audio-classification-task1

Audio Classification Project

A comprehensive audio classification pipeline built using PyTorch, integrating modern deep‑learning techniques such as transfer learning, mel‑spectrogram feature extraction, data augmentation, SpecAugment, early stopping, and ResNet‑18 fine‑tuning.

This project is based on the FreeCodeCamp PyTorch Tutorial by omaratef3221, with several enhancements inspired by the following resources:

• https://github.com/omaratef3221/pytorch_tutorials
• Transfer learning with ResNet‑18: https://ngaif.com/guide-to-transfer-learning-with-pytorch/
• SpecAugment and audio augmentation: https://docs.pytorch.org/audio/main/tutorials/audio_feature_augmentation_tutorial.html
• Early stopping reference: https://pythonguides.com/pytorch-early-stopping/

This README explains the project structure, dataset pipeline, model design, training loop, techniques used, and reasons for choosing them.

 Project Overview

This project performs multi‑class audio classification using deep learning. It converts raw audio into log‑mel‑spectrograms, applies augmentations, and trains a ResNet‑18 model (pretrained on ImageNet) to classify environmental sounds.

The workflow includes:

1. Dataset loading & preprocessing
2. Mel‑spectrogram extraction
3. Log1p compression
4. SpecAugment techniques
5. Model architecture (ResNet‑18 transfer learning)
6. Training/validation loops
7. Early stopping & best‑model saving
8. Evaluation & inference

---

---

Features & Techniques Used

1. Audio Preprocessing (Mel‑Spectrograms)

• Converts audio to 80‑dim mel‑spectrograms using librosa

• Parameters used:

  • n_fft = 1024
  • hop_length = 256
  • win_length = 1024

Reason: Mel‑spectrograms resemble how humans perceive sound and are the standard input for audio deep‑learning models.

---

2. Log1p Compression

log_mel = torch.log1p(mel)

Reason: Reduces dynamic range of spectrograms and stabilizes training.

---

3. SpecAugment (Time & Frequency Masking)

Added using PyTorch's official API:

• TimeMasking
• FrequencyMasking

Reason: Improves model robustness against noise & overfitting by simulating real‑world variations.

---

4. Transfer Learning with ResNet‑18

• Loaded pretrained ResNet‑18
• Modified first Conv2D layer to accept 1‑channel inputs
• Replaced final FC layer with nn.Linear(num_features, num_classes)

Reason: ResNet‑18 learns high‑level features efficiently and performs extremely well on spectrogram images.

---

5. Early Stopping

Monitors validation loss, stops when no improvement after patience epochs.

Reason: Prevents overfitting and reduces unnecessary training time.

---

6. Custom Dataset Class

Handles:

• Loading audio
• Computing mel‑spectrograms
• Padding/cropping to fixed frame length
• Applying SpecAugment

Reason: Provides complete control over preprocessing, making the pipeline reusable.

---

7. Training Loop

Includes:

• Learning rate scheduling
• Loss monitoring
• Model saving

Reason: Smooth and stable optimization.

---

 Training

• Optimizer: AdamW
• Loss: CrossEntropyLoss
• Batch size : 16
• Epochs: 25

---

References

1. FreeCodeCamp PyTorch tutorial (base code): https://github.com/omaratef3221/pytorch_tutorials

2. Transfer Learning with ResNet‑18: https://ngaif.com/guide-to-transfer-learning-with-pytorch/

3. SpecAugment documentation:

   • https://docs.pytorch.org/audio/main/tutorials/audio_feature_augmentation_tutorial.html

4. Early Stopping guide:

   • https://pythonguides.com/pytorch-early-stopping/