s444337/praca-magisterska
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'seq2seq_model' of s444337/praca-magisterska into master

Browse Source 
seq2seq model
This commit is contained in:
Cezary Pukownik 2019-10-28 09:47:52 +00:00 committed by Gogs
commit 76db0a4b9d
16 changed files with 937 additions and 967 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

BIN
project/__pycache__/midi_processing.cpython-36.pyc Normal file
View File

Binary file not shown.

BIN
project/__pycache__/model.cpython-36.pyc Normal file
View File

Binary file not shown.

BIN
project/data/models/piano_best.pickle
View File

Binary file not shown.

BIN
project/data/output/distgit.mid
View File

Binary file not shown.

BIN
project/data/output/distgit.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 7.8 KiB

BIN
project/data/output/piano_lstm.mid
View File

Binary file not shown.

BIN
project/data/output/piano_lstm.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 7.8 KiB

74
project/extract.py Normal file
View File

@ -0,0 +1,74 @@
import os
import sys
import argparse
import pickle
from midi_processing import extract_data, analyze_data
def make_folder_if_not_exist(path):
try:
os.mkdir(path)
except:
pass
def parse_argv():
parser = argparse.ArgumentParser()
parser.add_argument('midi_pack', help='folder name for midi pack in midi_packs folder', type=str)
parser.add_argument('--n', help='name for experiment', type=str)
parser.add_argument('--b', help='lengh of sequence in bars', type=int)
parser.add_argument('-a', help='analize data', action='store_true')
args = parser.parse_args()
return args
def ask_for_workflow():
'''MODEL WORKFLOW DIALOG'''
number_of_instruments = int(input('Please specify number of instruments\n'))
model_workflow = dict()
for i in range(number_of_instruments):
input_string = input('Please specify a workflow step <Instrument> [<Second Instrument>] <mode> {m : melody, a : arrangment}\n')
tokens = input_string.split()
if tokens[-1] == 'm':
model_workflow[i] = (tokens[0], 'melody')
elif tokens[-1] == 'a':
model_workflow[i] = ((tokens[1], tokens[0]), 'arrangment')
else:
raise ValueError("The step definitiom must end with {'m', 'a'}");
make_folder_if_not_exist(os.path.join('training_sets', EXPERIMENT_NAME))
pickle.dump(model_workflow, open(os.path.join('training_sets', EXPERIMENT_NAME, 'workflow.pkl'),'wb'))
return model_workflow
def extract_from_folder(model_workflow):
for key, (instrument, how) in model_workflow.items():
if how == 'melody':
instrument_name = instrument
else:
instrument_name = instrument[1]
make_folder_if_not_exist(os.path.join('training_sets', EXPERIMENT_NAME))
save_path = os.path.join('training_sets', EXPERIMENT_NAME, instrument_name.lower() + '_data.pkl')
x_train, y_train, program = extract_data(midi_folder_path=os.path.join('midi_packs', MIDI_PACK_NAME),
how=how,
instrument=instrument,
bar_in_seq=BARS_IN_SEQ)
pickle.dump((x_train, y_train, program), open(save_path,'wb'))
if __name__ == '__main__':
args = parse_argv()
MIDI_PACK_NAME = args.midi_pack
EXPERIMENT_NAME = args.n
BARS_IN_SEQ = args.b
if not EXPERIMENT_NAME:
EXPERIMENT_NAME = MIDI_PACK_NAME
if not BARS_IN_SEQ:
BARS_IN_SEQ = 4
ANALIZE = args.a
if ANALIZE:
analyze_data(os.path.join('midi_packs', MIDI_PACK_NAME))
else:
extract_from_folder(ask_for_workflow())

166
project/generate.py
View File

@ -1,96 +1,94 @@
#!python3
#!/usr/bin/env python3
''' This module generates a sample, and create a midi file.
Usage:
>>> ./generate.py [trained_model_path] [output_path]
'''
import settings
import sys
import random
from midi_processing import MultiTrack, SingleTrack, Stream
from model import Seq2SeqModel, seq_to_numpy
from tqdm import tqdm
import argparse
import os
import pickle
import numpy as np
import tensorflow as tf
import pypianoroll as roll
import matplotlib.pyplot as plt
from tqdm import trange, tqdm
from music21 import converter, instrument, note, chord, stream
from keras.layers import Input, Dense, Conv2D
from keras.models import Model
from keras.layers import Input, Dense, Conv2D, Flatten, LSTM, Dropout, TimeDistributed, RepeatVector
from keras.models import Model, Sequential
parser = argparse.ArgumentParser()
parser.add_argument('n', help='name for experiment', type=str)
parser.add_argument('s', help='session name', type=str)
parser.add_argument('--i', help='number of midis to generate', type=int)
parser.add_argument('--l', help='latent_dim_of_model', type=int)
parser.add_argument('--m', help="mode {'from_seq', 'from_state}'", type=str)
args = parser.parse_args()
EXPERIMENT_NAME = args.n
SESSION_NAME = args.s
GENERETIONS_COUNT = args.i
LATENT_DIM = args.l
MODE = args.m
def choose_by_prob(list_of_probs):
''' This functions a list of values and assumed
that if the value is bigger it should by returned often
if not GENERETIONS_COUNT:
GENERETIONS_COUNT = 1
if not LATENT_DIM:
LATENT_DIM = 256
It was crated to give more options to choose than argmax function,
thus is more than one way that you can develop a melody.
Returns a index of choosen value from given list.
'''
sum_prob = np.array(list_of_probs).sum()
prob_normalized = [x/sum_prob for x in list_of_probs]
cumsum = np.array(prob_normalized).cumsum()
prob_cum = cumsum.tolist()
random_x = random.random()
for i, x in enumerate(prob_cum):
if random_x < x:
return i
if not MODE:
MODE = 'from_seq'
trained_model_path = sys.argv[1]
output_path = sys.argv[2]
model_workflow = pickle.load(open(os.path.join('training_sets', EXPERIMENT_NAME, 'workflow.pkl'),'rb'))
# load model and dictionary that can translate back index_numbers to notes
# this dictionary is generated with model
print('Loading... {}'.format(trained_model_path))
model = pickle.load(open(trained_model_path, 'rb'))
int_to_note, n_vocab, seq_len = pickle.load(open('{}_dict'.format(trained_model_path), 'rb'))
seed = [random.randint(0,n_vocab) for x in range(seq_len)]
music = []
print('Generating...')
for i in trange(124):
predicted_vector = model.predict(np.array(seed).reshape(1,seq_len,1))
# using best fitted note
# predicted_index = np.argmax(predicted_vector)
# using propability distribution for choosing note
# to prevent looping
predicted_index = choose_by_prob(predicted_vector)
music.append(int_to_note[predicted_index])
seed.append(predicted_index)
seed = seed[1:1+seq_len]
print('Saving...')
offset = 0
output_notes = []
for _event in tqdm(music):
event, note_len = _event.split(';')
if (' ' in event) or event.isdigit():
notes_in_chord = event.split(' ')
notes = []
for current_note in notes_in_chord:
new_note = note.Note(current_note)
new_note.storedInstrument = instrument.Piano()
notes.append(new_note)
new_chord = chord.Chord(notes)
new_chord.offset = offset
output_notes.append(new_chord)
band = dict()
for key, value in model_workflow.items():
if isinstance(value[0], str):
instrument = value[0]
generator = None
else:
new_note = note.Note(event)
new_note.offset = offset
new_note.storedInstrument = instrument.Piano()
output_notes.append(new_note)
instrument = value[0][1]
generator = value[0][0]
band[instrument] = [None, None, generator]
offset += float(note_len)
'''LOAD MODELS'''
for instrument in tqdm(band):
data_path = os.path.join('training_sets', EXPERIMENT_NAME, instrument.lower() + '_data.pkl')
model_path = os.path.join('models', EXPERIMENT_NAME, instrument.lower() + '_model.h5')
x_train, y_train, program = pickle.load(open(data_path,'rb'))
model = Seq2SeqModel(LATENT_DIM, x_train, y_train)
model.load(model_path)
band[instrument][0] = model
band[instrument][1] = program
midi_stream = stream.Stream(output_notes)
for midi_counter in tqdm(range(GENERETIONS_COUNT)):
''' MAKE MULTIINSTRUMENTAL MUSIC !!!'''
notes = dict()
midi_stream.write('midi', fp='{}.mid'.format(output_path))
for instrument, (model, program, generator) in band.items():
if generator == None:
notes[instrument] = model.develop(mode=MODE)
else:
input_data = seq_to_numpy(notes[generator],
model.transformer.x_max_seq_length,
model.transformer.x_vocab_size,
model.transformer.x_transform_dict)
notes[instrument] = model.predict(input_data)[:-1]
print('Done!')
'''COMPILE TO MIDI'''
generated_midi = MultiTrack()
for instrument, (model, program, generator) in band.items():
if instrument == 'Drums':
is_drums = True
else:
is_drums = False
stream = Stream(first_tick=0, notes=notes[instrument])
track = SingleTrack(name=instrument ,program=program, is_drum=is_drums, stream=stream)
generated_midi.tracks.append(track)
# make folder for new experiment
try:
os.mkdir(os.path.join('generated_music', EXPERIMENT_NAME))
except:
pass
try:
os.mkdir(os.path.join('generated_music', EXPERIMENT_NAME, SESSION_NAME))
except:
pass
save_path = os.path.join('generated_music', EXPERIMENT_NAME, SESSION_NAME, f'{EXPERIMENT_NAME}_{midi_counter}_{MODE}_{LATENT_DIM}.mid')
generated_midi.save(save_path)
# print(f'Generated succefuly to {save_path}')

