G2Net

This project involves training and predicting with a machine learning model using noisy signal data. The steps below outline the required environment setup, dependencies, and instructions for data preparation, model training, and result prediction.

Environment Setup

Ensure the following tools are installed and configured in your environment:

• CUDA
• Apex

Required Packages

Install the following Python packages before running the scripts:

pip install timm numpy omegaconf pandas pyfstat pytorch_lightning scikit_learn torch tqdm wandb

Run Instructions

Step 0: Download Raw Data

Download the raw data (approximately 200GB). This may take some time, so please be patient.

Step 1: Generate Signal Images with Noise

Run the script to generate signal images by adding noise to clean signals.

python scripts/simulate_signals.py resources/competition/timestamps.pkl

Step 2: Combine Gaussian Noise with Pure Signals

Generate random Gaussian background noise and combine it with pure signals.

python scripts/synthesize_external_psds.py resources/external/train/signals

Step 3: Convert HDF5 Data Format

Convert HDF5 data files to the required input format for the model.

python extract_psds_from_hdf5.py ../input/train/test_hdf5_directory

Step 4: Train the Model

Train the machine learning model using the specified configuration file.

python src/train.py config/convnext_small_in22ft1k.yaml

Step 5: Predict Results

Use the trained model to generate predictions.

python src/predict.py convnext_small_in22ft1k-6f6648-last.pt --use-flip-tta

This will produce submission1.csv.

Step 6: Train and Predict with Model 2

Train a second model and generate additional predictions.

python src/g2net-augmentation.py

This will produce submission2.csv.

Step 7: Model Ensemble

Combine the results from both models for final submission.

python src/combine.py

This will generate the final submission.csv.