This project implements a deep learning pipeline for classifying environmental sounds from the ESC-50 dataset. It features a custom convolutional neural network (AutoCNN) built with PyTorch, using residual blocks for improved learning. The solution includes data augmentation, training on GPU using Modal, and a scalable inference API.

ML/Backend:

• PyTorch
• Torchaudio
• Modal (GPU-accelerated training & inference)
• FastAPI (for serving inference endpoints)

Client (optional visualization layer):

• Next.js
• Tailwind CSS
• TypeScript
• Shadcn UI

✅ Custom CNN architecture with residual connections
✅ ESC-50 dataset ingestion and preprocessing
✅ Data augmentation with Mixup and spectrogram masking
✅ Fully managed training on GPU using Modal
✅ Inference API with real-time audio classification
✅ Intermediate feature map visualization for debugging and interpretability
✅ Example endpoint for testing predictions from WAV files

# Modal API Key
NEXT_PUBLIC_MODAL_API=

This project uses the ESC-50 dataset, which contains 50 environmental sound categories. The training pipeline automatically downloads and prepares the dataset during Modal app initialization.

# Clone the repo
git clone https://github.com/JoelDeonDsouza/Auto_CNN.git
cd auto-cnn

# Install dependencies
pip install -r requirements.txt

You can launch a training job on Modal with GPU acceleration.

modal run train.py

The trained model will be saved in a Modal-managed volume.

Deploy the inference API:

modal deploy main.py

Test an inference request locally:

modal run main.py

Put your WAV files in the audio-tests/ directory. An example (chirpingBirds.wav) is included for testing.

• AutoCNN: Custom CNN model with residual blocks
• ESC50Dataset: PyTorch Dataset class for ESC-50
• train.py: Training loop with augmentation, optimizer, and TensorBoard logging
• main.py: FastAPI-powered inference endpoint
• modal: Manages GPU workloads and deploys endpoints

The AutoCNN model follows a ResNet-inspired structure with four convolutional stages, followed by global pooling and a linear classifier.