133
project/midi.py
View File

@ -1,133 +0,0 @@
#!python3
#!/usr/bin/env python3
''' This module contains functions to endocing midi files into data samples
that is prepared for model training.
midi_folder_path - the path to directiory containing midi files
output_path - the output path where will be created samples of data
Usage:
>>> ./midi.py <midi_folder_path> <output_path> <sequence_lenth>
'''
import settings
import pypianoroll as roll
import numpy as np
import os
from tqdm import tqdm
from math import floor
import sys
from collections import defaultdict
import pickle
from music21 import converter, instrument, note, chord, stream
import music21
class MidiParseError(Exception):
"""Error that is raised then midi file cannot be parsed"""
pass
def parse_argv(argv):
'''This function is parsing given arguments when running a midi script.
Returns a tuple consinting of midi_folder_path, output_path, seq_len'''
try:
midi_folder_path = argv[1]
output_path = argv[2]
seq_len = int(argv[3])
return midi_folder_path, output_path, seq_len
except IndexError:
raise AttributeError('You propably didnt pass parameters to run midi.py script.\
>>> ./midi.py <midi_folder_path> <output_path> <sequence_lenth>')
def to_sequence(midi_path, seq_len):
''' This function is supposed to be used on one midi file in directory loop.
Its encoding midi files, into sequances of given lenth as a train_X,
and the next note as a train_y. Also splitting midi samples into
instrument group.
Use for LSTM neural network.
Parameters:
- midi_path: path to midi file
- seq_len: lenght of sequance before prediction
Returns: Tuple of train_X, train_y dictionaries consisinting of samples of song grouped by instruments
'''
seq_by_instrument = defaultdict( lambda : [] )
try:
midi_file = music21.converter.parse(midi_path)
except music21.midi.MidiException:
raise MidiParseError
stream = music21.instrument.partitionByInstrument(midi_file)
for part in stream:
for event in part:
if part.partName != None:
if isinstance(event, music21.note.Note):
to_export_event = '{};{}'.format(str(event.pitch), float(event.quarterLength))
seq_by_instrument[part.partName].append(to_export_event)
elif isinstance(event, music21.chord.Chord):
to_export_event = '{};{}'.format(' '.join(str(note) for note in event.pitches), float(event.quarterLength))
seq_by_instrument[part.partName].append(to_export_event)
X_train_by_instrument = defaultdict( lambda : [] )
y_train_by_instrument = defaultdict( lambda : [] )
for instrument, sequence in seq_by_instrument.items():
for i in range(len(sequence)-(seq_len)) :
X_train_by_instrument[instrument].append(np.array(sequence[i:i+seq_len])) # <seq lenth
y_train_by_instrument[instrument].append(np.array(sequence[i+seq_len]))
return X_train_by_instrument, y_train_by_instrument
def colect_samples(midi_folder_path, seq_len):
'''This function is looping throuth given directories and
collecting samples from midi files.
Parameters: midi_folder_path - a path to directory with midi files
seq_len - a lenth of train_X sample that tells
how many notes is given do LSTM to predict the next note.
Returns: Tuple of train_X, train_y dictionaries consisinting
of samples of all songs in directory grouped by instruments.
'''
print('Collecting samples...')
train_X = defaultdict( lambda : [] )
train_y = defaultdict( lambda : [] )
for directory, subdirectories, files in os.walk(midi_folder_path):
for midi_file in tqdm(files):
midi_file_path = os.path.join(directory, midi_file)
try:
_X_train, _y_train = to_sequence(midi_file_path, seq_len)
except MidiParseError:
continue
for (X_key, X_value), (y_key, y_value) in zip(_X_train.items(), _y_train.items()):
train_X[X_key].extend(np.array(X_value))
train_y[y_key].extend(np.array(y_value))
return train_X, train_y
def save_samples(output_path, samples):
'''This function save samples to npz packages, splitted by instrument.'''
print('Saving...')
if not os.path.exists(output_path):
os.makedirs(output_path)
train_X, train_y = samples
for (X_key, X_value), (y_key, y_value) in tqdm(zip(train_X.items(), train_y.items())):
if X_key == y_key:
np.savez_compressed('{}/{}.npz'.format(output_path, X_key), np.array(X_value), np.array(y_value))
def main():
midi_folder_path, output_path, seq_len = parse_argv(sys.argv)
save_samples(output_path, colect_samples(midi_folder_path, seq_len))
print('Done!')
if __name__ == '__main__':
main()

