Interactive Music Prediction System (IMPS)
IMPS is a system for predicting musical control data in live performance. It uses a mixture density recurrent neural network (MDRNN) to observe control inputs over multiple time steps, predicting the next value of each step, and the time that expects the next value to occur. It provides an input and output interface over OSC and can work with musical interfaces with any number of real valued inputs (we’ve tried from 1-8). Several interactive paradigms are supported for call-response improvisation, as well as independent operation, and “filtering” of the performer’s input. Whenever you use IMPS, your input data is logged to build up a training corpus and a script is provided to train new versions of your model.
Here’s a demonstration video showing how IMPS can be used with different musical interfaces.
IMPS is written in Python with Keras and TensorFlow Probability, so it should work on any platform where Tensorflow can be installed. Python 3 is required.
First you should clone this repository or download it to your computer:
git clone https://github.com/cpmpercussion/imps.git cd imps
The python requirements can be installed as follows:
pip install -r requirements.txt
Some people like to keep Python packages separate in virtual environments, if that’s you, here’s some terminal commands to install:
virtualenv --system-site-packages -p python3 venv source venv/bin/activate pip install -r requirements.txt
How to use
There are four steps for using IMPS. First, you’ll need to setup your musical interface to send it OSC data and receive predictions the same way. Then you can log data, train the MDRNN, and make predictions using our provided scripts.
1. Connect music interface and synthesis software
- A music interface that can output data as OSC.
- Some synthesiser software that can take OSC as input.
These could be the same piece of software or hardware!
You need to decide on the number of inputs (or dimension) for your predictive model. This is the number of continuous outputs from your interface plus one (for time). So for an interface with 8 faders, the dimension will be 9.
Now you need your music interface to send messages to IMPS over OSC. The default address for IMPS is: localhost:5001. The messages to IMPS should have the OSC address
/interface, and then a float between 0 and 1 for each continuous output on your interface, e.g.:
/interface 0 0.5 0.23 0.87 0.9 0.7 0.45 0.654
For an 8-dimensional interface.
Your synthesiser software or interface needs to listen for messages from the IMPS system as well. These have the same format with the OSC address
/prediction. You can interpret these as interactions predicted to occur right when the message is sent.
Here’s an example diagram for our 8-controller example, the xtouch mini controller.
In this example we’ve used Pd to connect the xtouch mini to IMPS and to synthesis sounds. Our Pd mapping patch takes data from the xtouch and sends
/interface OSC messages to IMPS, it also receives
/prediction OSC message back from IMPS whenever they occur. Of course, whenever the user performs with the controller, the mapping patch sends commands to the synthesiser patch to make sound. Whenever
/prediction messages are received, these also trigger changes in the synth patch, and we also send MIDI messages back to the xtouch controller to update its lights so that the performer knows what IMPS is predicting.
So what happens if IMPS and the performer play at the same time? In this example, it doesn’t make sense for both to control the synthesiser at the same time, so we set IMPS to run in “call and response” mode, so that it only makes predictions when the human has stopped performing. We could also set up our mapping patch to use prediction messages for a different synth and use one of the simultaneous performance modes of IMPS.
2. Log some training data
You use the
predictive_music_model command to log training data. If your interface has N inputs the dimension is N+1:
python predictive_music_model.py --dimension=(N+1) --log
This command creates files in the
logs directory with data like this:
2019-01-17T12:37:38.109979,interface,0.3359375,0.296875,0.5078125 2019-01-17T12:37:38.137938,interface,0.359375,0.296875,0.53125 2019-01-17T12:37:38.160842,interface,0.375,0.3046875,0.1953125
These CSV files have the format: timestamp, source of message (interface or rnn), x_1, x_2, …, x_N.
You can log training data without using the RNN with the
o switch (user only) if you like, or use a partially trained RNN and then collect more data.
python predictive_music_model.py --dimension=(N+1) --log -o
Every time you run the
predictive_music_model, a new log file is created so that you can build up a significant dataset!
3. Train an MDRNN
There’s two steps for training: Generate a dataset file, and train the predictive model.
python generate_dataset --dimension=(N+1)
This command collates all logs of dimension N+1 from the logs directory and saves the data in a compressed
.npz file in the datasets directory. It will also print out some information about your dataset, in particular the total number of individual interactions. To have a useful dataset, it’s good to start with more than 10,000 individual interactions but YMMV.
To train the model, use the
train_predictive_music_model command—this can take a while on a normal computer, so be prepared to let your computer sit and think for a few hours! You’ll have to decide what size model to try to train:
xl. The size refers to the number of LSTM units in each layer of your model and roughly corresponds to “learning capacity” at a cost of slower training and predictions.
It’s a good idea to start with an
s model, and the larger models are more relevant for quite large datasets (e.g., >1M individual interactions).
python train_predictive_music_model.py --dimension=(N+1) --modelsize=xs --earlystopping
It’s a good idea to use the “earlystopping” parameter to stop training after the model stops improving for 10 epochs.
4. Perform with your predictive model
Now that you have a trained model, you can run this command to start making predictions:
python predictive_music_model.py -d=(N+1) --modelsize=xs --log -c
--log switch logs all of your interactions as well as predictions for later re-training. (The dataset generator filters out RNN records so that you only train on human sourced data).
PS: the three scripts respond to the
--help switch to show command line options. If there’s something not documented or working, it would be great if you add an issue above to let me know or get in touch on Twitter at @cpmpercussion.
More about Mixture Density Recurrent Neural Networks
IMPS uses a mixture density recurrent neural network MDRNN to make predictions. This machine learning architecture is set up to predict the next in a sequence of multi-valued elements. The recurrent neural network uses LSTM units to remember information about past inputs and use this to help make decisions. The mixture density model at the end of the network allows continuous multi-valued elements to be sampled from a rich probability distribution.
The network is illustrated here—every time IMPS receives an interaction message from your interface, it is sent to thorugh the LSTM layers to produce the parameters of a Gaussian mixture model. The predicted next interaction is sampled from this probability model.
The MDRNN is written in Keras and uses the keras-mdn-layer package. There’s more info and tutorials about MDNs on that github repo.