Deep Learning Project for Playlist Recommendation based on a User's current session

Dependencies/ Uses:

• Librosa
• Numpy
• SKLearn
• Theano
• Lasagne
• NoLearn

System:

• Convolutional Neural Network (CNN) - Deep Learning for Feature Extraction.
  • 3 ConvLayers and 2 FC-Layers.
• Support Vector Machine (SVM) - Machine Learning for Classification.
  • Linear, SVC.

Algorithm/ Data Flow:

• User provides path to music folder/ library.
• System retrieves all music files that end in .mp3 and stores in array.
• System randomly seperates the music files array into training and testing arrays.
• System "plays"/ displays individual songs to the user.
• User labels each song as "liked" or "disliked".
• System reads in each song, picks a 20s segment out 60s into the song and decodes .mp3 files to .wav.
• System uses the .wav of each file to represent into corresponding Mel Spectrograms.
• System scales the spectrograms by amplitude.
• System proceeds to slice the spectrograms up into smaller pieces of the spectrogram for data augmentation purposes.
• System defines the CNN parameters and proceeds to train the model using the training set.
• System uses Theano to extract features from each spectrogram slice passed through the model for both sets.
• System appends all the spectrogram slice's features together to represent that music file and reshapes the features accordingly.
• System trains SVM model with extracted features and known labels from user input from the training set.
• System iterates testing set and attempts to predict by classifying each testing song's features as "liked" or "disliked".