deep-jammer | Deep Jammer is a deep neural network

To set up Deep Jamer, we just need to do a few steps:.
First, make sure that you're using Python2.7 or else you'll get a lot of errors. Deep Jammer doesn't work with Python3 since we don't use print as a function (and probably other stuff). Yup, we totally regret that decision.
Since the Keras version trains very slowly for some reason, let's stick to the Theano version. I'm honestly not even sure if the Keras version is in a working state. We haven't touched it in a while. From the root directory of the project, run the following command:
Before we can run Deep Jammer, we need to install a few Python packages, some of which you may already have. To get every package, we just run the command below. While we're pretty sure that we've included everything, let us know what we missed.
Next, we need to create a repository. A repository is a collection of music in the representation understood by Deep Jammer. To do this, we need to run the command below. This could take a while just as a heads up. Note that repository_handler.py uses all of the MIDI files stored in the pieces directory and saves the repository in the repositories directory. By default, w've included a lot of MIDI files in the pieces directory. This means all you have to do is generate a repository. If you run into any permission issues, just add sudo to the front of the command. This should solve any problems. This happens because it needs to read MIDI files and write a repository file to your file system.
If that command ran successfully, you should see the following output:
Now that we've built a repository, we can train Deep Jammer. All we need to do is run the command below. Again, add sudo to the front of the command if you run into any permission issues. Theano needs to access a few files in the ~/.theano directory.
If Deep Jammer runs properly, you should see the output below. At every 5 epochs, we print out a summary of the current loss. We also save the current configuration (i.e., the current weights) and compose a sample. If you want, you can adjust how often we save the current configuration or compose a sample piece by modifying the variables SUMMARY_THRESHOLD and CHECKPOINT_THRESHOLD in the deep_jammer.py file. More importantly, we store every loss in a file called loss-history.txt.
Now that we've trained Deep Jammer, let's adjust some of the parameters of our neural network just so you know how. At the top of deep_jammer.py, you'll see two variables: TIME_MODEL_LAYERS and NOTE_MODEL_LAYERS. Basically, these arrays represents the number of nodes in each time layer and each note layer. You can adjust them to whatever you want. We found that the current configuration is the sweet spot though. While there are other variables that you can adjust, they aren't as important. Feel free to play around with them if you want.
In addition to adjusting the parameters of our neural network, we can change the the number of epochs and the batch size. By default, the number of epochs is 200 and the batch size is 5. You'll probably need more epochs to generate good music. We used 5000 epochs in our experiments. That said, all we have to do is add two optional flags to the command we ran before. Check it out below: