LOP

Requirements

You need to download the toolbox https://github.com/aciditeam/acidano and add it to your python path. To do this, you can either :

• place the acidano folder in the lop/Source

• or type in a shell

    PYTHONPATH=”${PYTHONPATH}:/path/to/acidano/folder”
  

A scenario

Building the database

The first step is to build the database, i.e. read a corpus of midi file in an orchestra and a piano matrices, split them into train/valid/test matrices and save them in a folder (for instance lop/Data) as .npy files. This is done by executing :

cd Source/Database; python build_data.py

The train/valid/test split is defined through indexing files, whose path is passed as an argument of the main function in build_data.py. A database can be found at https://qsdfo.github.io/LOP/database/LOP_database_06_09_17.zip.
We simplify orchestrations by mapping rare instruments to more common ones through the simplify_instrumentation table located in build_data_aux.py that you might want to edit.

Training

Is made by calling

cd Source; python main_random_search.py

Parameters at the very beginning of the file are used for execution context and should not be touched, except DEFINED_CONFIG and CONFIG_ID

N_HP_CONFIG = 1      # In case you run hyper-parameter search
LOCAL = True            # Was used for Guillimin/non-Guillimin, should be set to True
DEBUG = False          
CLEAN = False            # Remove empty folders in the result folders. If you want to run several configs in parallel, set it to False
DEFINED_CONFIG = True    # Choose between hyper-parameter search, or run for a particular configuration.
CONFIG_ID = '/13'       # Name of the folder in which results will be stored. If you choose an  existing folder, raise an error. Useless for Hyper-parameter runs.

Training parameters

The training parameters (model, optim method, granularity, type of units and quantization) are defined in the commands table :

commands = [
    'FGgru',
    'adam_L2',
    'event_level',
    'binary',
    '100'           # Set to 100 when using event_level
]

Defining a specific configuration for a model

DEFINED_CONFIG allows to run a configurations defined in the dictionnary configs at the top of main_random_search.py.

Generating

Call

python result_folder_main.py

The main function of this script takes as input a result folder's path and generate trained weights' plots, midi files (short or whole track), statistics...

## Future

• Adapt to Scikit framework
• Adapt to scikit-multilearn