YouTube cantopop title parser

We can download a YouTube video with youtube_dl. I usually do this to collect cantopop in MP3 format but the issue will be the id3 tags.

This is a script to train a MLP to figure out the artist and song title from the YouTube video title. I hand crafted the features (should try word2vec but I did not) and feed into a simple 3-layer MLP to identify tokens.

The training data is in titles.txt and I used crawler-dbm.py to preprocess the data into feat.pickle. Then running train.py will train a MLP (using scikit-learn) for the purpose, which is then saved as mlp-trained.pickle.

When we have a trained model, we can tag all MP3 files based on their filename (as if youtube_dl give you by default):

python deploy.py [files...]

The ID3v2 access is using mutagen library.

An alternative version is built using Keras/tensorflow as well: train_keras.py and deploy_keras.py. The model of MLP is same as scikit-learn but the code is slower to initialize.