Music Classification

Using the GTZAN dataset (10 genres; audio files, precomputed mel spectrogram images and extracted features), the contestant classifies music by genre from the spectrogram images. The notebook downloads the data, demonstrates waveform and spectrogram plots, splits each genre into 20 test images and the rest for training, and provides a baseline CNN and plotting helpers.

Exercise 1: plot one spectrogram per genre. Exercise 2: train a CNN from scratch reaching at least 75% accuracy on the given test set and describe the settings used. Exercise 3: retrain with an augmentation method proposed by Park et al. (2019) (SpecAugment) and compare. Exercise 4: fine-tune an ImageNet-pretrained DenseNet or ResNet, following Palanisamy et al. (2020). Exercise 5: extract the inputs of the last dense layer as features, project them to 2D and plot them coloured by genre.

Exercise 6: quantify genre similarity with cosine similarity between feature vectors, optionally proposing a better measure. Exercise 7: visualise an arbitrary test image with integrated gradients (Sundararajan et al., 2017) beside the original. Exercise 8: train or fine-tune a Vision Transformer and try to beat the CNN of Exercise 2. Exercise 9: train or fine-tune a generative model on the spectrograms, generate new spectrograms and convert them back to audio.