Skip to content

Latest commit

 

History

History

matterhorn

README.md

🏆 Matternhorn Approach 🏆

📋 Overview

This repository contains a modularized script that our team used to solve the Bone Challenge. Our approach involved using a deep learning model based on the EfficientNet family of models to predict the growth plane of a bone using NIfTI (nii) files. The project includes customizable parameters for various hyperparameters, allowing for different approaches to solving the problem.

🌟 Features

  • EfficientNet Models: Utilizes the EfficientNet family for building the deep learning models.
  • Parametrization: Allows for the customization of hyperparameters to explore various configurations.
  • Reproducible Research: Designed with reproducible research in mind, tracking all runs in a run_history.json file. Each run is uniquely identified by a run_id and can only be executed on a committed git repository, ensuring that the script used and all hyperparameters are saved for each run_id.

🛠️ Installation

  1. Create and activate a python environment.
  2. Run pip install -r requirements.txt

💡 Usage

To train the model, run the script with the desired parameters. Here is an example command:

python train_model.py --param1 value1 --param2 value2

⚙️ Parameters

The following parameters can be used to customize the training process. Default are the ones that we used in our final submission.

  • --num-folds: int (default: 5) Number of folds for cross-validation.

  • --fold: int (default: None) If specified, run only one fold. Provide the fold number.

  • --epochs: int (default: 2000) Total number of epochs for training.

  • --central-prop: float (default: 0.40) Central proportion for training data.

  • --central-prop-val: float (default: 0.40) Central proportion for validation data.

  • --batch-size: int (default: 16) Size of the mini-batch used in training.

  • --nr-val-eval: int (default: 10) Number of validation evaluations per epoch.

  • --lr: float (default: 3e-4) Learning rate for the optimizer.

  • --model-type: str (choices: "B1", "B2", "B3", "B4", "B5", "B6", "B7", default: "B3") Type of the model to use for training.

  • --summary-path: str Path to the CSV file where training summaries will be saved.

  • --nii-path: str Path to the NII files used for training.

  • --seed: int (default: 42) Random seed for reproducibility.

  • --train-mode: str (choices: "full", "half", default: "half") Training precision mode: full or half precision.

  • --zaxis-mode: str (choices: "crop", "interpolate", "mixed", default: "mixed") Mode for z-axis handling: Crop, Interpolate, or Mixed.

  • --planes-index-selection: str (choices: "random", "consecutive", default: "random") Method for selecting planes index: Random or Consecutive.

  • --num-input-channels: int (default: 9) Number of input channels for the network.

  • --augmented-dataset-root-dir: str Root directory for the augmented dataset.

  • --use-existing-augmented-dir: bool (default: False) Whether to use existing augmented datasets. If set, existing datasets will be used; otherwise, new ones will be generated.

  • --augmentation-train-count: int (default: 150) Number of augmentations per file for training.

  • --augmentation-valid-count: int (default: 10) Number of augmentations per file for validation.

  • --num-workers: int (default: 4) Number of workers for the DataLoader.

Test reproducible

Calculate the predictions

#!/bin/bash

# Define variables
MODEL_PATH=""
NII_PATH=""

MODEL_FILES=(
    "models/fold_0_epoch_1978_model_best_val_score_0.591_val_loss_10.1503.pt"
    "models/fold_1_epoch_1710_model_best_val_score_0.566_val_loss_11.6527.pt"
    "models/fold_2_epoch_806_model_best_val_score_0.602_val_loss_17.8721.pt"
    "models/fold_3_epoch_887_model_best_val_score_0.553_val_loss_23.7504.pt"
    "models/fold_4_epoch_1844_model_best_val_score_0.586_val_loss_16.3064.pt"
)

# Loop through each model file and run the prediction
for MODEL_FILE in "${MODEL_FILES[@]}"; do
    python predict.py --model-file "$MODEL_FILE" --nii-path "$NII_PATH" --nr-val-eval 3
done

Join the results

python join.py --output results/0.csv --stat median pred/2bed35d1-028e-463f-b0c3-44c6da60216d/
python join.py --output results/1.csv --stat median pred/357db53d-67ba-4672-9bad-934a478a682c/
python join.py --output results/2.csv --stat median pred/35a940e5-3344-4c14-a001-d5c4675717e0/
python join.py --output results/3.csv --stat median pred/f34d1167-3b2a-4b60-907f-429c7fbc391e/
python join.py --output results/4.csv --stat median pred/b69c6e0f-a68c-4d94-832f-30e9cf58a7c2/
python join.py --output results.csv --stat mean results/
cat y_true_test.csv | cut -d , -f 3,4 > y_true_test_only_preds.csv
python evaluate.py results.csv y_true_test_only_preds.csv

Test results

The output obtained from test set evaluation using B3-300 is:

Mean Score: 0.682 Standard Deviation of Score: 0.334 Sum of Scores: 8.87

Arch Size Mean scr Std scr Fold 1 Fold 2 Fold 3 Fold 4 Fold 5
B0 300 0.578 0.017 0.605 0.577 0.580 0.559 0.570
B1 300 0.604 0.053 0.667 0.616 0.585 0.626 0.526
B2 300 0.587 0.022 0.573 0.560 0.611 0.581 0.609
B3 300 0.580 0.020 0.591 0.566 0.602 0.553 0.586
B4 300 0.590 0.033 0.629 0.569 0.564 0.622 0.565
B4 380 0.585 0.046 0.589 0.623 0.514 0.574 0.627