697
project/midi_processing.py
View File

@ -1,324 +1,16 @@
#!/usr/bin/env python
# coding: utf-8
# In[1]:
import os
import sys
import pickle
import operator
import numpy as np
from collections import defaultdict
from collections import Counter
from random import randint
import pretty_midi as pm
from tqdm import tqdm
# In[98]:
TODO = '''
TODO: put methods of data extraction for seq2seq arangment model to multitrack class [DONE]
TODO: make functions for data extraction for seq2seq model for riff/melody generation [DONE]
'''
# In[367]:
# '''return a dictionary with tracks indexes grouped by instrument class'''
# tracks = file.tracks
# names = [track.name for track in tracks]
# uniqe_instruemnts = set(names)
# tracks_by_instrument = dict()
# for key in uniqe_instruemnts:
# tracks_by_instrument[key] = []
# for i, track in enumerate(tracks):
# tracks_by_instrument[track.name].append(i)
# tracks_by_instrument
# In[368]:
# def get_posible_pairs(instrument_x, instrument_y):
# '''it takes two lists, and return a list of tuples with every posible 2-element combination
# parameters:
# -----------
# instrument_x, instrument_y : string {'Guitar','Bass','Drums'}
# a string that represent a instrument class you want to look for in midi file.
# returns:
# ----------
# pairs: list of tuples
# a list of posible 2-element combination of two lists
# '''
# x_indexes = tracks_by_instrument[instrument_x]
# y_indexes = tracks_by_instrument[instrument_y]
# pairs = []
# # pairs = [(x,y) for x in x_indexes for y in y_indexes]
# for x in x_indexes:
# for y in y_indexes:
# pairs.append((x,y))
# return pairs
# In[369]:
# def get_common_bars_for_every_possible_pair(pairs)
# ''' for every possible pair of given instrument classes
# returns common bars from multitrack'''
# x_bars = []
# y_bars = []
# for x_track_index, y_track_index in pairs:
# _x_bars, _y_bars = get_common_bars(file.tracks[x_track_index], file.tracks[y_track_index])
# x_bars.extend(_x_bars)
# y_bars.extend(_y_bars)
# return x_bars, y_bars
# In[370]:
# def get_data_seq2seq_arrangment(self, bars_in_seq):
# ## This is the end of extracting data from midis to seq2seq arranging network.
# '''this method is returning a sequances of given lenth by rolling this lists of x and y for arrangemt generation'''
# x_seq = []
# y_seq = []
# for i in range(len(x_bars) - bars_in_seq + 1):
# x_seq_to_add = [note for bar in x_bars[i:i+bars_in_seq] for note in bar ]
# y_seq_to_add = [note for bar in y_bars[i:i+bars_in_seq] for note in bar ]
# x_seq.append(x_seq_to_add)
# y_seq.append(y_seq_to_add)
# len(x_seq), len(y_seq)
# # get_bar_len(y_seq[0])
# In[371]:
# def get_track_by_instrument(self):
# '''return a dictionary with tracks indexes grouped by instrument class'''
# tracks = self.tracks
# names = [track.name for track in tracks]
# uniqe_instruemnts = set(names)
# tracks_by_instrument = dict()
# for key in uniqe_instruemnts:
# tracks_by_instrument[key] = []
# for i, track in enumerate(tracks):
# tracks_by_instrument[track.name].append(i)
# return tracks_by_instrument
# In[372]:
# def get_data_seq2seq_melody(self,instrument_class, x_seq_len=4)
# '''return a list of bars with content for every track with given instrument class for melody generaiton'''
# instrument_tracks = tracks_by_instrument[instrument_class]
# for track_index in instrument_tracks:
# # make below as function: get_bars_with_content
# bars = file.tracks[track_index].stream_to_bars()
# bars_indexes_with_content = get_bar_indexes_with_content(bars)
# bars_with_content = [bars[i] for i in get_bar_indexes_with_content(bars)]
# # make below as function: get_sequances_from_bars (for seq2seq melody generator)
# x_seq = []
# y_bar = []
# for i in range(len(bars_with_content)-seq_len-1):
# _x_seq = bars_with_content[i:i+seq_len]
# _y_bar = bars_with_content[i+seq_len]
# x_seq.append(_x_seq)
# y_bar.append(_y_bar)
# len(x_seq), len(y_bar)
# # print( ' x:' ,x_seq[1],'\n', 'y: ', y_bar[1],'\n', 'seq: ',bars_with_content[1:6])
# In[15]:
def get_bar_indexes_with_content(bars):
'''this method is looking for non-empty bars in the tracks bars
the empty bar consist of only rest notes.
returns: a set of bars indexes with notes
'''
bars_indexes_with_content = set()
for i, bar in enumerate(bars):
if bar_has_content(bar):
bars_indexes_with_content.add(i)
return bars_indexes_with_content
# In[4]:
def get_bars_with_content(bars):
'''this method is looking for non-empty bars in the tracks bars
the empty bar consist of only rest notes.
returns: a set of bars with notes
'''
bars_with_content = []
for bar in bars:
if bar_has_content(bar):
bars_with_content.append(bar)
return bars_with_content
# In[5]:
def get_common_bars(track_x,track_y):
'''return common bars, for two tracks is song
return X_train, y_train list of
'''
bars_x = track_x.stream_to_bars()
bars_y = track_y.stream_to_bars()
bwc_x = get_bar_indexes_with_content(bars_x)
bwc_y = get_bar_indexes_with_content(bars_y)
common_bars = bwc_x.intersection(bwc_y)
common_bars_x = [bars_x[i] for i in common_bars]
common_bars_y = [bars_y[i] for i in common_bars]
return common_bars_x, common_bars_y
# In[6]:
def get_bar_len(bar):
"""calculate a lenth of a bar
parameters:
bar : list
list of "notes", tuples like (pitches, len)
"""
time = 0
for note in bar:
time += note[1]
return time
# In[7]:
def bar_has_content(bar):
'''check if bar has any musical information, more accurate
it checks if in a bar is any non-rest event like note, or chord
parameters:
-----------
bar: list
list of notes
return:
-------
bool:
True if bas has concent and False of doesn't
'''
bar_notes = len(bar)
count_rest = 0
for note in bar:
if note[0] == (-1,):
count_rest += 1
if count_rest == bar_notes:
return False
else:
return True
# In[8]:
def round_to_sixteenth_note(x, base=0.25):
'''round value to closest multiplication by base
in default to 0.25 witch is sisteenth note accuracy
'''
return base * round(x/base)
# In[9]:
def parse_pretty_midi_instrument(instrument, resolution, time_to_tick, key_offset):
''' arguments: a prettyMidi instrument object
return: a custom SingleTrack object
'''
first_tick = None
prev_tick = 0
prev_note_lenth = 0
max_rest_len = 4.0
notes = defaultdict(lambda:[set(), set()])
for note in instrument.notes:
if first_tick == None:
# first_tick = round_to_sixteenth_note(time_to_tick(note.start)/resolution)
first_tick = 0
tick = round_to_sixteenth_note(time_to_tick(note.start)/resolution)
# add rest if needed
if prev_tick != None:
act_tick = prev_tick + prev_note_lenth
if act_tick < tick:
rest_lenth = tick - act_tick
while rest_lenth > max_rest_len:
notes[act_tick] = [{-1},{max_rest_len}]
act_tick += max_rest_len
rest_lenth -= max_rest_len
notes[act_tick] = [{-1},{rest_lenth}]
note_lenth = round_to_sixteenth_note(time_to_tick(note.end-note.start)/resolution)
if -1 in notes[tick][0]:
notes[tick] = [set(), set()]
if instrument.is_drum:
notes[tick][0].add(note.pitch)
else:
notes[tick][0].add(note.pitch+key_offset)
notes[tick][1].add(note_lenth)
prev_tick = tick
prev_note_lenth = note_lenth
notes = [(tuple(e[0]), max(e[1])) for e in notes.values()]
name = 'Drums' if instrument.is_drum else pm.program_to_instrument_class(instrument.program)
