware requirements below serve as a guideline and can be adjusted based on available resources. Some decent cloud options are Lambda Labs, Runpod, and Paperspace.

Hardware Requirements

Component	Recommended
OS	Ubuntu 22.04
RAM	≥ 32 GB
Disk Space	≥ 200 GB
CPU Cores	≥ 8
CUDA Version	12.4
GPU	NVIDIA A100 (80GB)

Miniconda Install

1. Get bash script for miniconda download from here

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

2. Exectute setup script

chmod u+x ./Miniconda3-latest-Linux-x86_64.sh
./Miniconda3-latest-Linux-x86_64.sh

Miniconda Environment Setup

1. Set channels

conda config --set channel_priority flexible
conda config --remove channels defaults
conda config --add channels conda-forge
conda config --add channels nvidia
conda config --add channels pytorch
conda config --show channels

2. Create environment

conda create --prefix ../byu_env python=3.10
conda activate ../byu_env/

3. Install packages

# Conda packages
conda install pytorch==2.5.1 torchaudio==2.5.1 torchvision==0.20.1 pytorch-cuda==11.8 monai==1.4.0 wandb==0.19.6 tqdm==4.67.1

# Pip packages
pip install -r requirements.txt

Directory Structure

Before training, you will need to manually create the data directory.

1. The raw competition data should be placed under ./data/raw/.

2. The external tomograms from here should be placed under ./data/processed/fold_-100/.

3. r3d18_KM_200ep.pt, r3d200_KM_200ep.pt and folds_all.csv can be downloaded from here and placed accordingly.

The directory structure should be as follows:

./data/
├── checkpoints/
├── model_zoo/
│   ├── r3d18_KM_200ep.pt
│   └── r3d200_KM_200ep.pt
├── processed/
│   ├── fold_-100/
│   │   ├── aba2013-04-06-7.npy
│   │   ├── aba2013-04-06-8.npy
│   │   └── ...
│   └── folds_all.csv
├── raw/
│   ├── test/
│   └── train/
└── sample_submission.csv
└── train_labels.csv

Preprocessing

Once everything is setup correctly, you can process the competition data. This will populate the ./data/processed/ directory.

python -m src.pre.run

Training

To train a model, run the following commands. Each model takes 35-40 hours to train. You can train for 250 epochs and get the same performance.

You can repeat this step to train multiple models for an ensemble.

# Make executable
chmod u+x run.sh

# Run in background
nohup ./run.sh > nohup.out &