DL and GenAI Project - Messy Mashup

Music genre classification under noisy mixing conditions using the Messy Mashup competition dataset.

Competition

Jan 2026 DLGenAI Project - Messy Mashup
https://www.kaggle.com/competitions/jan-2026-dl-gen-ai-project/

The challenge is to classify music genres from mashups created by mixing instrument stems from different songs of the same genre, with added environmental noise.

Genres: blues, classical, country, disco, hiphop, jazz, metal, pop, reggae, rock

Repository Structure

├── notebooks/
│   ├── milestone-1.ipynb          # Milestone 1: EDA and baseline
│   └── kaggle_notebook.ipynb      # Kaggle submission notebook
├── src/
│   ├── train.py
│   ├── inference.py
│   └── utils.py
├── reports/
├── models/
├── requirements.txt
├── .gitignore
└── README.md

Milestones

Milestone 1 - EDA and Baseline

• Dataset exploration and validation (completeness, corruption, size analysis)
• Silence detection across training stems
• Audio stem mixing and peak normalization
• Baseline heuristic methods using librosa

Dataset

The dataset is available on Kaggle and contains:

• genres_stems/ - 10 genres x 100 songs x 4 stems (drums, vocals, bass, others)
• ESC-50-master/ - Environmental noise dataset
• mashups/ - 3,020 test audio mashups
• test.csv - Test index file
• sample_submission.csv - Submission template

Milestone 2 - Dataset & Baseline

• Synthetic mashup generation from genre stems
• ESC-50 noise injection pipeline
• Mel-spectrogram feature extraction
• Baseline model evaluation

Milestone 3 - CNN Model

• Custom CNN architecture (4 conv blocks)
• Training with Adam optimizer and StepLR scheduler
• WandB experiment tracking

Milestone 4 - Transfer Learning & RNN

• ResNet18 fine-tuning on mel-spectrograms
• CRNN with bidirectional GRU for temporal features
• Model ensemble with weighted averaging

Notebooks

Notebook	Description
01_eda.ipynb	Dataset exploration and visualization
02_data_generation.ipynb	Synthetic mashup generation
03_model_cnn.ipynb	CNN from scratch
04_model_resnet.ipynb	ResNet18 fine-tuning
05_model_crnn.ipynb	CRNN with bidirectional GRU
06_ensemble.ipynb	Model ensemble and comparison
kaggle_notebook.ipynb	Final Kaggle submission

Model Performance

(Run notebooks/06_ensemble.ipynb to generate final metrics and replace this section)

Model	Macro F1	Accuracy
GenreCNN	TBD	TBD
ResNet18	TBD	TBD
GenreCRNN	TBD	TBD

CRNN typically performs best due to its ability to capture temporal dynamics.