return SingleTrack(name, instrument.program, instrument.is_drum, Stream(first_tick,notes) )
# In[10]:
def remove_duplicated_sequences(xy_tuple):
x = xy_tuple[0]
y = xy_tuple[1]
x_freeze = [tuple(seq) for seq in x]
y_freeze = [tuple(seq) for seq in y]
unique_data = list(set(zip(x_freeze,y_freeze)))
x_unique = [seq[0] for seq in unique_data]
y_unique = [seq[1] for seq in unique_data]
return x_unique, y_unique
# In[11]:
# TODO: Stream class is no logner needed <- remore from code and make just SingleTrack.notes instead on SingleTrack.stream.notes
class Stream():
def __init__ (self, first_tick, notes):
@ -327,11 +19,7 @@ class Stream():
def __repr__(self):
return '<Stream object with {} musical events>'.format(len(self.notes))
# In[12]:
class SingleTrack():
'''class of single track in midi file encoded from pretty midi library
@ -352,6 +40,7 @@ class SingleTrack():
self.program = program
self.is_drum = is_drum
self.stream = stream
self.is_melody = self.check_if_melody()
def __repr__(self):
return "<SingleTrack object. Name:{}, Program:{}, is_drum:{}>".format(self.name, self.program, self.is_drum)
@ -387,13 +76,21 @@ class SingleTrack():
def stream_to_bars(self, beat_per_bar=4):
'''it takes notes and split it into equaly time distibuted sequances
if note is between bars, the note is splited into two notes, with time sum equal to the note between bars.
arguments:
----------
stream: list of "notes"
return:
-------
bars: list: list of lists of notes, every list has equal time. in musical context it returns bars
'''
# TODO: if last bar of sequance has less notes to has time equal given bar lenth it is left shorter
# fill the rest of bar with rests
# FIXME: there is a problem, where note is longer that bar and negative time occured
# split note to max_rest_note, the problem occured when note is longer then 2 bars
notes = self.stream.notes
bars = []
time = 0
@ -408,8 +105,15 @@ class SingleTrack():
bars.append([])
if add_tail:
bars[bar_index].append(tail_note)
time += note_len(tail_note)
tail_pitch = note_pitch(tail_note)
while tail_note_len > beat_per_bar:
bars[bar_index].append((tail_pitch, beat_per_bar))
tail_note_len -= beat_per_bar
bar_index += 1
bars.append([])
bars[bar_index].append((tail_pitch, tail_note_len))
time += tail_note_len
add_tail = False
time += note_len(note)
@ -434,12 +138,45 @@ class SingleTrack():
else:
bars[bar_index].append(note)
return bars
# In[99]:
return bars
def check_if_melody(self):
'''checks if Track object could be a melody
it checks if percentage of single notes in Track.stream.notes is higher than treshold of 90%
TODO: and there is at least 3 notes in bar per average
'''
events = None
single_notes = None
content_lenth = None
for note in self.stream.notes:
if self.name not in ['Bass','Drums']:
events = 0
content_lenth = 0
single_notes = 0
if note[0][0] != -1: # if note is not a rest
events += 1
content_lenth += note[1]
if len(note[0]) == 1: # if note is a single note, not a chord
single_notes += 1
if events != None:
if events == 0 or content_lenth == 0:
return False
else:
single_notes_rate = single_notes/events
density_rate = events/content_lenth
if single_notes_rate >= 0.9 and density_rate < 2:
self.name = 'Melody'
return True
else:
return False
else:
return False
class MultiTrack():
'''Class that represent one midi file
atributes:
@ -452,13 +189,14 @@ class MultiTrack():
def __init__(self, path=None, tempo=100):
self.tempo = tempo
self.pm_obj = pm.PrettyMIDI(path, initial_tempo=self.tempo)
self.pm_obj = pm.PrettyMIDI(path, initial_tempo=self.tempo) # changename to self.PrettyMIDI
self.res = self.pm_obj.resolution
self.time_to_tick = self.pm_obj.time_to_tick
self.name = path
self.tracks = [parse_pretty_midi_instrument(instrument, self.res, self.time_to_tick, self.get_pitch_offset_to_C() ) for instrument in self.pm_obj.instruments]
self.tracks_by_instrument = self.get_track_by_instrument()
# TODO: this function is deprecated <- remove from code
def get_multiseq(self):
'''tracks: list of SingleTrack objects
reaturn a dictionary of sequences for every sequence in SingleTrack
@ -475,6 +213,14 @@ class MultiTrack():
multiseq[key].append((single_track.name,value))
return multiseq
def get_programs(self, instrument):
program_list = []
for track in self.tracks:
if track.name == instrument:
program_list.append(track.program)
return program_list
def get_pitch_offset_to_C(self):
'''to get better train resoult without augmenting midis to all posible keys
@ -526,7 +272,10 @@ class MultiTrack():
return x_bars, y_bars
def get_data_seq2seq_arrangment(self, x_instrument, y_instrument, bars_in_seq=4):
'''this method is returning a sequances of given lenth by rolling this lists of x and y for arrangemt generation'''
'''this method is returning a sequances of given lenth by rolling this lists of x and y for arrangemt generation
x and y has the same bar lenth, and represent the same musical phrase playd my difrent instruments (tracks)
'''
x_seq = []
y_seq = []
x_bars, y_bars = self.get_common_bars_for_every_possible_pair(x_instrument, y_instrument)
@ -540,7 +289,12 @@ class MultiTrack():
return x_seq, y_seq
def get_data_seq2seq_melody(self,instrument_class, x_seq_len=4):
'''return a list of bars with content for every track with given instrument class for melody generaiton'''
'''return a list of bars with content for every track with given instrument class for melody generaiton
x_seq_len and y_seq_len
x previous sentence, y next sentence of the same melody line
'''
instrument_tracks = self.tracks_by_instrument[instrument_class]
@ -573,13 +327,7 @@ class MultiTrack():
'''
x_indexes = self.tracks_by_instrument[instrument_x]
y_indexes = self.tracks_by_instrument[instrument_y]
# pairs = []
pairs = [(x,y) for x in x_indexes for y in y_indexes]
# for x in x_indexes:
# for y in y_indexes:
# pairs.append((x,y))
return pairs
def show_map(self):
@ -595,13 +343,217 @@ class MultiTrack():
track_str += '_'
print(track.name[:4],':', track_str)
def stream_to_bars(notes, beat_per_bar=4):
'''it takes notes and split it into equaly time distibuted sequances
if note is between bars, the note is splited into two notes, with time sum equal to the note between bars.
arguments:
stream: list of "notes"
return:
bars: list: list of lists of notes, every list has equal time. in musical context it returns bars
'''
# TODO: if last bar of sequance has less notes to has time equal given bar lenth it is left shorter
# fill the rest of bar with rests
# FIXME: there is a problem, where note is longer that bar and negative time occured
# split note to max_rest_note, the problem occured when note is longer then 2 bars - FIXED
bars = []
time = 0
bar_index = 0
add_tail = False
note_pitch = lambda note: note[0]
note_len = lambda note: note[1]
for note in notes:
try:
temp = bars[bar_index]
except IndexError:
bars.append([])
if add_tail:
tail_pitch = note_pitch(tail_note)
while tail_note_len > beat_per_bar:
bars[bar_index].append((tail_pitch, beat_per_bar))
tail_note_len -= beat_per_bar
bar_index += 1
bars[bar_index].append((tail_pitch, tail_note_len))
time += tail_note_len
