This GitHub organization showcases research projects on Unsupervised Domain Adaptation (UDA) conducted during a research internship at the Indian Institute of Technology (IIT) Mandi under the School of Computing and Electrical Engineering. Our work focuses on advancing domain adaptation techniques for audio scene classification, with applications of gradient reversal layers and cycle self-training methodologies.

Our team worked under the guidance of Dr. Padmanabhan Rajan and Abhishek Dileep (Research Scholar), exploring approaches to domain adaptation challenges in acoustic scene classification. The primary research areas include:

• Unsupervised Domain Adaptation techniques for cross-device audio classification
• Audio scene classification using transformer models
• Transfer learning across different recording devices and environments
• Feature extraction methods using pre-trained audio transformers (PaSST)
• Cross-device domain adaptation for real-world acoustic applications

First Project: Implementation of unsupervised domain adaptation using Gradient Reversal Layer (GRL) with PaSST feature extractors for cross-device acoustic scene classification.

Key Features:

• Domain-Adversarial Neural Networks (DANN) with gradient reversal layers
• PaSST Integration: Pre-trained audio transformers as feature extractors
• Cross-Device Evaluation: Device A → Devices B,C,S1-S6 adaptation
• Computer Vision Validation: SVHN→MNIST implementation for method verification

Technical Highlights:

• Architecture: PaSST (768-dim) → 768→512→256 adaptation layers → 10-class classifier
• Dataset: DCASE TAU 2020 Mobile (64 hours, 10 acoustic scenes)
• Results: 70.37% overall accuracy, 22.76% improvement over source-only training
• Innovation: Novel combination of audio transformers with adversarial domain adaptation

Technologies:

• Python, PyTorch, Jupyter Notebooks
• PaSST (Patchout Audio Spectrogram Transformer)
• DCASE TAU 2020 dataset
• Gradient Reversal Layer implementation

Second Project: Implementation of Unsupervised Domain Adaptation using Cycle Self-Training (CST) with enhanced loss functions and consistency regularization.

Key Features:

• Cycle Self-Training (CST) with source-target-source consistency
• Enhanced Loss Function: Combined Tsallis entropy, CST loss, and FixMatch
• FixMatch Integration: Consistency regularization with strong/weak augmentations
• SAM Optimizer: Sharpness-Aware Minimization for improved generalization

Technical Highlights:

• Architecture: PaSST + CST framework with multi-component loss
• Loss Components: CE + Tsallis (λ₁=0.1) + CST (λ₂=0.3) + FixMatch (λ₃=0.4)
• Results: 66.57% overall accuracy with improved target domain performance
• Innovation: First application of cycle self-training to audio domain adaptation

Technologies:

• Python, PyTorch, Jupyter Notebooks
• Cycle Self-Training methodology
• FixMatch consistency regularization
• SAM (Sharpness-Aware Minimization) optimizer
• Audio-specific augmentations and transformations

Our research team consists of two groups working collaboratively under expert supervision:

• Ronn Mathew Sino
• Noah Thomas

• K J Theophene Xavier Lynn
• Advaith S Menon

• Dr. Padmanabhan Rajan - Associate Professor, School of Computing and Electrical Engineering, IIT Mandi
• Abhishek Dileep - Research Scholar, IIT Mandi
• Dr. Divya James - Associate Professor, Rajagiri School of Engineering & Technology (Advisor)

• Python - Primary programming language
• PyTorch - Deep learning framework
• Jupyter Notebooks - Interactive development environment
• NumPy, Pandas - Data processing and analysis
• Librosa - Audio processing and feature extraction

• PaSST - Pre-trained audio transformers
• DALib - Domain adaptation algorithms
• Custom Implementations - GRL, CST, and supporting utilities

• DCASE TAU 2020 - Primary evaluation dataset
• Audio Transformations - Augmentation and preprocessing pipelines
• Cross-Device Evaluation - Multi-device testing framework

For inquiries about our research, collaboration opportunities, or technical questions:

• Ronn Mathew Sino: GitHub Profile
• Noah Thomas: GitHub Profile
• K J Theophene Xavier Lynn: GitHub Profile

• Dr. Padmanabhan Rajan - Associate Professor, IIT Mandi
• School of Computing and Electrical Engineering, IIT Mandi

• Institution: Indian Institute of Technology (IIT) Mandi
• Department: School of Computing and Electrical Engineering
• Research Focus: Machine Learning, Domain Adaptation, Audio Signal Processing

We express our sincere gratitude to:

• IIT Mandi and the School of Computing and Electrical Engineering for providing research opportunities and computational resources
• Dr. Padmanabhan Rajan for expert guidance and supervision throughout the internship
• Abhishek Dileep for mentorship and technical insights
• DCASE Community for providing high-quality datasets and evaluation frameworks
• Open Source Community for foundational implementations (DANN-pytorch, PaSST, CST)

These projects are developed for academic and research purposes. Please refer to individual repository licenses:

• Research use: Freely available for academic research
• Commercial use: Requires permission and proper attribution
• Dataset usage: Subject to DCASE TAU 2020 license terms

Last Updated: June 2025 | Projects: 3 Completed