When I first heard about magenta and machine learning for musical generation I was very excited. Soon after I realized that if I wanted to train my own model I was going to have to get a better computer and plan on doing a lot of web scraping of midi files…

Then I found out they have pre-trained models that we can play with. OK… not as cool as training your own but still interesting. So I thought I would git it a try:

First we need to run these commands to set up the environment:

sudo apt-get install build-essential libasound2-dev libjack-dev portaudio19-dev
virtualenv magenta
source magenta/bin/activate
pip3 install magenta

Ok now that we have our environment set up we can generate some melodies right? No! We need to download some pre-trained models from magenta. I got the basic_rnn, the loockback_rnn and the attention_rnn. Download the ones you want to use.

Now you can generate some basic melodies. Try the following:

melody_rnn_generate --config=basic_rnn --bundle_file=basic_rnn.mag --output_dir=/path/to/output/dir --num_outputs=10 --num_steps=128 --primer_melody="[60]"

This primes magenta with the midi note 60 which is middle C. This is not meant to be a complete guide so play with the other parameters and check the Docs for more info. This should at least get you started though. You can read more about RNN (Reccurent Neural Nets) here and here

The lookback model sounds a little better in my opinion:

melody_rnn_generate --config=lookback_rnn --bundle_file=lookback_rnn.mag --output_dir=~/Documents/magentaTutorial/generated --num_outputs=10 --num_steps=128 --primer_melody="[60]"

But the attention model sounds the best:

melody_rnn_generate --config=attention_rnn --bundle_file=attention_rnn.mag --output_dir=~/Documents/magentaTutorial/generated --num_outputs=10 --num_steps=128 --primer_melody="[60]"

You can also feed it a midi file for priming:

melody_rnn_generate --config=attention_rnn --bundle_file=attention_rnn.mag --output_dir=~/Documents/magentaTutorial/generated --num_outputs=10 --num_steps=128 --primer_midi=/path/to/midifile.mid

You can also check out my git hub repo with all the pretrained models and sample commands here