add_tail = False
time += note_len(note)
if time == beat_per_bar:
bars[bar_index].append(note)
time = 0
bar_index += 1
# In[104]:
elif time > beat_per_bar: # if note is between bars
between_bars_note_len = note_len(note)
tail_note_len = time - beat_per_bar
leading_note_len = between_bars_note_len - tail_note_len
leading_note = (note_pitch(note), leading_note_len)
bars[bar_index].append(leading_note)
tail_note = (note_pitch(note), tail_note_len)
add_tail = True
time = 0
bar_index += 1
else:
bars[bar_index].append(note)
return bars
def extract_data(midi_folder_path=None, how=None, instrument=None, remove_duplicates=True):
def get_bar_len(bar):
"""calculate a lenth of a bar
parameters:
bar : list
list of "notes", tuples like (pitches, len)
"""
time = 0
for note in bar:
time += note[1]
return time
def get_common_bars(track_x,track_y):
'''return common bars, for two tracks is song
return X_train, y_train list of
'''
bars_x = track_x.stream_to_bars()
bars_y = track_y.stream_to_bars()
bwc_x = get_bar_indexes_with_content(bars_x)
bwc_y = get_bar_indexes_with_content(bars_y)
common_bars = bwc_x.intersection(bwc_y)
common_bars_x = [bars_x[i] for i in common_bars]
common_bars_y = [bars_y[i] for i in common_bars]
return common_bars_x, common_bars_y
def get_bar_indexes_with_content(bars):
'''this method is looking for non-empty bars in the tracks bars
the empty bar consist of only rest notes.
returns: a set of bars indexes with notes
'''
bars_indexes_with_content = set()
for i, bar in enumerate(bars):
if bar_has_content(bar):
bars_indexes_with_content.add(i)
return bars_indexes_with_content
def get_bars_with_content(bars):
'''this method is looking for non-empty bars in the tracks bars
the empty bar consist of only rest notes.
returns: a set of bars with notes
'''
bars_with_content = []
for bar in bars:
if bar_has_content(bar):
bars_with_content.append(bar)
return bars_with_content
def bar_has_content(bar):
'''check if bar has any musical information, more accurate
it checks if in a bar is any non-rest event like note, or chord
parameters:
-----------
bar: list
list of notes
return:
-------
bool:
True if bas has concent and False of doesn't
'''
bar_notes = len(bar)
count_rest = 0
for note in bar:
if note[0] == (-1,):
count_rest += 1
if count_rest == bar_notes:
return False
else:
return True
def round_to_sixteenth_note(x, base=0.25):
'''round value to closest multiplication by base
in default to 0.25 witch is sisteenth note accuracy
'''
return base * round(x/base)
def parse_pretty_midi_instrument(instrument, resolution, time_to_tick, key_offset):
''' arguments: a prettyMidi instrument object
return: a custom SingleTrack object
'''
first_tick = None
prev_tick = 0
prev_note_lenth = 0
max_rest_len = 4.0
notes = defaultdict(lambda:[set(), set()])
for note in instrument.notes:
if first_tick == None:
first_tick = 0
tick = round_to_sixteenth_note(time_to_tick(note.start)/resolution)
if prev_tick != None:
act_tick = prev_tick + prev_note_lenth
if act_tick < tick:
rest_lenth = tick - act_tick
while rest_lenth > max_rest_len:
notes[act_tick] = [{-1},{max_rest_len}]
act_tick += max_rest_len
rest_lenth -= max_rest_len
notes[act_tick] = [{-1},{rest_lenth}]
note_lenth = round_to_sixteenth_note(time_to_tick(note.end-note.start)/resolution)
if -1 in notes[tick][0]:
notes[tick] = [set(), set()]
if instrument.is_drum:
notes[tick][0].add(note.pitch)
else:
notes[tick][0].add(note.pitch+key_offset)
notes[tick][1].add(note_lenth)
prev_tick = tick
prev_note_lenth = note_lenth
notes = [(tuple(e[0]), max(e[1])) for e in notes.values()]
name = 'Drums' if instrument.is_drum else pm.program_to_instrument_class(instrument.program)
return SingleTrack(name, instrument.program, instrument.is_drum, Stream(first_tick,notes) )
def remove_duplicated_sequences(xy_tuple):
''' removes duplicated x,y sequences
parameters:
-----------
xy_tuple: tuple of lists
tuple of x,y lists that represens sequances in training set
return:
------
x_unique, y_unique: tuple
a tuple of cleaned x, y traing set
'''
x = xy_tuple[0]
y = xy_tuple[1]
x_freeze = [tuple(seq) for seq in x]
y_freeze = [tuple(seq) for seq in y]
unique_data = list(set(zip(x_freeze,y_freeze)))
x_unique = [seq[0] for seq in unique_data]
y_unique = [seq[1] for seq in unique_data]
return x_unique, y_unique
def extract_data(midi_folder_path=None, how=None, instrument=None, bar_in_seq=4, remove_duplicates=True):
'''extract musical data from midis in given folder, to x_train, y_train lists on sequences
parameters:
@ -628,49 +580,74 @@ def extract_data(midi_folder_path=None, how=None, instrument=None, remove_duplic
notes:
------
extracted data is transposed to the key od C
extracted data is transposed to the key of C
duplicated x,y pairs are removed
'''
if how not in {'melody','arrangment'}:
raise ValueError('how parameter must by one of {melody,arrangment} ')
raise ValueError('how parameter must by one of {melody, arrangment} ')
x_train = []
y_train = []
programs_for_instrument = []
from collections import Counter
for directory, subdirectories, files in os.walk(midi_folder_path):
for midi_file in tqdm(files):
for midi_file in tqdm(files, desc='Exporting: {}'.format(instrument)):
midi_file_path = os.path.join(directory, midi_file)
try:
mt = MultiTrack(midi_file_path)
# get programs
mt.get_programs(instrument)
if how=='melody':
x ,y = mt.get_data_seq2seq_melody(instrument)
x ,y = mt.get_data_seq2seq_melody(instrument, bar_in_seq)
programs_for_instrument.extend(mt.get_programs(instrument))
if how=='arrangment':
x ,y = mt.get_data_seq2seq_arrangment(instrument[0], instrument[1])
x ,y = mt.get_data_seq2seq_arrangment(instrument[0], instrument[1], bar_in_seq)
programs_for_instrument.extend(mt.get_programs(instrument[1]))
x_train.extend(x)
y_train.extend(y)
except:
continue
most_recent_program = most_recent(programs_for_instrument)
if remove_duplicates:
x_train, y_train = remove_duplicated_sequences((x_train, y_train))
return x_train , y_train
# In[109]:
def main():
'''extract data from midis
return x_train , y_train, most_recent_program
def most_recent(list):
occurence_count = Counter(list)
return occurence_count.most_common(1)[0][0]
def analyze_data(midi_folder_path):
'''Show usage of instumets in midipack
parameters:
-----------
midi_folder_path : string
a path to directory where midi files are stored
'''
x_train, y_train = extract_data(midi_folder_path='WhiteStripes', how='arrangment', instrument=('Guitar','Bass'))
pickle.dump((x_train, y_train), open('Guitar_to_Bass_data.pkl','wb'))
return x_train, y_train
# In[107]:
if __name__=='__main__':
main()
instrument_count = dict()
instrument_programs = dict()
for directory, subdirectories, files in os.walk(midi_folder_path):
for midi_file in tqdm(files):
midi_file_path = os.path.join(directory, midi_file)
try:
mt = MultiTrack(midi_file_path)
for track in mt.tracks:
try:
instrument_count[track.name] += len(get_bars_with_content(track.stream_to_bars()))
except KeyError:
instrument_count[track.name] = 1
except Exception as e:
print(e)
for key, value in sorted(instrument_count.items(), key=lambda x: x[1], reverse=True):
print(value, 'of', key)

