Audio Classification Project README

Overview

This project involves the classification of musical instruments in audio files using a Convolutional Neural Network (CNN) trained on mel spectrograms. The project is organized into several components:

1. Data Loader: datav2.py contains the data loader class responsible for loading audio files.

2. CNN Architecture: modelv2.py includes the CNN architecture used for instrument classification.

3. Training: trainv2.py is where the model was trained using actual labels as targets and mel spectrograms as inputs. The trained model is saved as feedforwardnet.pth.

4. Inference: inference.py evaluates the trained model and generates inferences. The expected and predicted values are mapped in a file called inference_results.csv. An updated annotations file (updated_metadata.csv) containing pseudo labels is created based on these inferences.

5. Training with Pseudo Labels: train_pseudo.py involves training the model again, but this time using the pseudo labels obtained from the previous step. The trained model is saved as pseudo_model.pth.

6. Inference with Pseudo Labels: inference_pseudo.py evaluates the model trained with pseudo labels, and the inferences are compared against both actual labels and pseudo labels. The accuracy is recorded for both scenarios.

##datset - https://zenodo.org/records/3685367

Project Structure

The project directory structure is as follows:

- datav2.py
- modelv2.py
- trainv2.py
- inference.py
- updated_metadata.csv
- train_pseudo.py
- pseudo_model.pth
- inference_pseudo.py
- README.md