Register a new module The only restriction is that the given class has to implement the static method get_module_id Args:

Restore the parameters from the configuration group Args: args (parse_args() returned Obj): the parameters of the models (will be

Save the current module parameters Args: config_group (dict): dictionary where to write the configuration

Add the module to the command line parser All modules have to be registered before that call Args: group_args (ArgumentPa

Instantiate the chosen module This function can be called only once when initializing the module Args: args (Obj): the gl

Just print the current module configuration We use the args parameters because the function is called before build_module Arg

Compute the unit scale factor for the song The scale factor allow to have a tempo independent time unit, to represent the song as an array

Transform the song into a tempo independent song Warning: If the resolution of the song is is more fine that the given scale, some in

Forward/training step operation. Does not perform run on itself but just return the operators to do so. Those have then to be run by the

Args: args: parameters of the model

Args: args: parameters of the model

Return the list of the model files inside the model directory

Parse the argument to decide were to save/load the model This function is called at each checkpoint and the first time the model is load. If

Load samples from file Args: samples_path (str): The path where to load the model (all dirs should exist) Return:

Save samples to file Args: samples_path (str): The path where to save the model (all dirs should exist)

Save the model parameters and the variables Args: sess: the current session

Initialize the weights of the model

preprocessed song > batch

a RNN decoder See parent class for arguments details

In the case of the relative song, the input dim is the number of note on the scale (12) + 1 for the next token Return: in

Return the learning rate associated at the current training step Args: glob_step (int): Number of iterations since the beginning

See parent class for more details Args: dataset (list[Song]): the training/testing set name (str): indicate the datas

Return the initial cell state

Pre-process the data once globally Do it once globally. Args: old_song (Song): original song Returns:

Reconstruct the original raw song from the preprocessed data See parent class for details Some information will be lost compare to t

Save the predicted output songs using the given recorder Args: outputs (List[np.array]): The list of the predictions of the decod

Save the song on disk Args: piano_roll (np.array): a song object containing the tracks and melody filename (str): the

Create the computational graph

Create the computational graph

Create the database from the midi files

Parse the argument to decide on which device run the model Return: str: The name of the device on which run the program

Generate some songs The midi files will be saved on the same model_dir

Training loop

Parse the arguments from the given command line Args: args (list<str>): List of arguments to parse. If None, the default

Print the current params

Load/create the conversations data Done in two steps: * Extract the midi files as a raw song format * Transform this raw son

Load the some values associated with the current model, like the current glob_step value. Needs to be called before any other function becaus

Restore or reset the model, depending of the parameters If testing mode is set, the function has no effect If the destination directo

Save the params of the model, like the current glob_step value Warning: if you modify this function, make sure the changes mirror load_params

Create the test/train set from the loaded songs The dataset has been shuffled when calling this function (Warning: the shuffling is d

Record some result/generated songs during training. This allow to see the training progression and get an idea of what the network really lea

Convert all midi files of the given directory to mp3

Is called just before feeding the placeholder, allows additional pre-processing Args: target(Bool): is true if the bach a

Prepare the batches for the current epoch WARNING: The songs are not shuffled in this functions. We leave the choice to the batch_bui

Return the batches which initiate the RNN when generating The initial batches are loaded from a json file containing the first notes of the s

Return the name of the chosen module Name is defined by get_module_id Returns: str: the name of the chosen module

Return the learning rate associated at the current training step Args: glob_step (int): Number of iterations since the be

Return the chosen module Returns: Obj: the reference on the module instance

Return the chosen labels from the softmax distribution Allows to reconstruct the song

Return the list of the different modes Useful when parsing the command lines arguments

Return the previous sampling probability for the current step. If above, the RNN should use the previous step instead of the given input.

Return the previous sampling probability for the current step. If above, the RNN should use the previous step instead of the given input.

Convert the absolute midi position into the range given by MIDI_NOTES_RANGE Return int: The new position relative to the range (p

Return the list of the different testing modes Useful on when parsing the command lines arguments

Return the target weight for the given step i using the chosen policy

Return the target weight for the given step i using the chosen policy

Restore the module configuration

Extract data from midi file Args: filename (str): a valid img file Return: np.array: the piano roll associate

Launch the training and/or the interactive mode

Print modules current configuration

Create the batch associated with the song Args: raw_song (Song): The song to convert Return: RelativeBatch

Create the song associated with the network output Args: output (list[np.Array]): The ouput of the network (size batch_size*outpu

List all available modules for the current session This function should be called only once at the beginning of the program, before p

Save the modules configurations

Initialize from a mido message Args: msg (mido.MidiMessage): a valid control_change message

Convert given note into a absolute midi position Args: rel (int): The new position relative to the range (position on keyboard)

Run the cell at step t Args: prev_keyboard: keyboard configuration for the step t-1 (Ground truth or previous step) p

Function which apply the preprocessing Args: prev_output (tf.Tensor): the ouput on which applying the transformation Retu

Use TODO formula Args: prev_output (tf.Tensor): the ouput on which applying the transformation Return: tf.Ops

Predict the output from prev and scale the result on [-1, 1] Use sigmoid activation Args: X (tf.Tensor): the input

Args: args: parameters of the model

Args: args: parameters of the model

Args: args: parameters of the model

Args: name (str): the name of the module manager (useful for saving/printing)

Load all conversations Args: args: parameters of the model

Args: args: parameters of the model

Args: args: parameters of the model

Args: args: parameters of the model

Args: args: parameters of the model

Args: args: parameters of the model

All attribute are defined with respect with the previous one

Args: extracts(list[SongExtract]): Should be of length batch_size, or at least all from the same size