370
project/model.py Normal file
View File

@ -0,0 +1,370 @@
from __future__ import print_function
from midi_processing import stream_to_bars
from keras.models import Model, load_model
from keras.layers import Input, LSTM, Dense, LSTM, LSTMCell, TimeDistributed
import numpy as np
class Seq2SeqTransformer():
''' encoder/transforer
params:
-------
x_train, y_train - list of sequences
methods:
fit
transform'''
def __init__(self):
self.transform_dict = None
self.reverse_dict = None
self.vocab_x = None
self.vocab_y = None
def preprocess(self, x_train, y_train):
'''Converts training set do list and add special chars'''
_x_train = []
for i, seq in enumerate(x_train):
_x_train.append([])
for note in seq:
_x_train[i].append(note)
_y_train = []
for i, seq in enumerate(y_train):
_y_train.append([])
_y_train[i].append('<GO>')
for note in seq:
_y_train[i].append(note)
_y_train[i].append('<EOS>')
return _x_train, _y_train
def transform(self, x_train, y_train):
x_vocab = set([note for seq in x_train for note in seq])
y_vocab = set([note for seq in y_train for note in seq])
self.x_vocab = sorted(list(x_vocab))
self.y_vocab = ['<GO>','<EOS>']
self.y_vocab.extend(sorted(list(y_vocab)))
self.x_vocab_size = len(self.x_vocab)
self.y_vocab_size = len(self.y_vocab)
self.x_transform_dict = dict(
[(char, i) for i, char in enumerate(self.x_vocab)])
self.y_transform_dict = dict(
[(char, i) for i, char in enumerate(self.y_vocab)])
self.x_reverse_dict = dict(
(i, char) for char, i in self.x_transform_dict.items())
self.y_reverse_dict = dict(
(i, char) for char, i in self.y_transform_dict.items())
x_train, y_train = self.preprocess(x_train, y_train)
self.x_max_seq_length = max([len(seq) for seq in x_train])
self.y_max_seq_length = max([len(seq) for seq in y_train])
encoder_input_data = np.zeros(
(len(x_train), self.x_max_seq_length, self.x_vocab_size),
dtype='float32')
decoder_input_data = np.zeros(
(len(x_train), self.y_max_seq_length, self.y_vocab_size),
dtype='float32')
decoder_target_data = np.zeros(
(len(x_train), self.y_max_seq_length, self.y_vocab_size),
dtype='float32')
for i, (x_train, y_train) in enumerate(zip(x_train, y_train)):
for t, char in enumerate(x_train):
encoder_input_data[i, t, self.x_transform_dict[char]] = 1.
for t, char in enumerate(y_train):
decoder_input_data[i, t, self.y_transform_dict[char]] = 1.
if t > 0:
decoder_target_data[i, t - 1, self.y_transform_dict[char]] = 1.
return encoder_input_data, decoder_input_data, decoder_target_data
class Seq2SeqModel():
'''NeuralNerwork Seq2Seq model.
The network is created based on training data
'''
def __init__(self, latent_dim, x_train, y_train):
self.has_predict_model = False
self.has_train_model = False
self.x_train = x_train
self.y_train = y_train
self.latent_dim = latent_dim
self.transformer = Seq2SeqTransformer()
self.encoder_input_data, self.decoder_input_data, self.decoder_target_data = self.transformer.transform(self.x_train, self.y_train)
# ---------------
# SEQ 2 SEQ MODEL:
# INPUT_1 : encoder_input_data
# INPUT_2 : decodet_input_data
# OUTPUT : decoder_target_data
# ---------------
# ENCODER MODEL
#---------------
# 1 layer - Input : encoder_input_data
self.encoder_inputs = Input(shape=(None, self.transformer.x_vocab_size ))
# 2 layer - LSTM_1, LSTM
self.encoder = LSTM(latent_dim, return_state=True)
#self.encoder = LSTM(latent_dim, return_state=True)
# 2 layer - LSTM_1 : outputs
self.encoder_outputs, self.state_h, self.state_c = self.encoder(self.encoder_inputs)
self.encoder_states = [self.state_h, self.state_c]
# DECODER MODEL
#---------------
# 1 layer - Input : decoder_input_data
self.decoder_inputs = Input(shape=(None, self.transformer.y_vocab_size))
# 2 layer - LSTM_1, LSTM
self.decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
#self.decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
# 2 layer - LSTM_2 : outputs, full sequance as lstm layer
self.decoder_outputs, _, _ = self.decoder_lstm(self.decoder_inputs,
initial_state=self.encoder_states)
# 3 layer - Dense
self.decoder_dense = Dense(self.transformer.y_vocab_size, activation='softmax')
# 3 layer - Dense : outputs, full sequance as the array of one-hot-encoded elements
self.decoder_outputs = self.decoder_dense(self.decoder_outputs)
def init_train_model(self):
self.train_model = Model([self.encoder_inputs, self.decoder_inputs], self.decoder_outputs)
self.train_model.compile(optimizer='rmsprop', loss='categorical_crossentropy')
def fit(self, batch_size, epochs, callbacks):
if not self.has_train_model:
self.init_train_model()
self.has_train_model = True
history = self.train_model.fit([self.encoder_input_data, self.decoder_input_data], self.decoder_target_data,
batch_size=batch_size,
epochs=epochs,
callbacks=callbacks,
validation_split=0.2)
return history
def save(self, path):
self.train_model.save(path)
def load(self, path):
self.train_model = load_model(path)
self.has_train_model = True
self.encoder_inputs = self.train_model.layers[0].input
self.encoder = self.train_model.layers[2]
self.encoder_outputs, self.state_h, self.state_c = self.train_model.layers[2].output
self.encoder_states = [self.state_h, self.state_c]
self.decoder_inputs = self.train_model.layers[1].input
self.decoder_lstm = self.train_model.layers[3]
self.decoder_outputs, _, _ = self.train_model.layers[3].output
self.decoder_dense = self.train_model.layers[4]
self.decoder_outputs = self.train_model.layers[4].output
def init_predict_model(self):
# ENCODER MODEL <- note used in develop music process
# from encoder_input to encoder_states
# to give a context to decoder model
#---------------------------------
