IMPSY 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 IMPSY, 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 IMPSY can be used with different musical interfaces:
IMPSY 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 and we use Poetry for managing dependencies. IMPSY currently relies on Python 3.11, TensorFlow 2.15.0, TensorFlow Probability 0.23.0, and keras-mdn-layer 0.3.0. You can see the dependencies in pyproject.toml.
To install IMPSY, first ensure that you have a Python 3.11 installation available, then install Poetry. The poetry install instructions vary depending on your preferences for a python setup this is likely to work on Linux, macOS or Windows (WSL):
curl -sSL https://install.python-poetry.org | python3 -
Then you should clone this repository or download it to your computer:
git clone https://github.com/cpmpercussion/impsy.git
cd impsy
Then you can install the dependencies using Poetry:
poetry install
Finally, you can test that IMPSY works:
poetry run ./start_impsy.py --help
There are four steps for using IMPSY. 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.
You'll need:
- 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 IMPSY over OSC. The default address for IMPSY is: localhost:5001. The messages to IMPSY 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 IMPSY 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 IMPSY and to synthesis sounds. Our Pd mapping patch takes data from the xtouch and sends /interface OSC messages to IMPSY, it also receives /prediction OSC message back from IMPSY 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 IMPSY is predicting.
So what happens if IMPSY 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 IMPSY 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.
You use the run command to log training data. If your interface has N inputs the dimension is N+1:
poetry run ./start_impsy run --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 mode option to select "user" if you like, or use a partially trained RNN and then collect more data.
poetry run ./start_impsy run --dimension (N+1) --log -mode user
Every time you use IMPS' "run" command, a new log file is created so that you can build up a significant dataset!
There's two steps for training: Generate a dataset file, and train the predictive model.
Use the dataset command:
poetry run ./start_impsy 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 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: xs, s, m, l, 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 xs or s model, and the larger models are more relevant for quite large datasets (e.g., >1M individual interactions).
poetry run ./start_impsy train --dimension (N+1) --modelsize s
It's a good idea to use the earlystopping option to stop training after the model stops improving for 10 epochs.
Now that you have a trained model, you can run this command to start making predictions:
poetry run ./start_impsy run --dimension (N+1) --modelsize xs --log
The --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: all the IMPSY commands 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.
IMPSY 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 IMPSY 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.