For windows runs

Its best to switch into wsl for faster runs (23->3min/epoch, 0.05->0.5it/s).

• Sanity check hanged,
• when disabled epoch bar never showed, it just loaded infinitely.
• Commenting out num_workers helped: https://stackoverflow.com/a/70949425/4404911)
• but made runs much slower

Running in windows may still be possible by commenting lines with num_workers=num_workers, and persistent_workers=True but no guarantees.

See instructions for WSL:

wsl --install -d Ubuntu-24.04

Get WSL compatible driver:

wsl -d Ubuntu-24.04

• sudo apt install python3.12-venv python3-pip ffmpeg
• Copy data folder
• Symlink main, .env, requirements

To run the project

Download dataset from BirdCLEF 2024, unzip it and place the files into ./data

Copy .env.template and rename it to .env and fill out the values for your wandb account

First run data_exploration.ipynb to

• view data analysis
• trim dataset based on data analysis

Then run main.ipynb to

• create spectrograms (and view them) from the remaining audios
• train on them
• run evaluations
• see application

To run on Komondor HPC

Set is_gradio_on to False in main.ipynb

Build sif on your computer

• singularity build --fakeroot Singularity.sif Singularity.def

Copy files to HPC & Run

• scp .sbatch komondor:~/nr_haml2025_scratch
• scp Singularity.sif komondor:~/nr_haml2025_scratch

Run on HPC:

• mkdir ~/models
• chmod u+x .sbatch
• sbatch .sbatch