self.encoder_model = Model(self.encoder_inputs, self.encoder_states)
# DECODER MODEL
# From states (context) to sequance by generating firts element from context vector
# and starting element <GO>. Then adding predicted element as input to next cell, with
# updated states (context) by prevously generated element.
#
# INPUT_1 : state_h
# INPUT_2 : state_c
# INPUT_3 : y_train sized layer, that will be recursivly generated starting from <GO> sign
#
# INPUT -> LSTM -> DENSE
#
# OUTPUT : one-hot-encoded element of sequance
# OUTPUT : state_h (updated)
# OUTPUT : state_c (updated)
# -------------
# 1 layer: TWO INPUTS: decoder_state_h, decoder_state_c
self.decoder_state_input_h = Input(shape=(self.latent_dim,))
self.decoder_state_input_c = Input(shape=(self.latent_dim,))
self.decoder_states_inputs = [self.decoder_state_input_h, self.decoder_state_input_c]
# 2 layer: LSTM_1 output: element of sequance, lstm cell states
self.decoder_outputs, self.state_h, self.state_c = self.decoder_lstm(
self.decoder_inputs,
initial_state = self.decoder_states_inputs
)
self.decoder_states = [self.state_h, self.state_c]
# 3 layer: Dense output: one-hot-encoded representation of element of sequance
self.decoder_outputs = self.decoder_dense(self.decoder_outputs)
self.decoder_model = Model(
[self.decoder_inputs] + self.decoder_states_inputs,
[self.decoder_outputs] + self.decoder_states)
self.has_predict_model = True
def predict(self, input_seq=None, mode=None):
if not self.has_predict_model:
self.init_predict_model()
self.has_predict_model = True
if mode == 'generate':
# create a random context as starting point
h = np.random.rand(1,self.latent_dim)*2 - 1
c = np.random.rand(1,self.latent_dim)*2 - 1
states_value = [h, c]
else:
# get context from input sequance
states_value = self.encoder_model.predict(input_seq)
# make the empty decoder_input_data
# and create the starting <GO> element of decoder_input_data
target_seq = np.zeros((1, 1, self.transformer.y_vocab_size))
target_seq[0, 0, self.transformer.y_transform_dict['<GO>']] = 1.
# sequance generation loop of decoder model
stop_condition = False
decoded_sentence = []
# time = 0
while not stop_condition:
# INPUT_1 : target_seq : started from empty array with start <GO> char
# and recursivly updated by predicted elements
# INPUT_2 : states_value :context from encoder model or randomly generated in develop mode
# this can give as a 2 * latent_dim parameters to play with in manual generation
# OUTPUT_1 : output_tokens : one hot encoded predicted element of sequance
# OUTPUT_2,3 : h, c : context updated by predicted element
output_tokens, h, c = self.decoder_model.predict(
[target_seq] + states_value)
# get most likly element index
# translate from index to final (in normal form) preidcted element
# append it to output list
sampled_token_index = np.argmax(output_tokens[0, -1, :])
sampled_char = self.transformer.y_reverse_dict[sampled_token_index]
decoded_sentence.append(sampled_char)
# time += sampled_char[1]
# or time>=16
if (sampled_char == '<EOS>' or len(decoded_sentence) > self.transformer.y_max_seq_length ):
stop_condition = True
target_seq = np.zeros((1, 1, self.transformer.y_vocab_size))
target_seq[0, 0, sampled_token_index] = 1.
states_value = [h, c]
return decoded_sentence
def develop(self, mode='from_seq'):
# music generation for seq2seq for melody
input_seq_start = random_seed_generator(16,
self.transformer.x_max_seq_length,
self.transformer.x_vocab_size,
self.transformer.x_transform_dict,
self.transformer.x_reverse_dict)
input_data = seq_to_numpy(input_seq_start,
self.transformer.x_max_seq_length,
self.transformer.x_vocab_size,
self.transformer.x_transform_dict)
# generate sequnce iterativly for melody
input_seq = input_seq_start.copy()
melody = []
for i in range(4):
if mode == 'from_seq':
decoded_sentence = self.predict(input_data)[:-1]
elif mode == 'from_state':
decoded_sentence = self.predict(mode='generate')[:-1]
else:
raise ValueError('mode must be in {from_seq, from_state}')
melody.append(decoded_sentence)
input_seq.extend(decoded_sentence)
input_bars = stream_to_bars(input_seq, 4)
input_bars = input_bars[1:5]
input_seq = [note for bar in input_bars for note in bar]
input_data = seq_to_numpy(input_seq,
self.transformer.x_max_seq_length,
self.transformer.x_vocab_size,
self.transformer.x_transform_dict)
melody = [note for bar in melody for note in bar]
return melody
def random_seed_generator(time_of_seq, max_encoder_seq_length, num_encoder_tokens, input_token_index, reverse_input_char_index):
time = 0
random_seq = []
items = 0
stop_sign = False
while (time < time_of_seq):
seed = np.random.randint(0,num_encoder_tokens-1)
note = reverse_input_char_index[seed]
time += note[1]
if time > time_of_seq:
note_time = note[1] - (time-time_of_seq)
trimmed_note = (note[0],note_time)
try:
seed = input_token_index[trimmed_note]
random_seq.append(trimmed_note)
items += 1
except KeyError:
time -= note[1]
continue
else:
random_seq.append(note)
items += 1
if items > max_encoder_seq_length:
time = 0
random_seq = []
items = 0
stop_sign = False
return random_seq
# seq to numpy array:
def seq_to_numpy(seq, max_encoder_seq_length, num_encoder_tokens, input_token_index):
input_data = np.zeros(
(1, max_encoder_seq_length, num_encoder_tokens),
dtype='float32')
for t, char in enumerate(seq):
try:
input_data[0, t, input_token_index[char]] = 1.
except KeyError:
char_time = char[1]
_char = ((-1,), char_time)
except IndexError:
break
return input_data

28
project/readme.md
View File

@ -1,26 +1,12 @@
## MUSIC GENERATION USING DEEP LEARNING
## AUTHOR: CEZARY PUKOWNIK
### Files:
- midi.py - code for data extraction, and midi convertion
- train.py - code for model definition, and training session
- generate.py - code for model loading, predicting ang saving to midi_dir
- settings.py - file where deafult settings are stored
- readme.md - this file
### Directories:
- data/midi - directory where input midi are stored
- data/models - directory where trained models are stored
- data/output - directory where generated music is stored
- data/samples - directory where extracted data from midi is stored
- data/samples.npz - deprecated
## How to use:
1. Use midi.py to export data from midi files
> ./midi.py [midi_folder_path] [output_path]
1. In folder ./midi_packs make folder with midi files you want train on
2. Use extract.py to export data from midis
> ./extract.py [str: midi_pack_name] [str: name_of_session] --b [int: seq_len] -a [analize data first]
3. Use train.py to train model
> ./train.py [str: name_of_session] --b [int: batch_size] --l [int: latent_space] --e [int: epochs] --i [str: instrument] -r [reset]
4. Use generate.py to generate music from models
> ./generate.py [str: name_of_session] --n [number of generations] --m [mode {'from_seq','from_state'}]
2. Use train.py to train a model (this can take a while)
> ./train.py [input_training_data] [model_save_path] [epochs]
3. Use generate.py to generate music from trained models
> ./generate.py [trained_model_path] [output_path] [treshold]

313
project/settings.py
View File

@ -1,313 +0,0 @@
# paths
midi_dir = 'data/midi'
samples_dir = 'data/samples'
samples_path = 'data/samples.npz'
sample_preview_path = 'data/samples_preview.png'
model_path = 'data/autoencoder_model.h5'
generated_sample_path = 'data/output/generated_bar.npz'
generated_midi_path = 'data/output/generated_midi.mid'
generated_pianoroll_path = 'data/output/pianoroll.png'
# export_settings
midi_resolution = 96
beat_resolution = 24
beats_per_sample = 1
ignore_note_lenght = False
#train_settings
epochs = 1
#extras
midi_program = {
# Piano
1 : 'Acoustic Grand Piano',
2 : 'Bright Acoustic Piano',
3 : 'Electric Grand Piano',
4 : 'Honky-tonk Piano',
5 : 'Electric Piano 1',
6 : 'Electric Piano 2',
7 : 'Harpsichord',
8 : 'Clavi',
# Chromatic Percussion
9 : 'Celesta',
10 : 'Glockenspiel',
11 : 'Music Box',
12 : 'Vibraphone',
13 : 'Marimba',
14 : 'Xylophone',
15 : 'Tubular Bells',
16 : 'Dulcimer',
# Organ
17 : 'Drawbar Organ',
18 : 'Percussive Organ',
19 : 'Rock Organ',
20 : 'Church Organ',
21 : 'Reed Organ',
22 : 'Accordion',
23 : 'Harmonica',
24 : 'Tango Accordion',
# Guitar
25 : 'Acoustic Guitar (nylon)',
26 : 'Acoustic Guitar (steel)',
27 : 'Electric Guitar (jazz)',
28 : 'Electric Guitar (clean)',
29 : 'Electric Guitar (muted)',
30 : 'Overdriven Guitar',
31 : 'Distortion Guitar',
32 : 'Guitar harmonics',
# Bass
33 : 'Acoustic Bass',
34 : 'Electric Bass (finger)',
35 : 'Electric Bass (pick)',
36 : 'Fretless Bass',
37 : 'Slap Bass 1',
38 : 'Slap Bass 2',
39 : 'Synth Bass 1',
40 : 'Synth Bass 2',
# Strings
41 : 'Violin',
42 : 'Viola',
43 : 'Cello',
44 : 'Contrabass',
45 : 'Tremolo Strings',
46 : 'Pizzicato Strings',
47 : 'Orchestral Harp',
48 : 'Timpani',
# Ensemble
49 : 'String Ensemble 1',
50 : 'String Ensemble 2',
51 : 'SynthStrings 1',
52 : 'SynthStrings 2',
53 : 'Choir Aahs',
54 : 'Voice Oohs',
55 : 'Synth Voice',
56 : 'Orchestra Hit',
# Brass
57 : 'Trumpet',
58 : 'Trombone',
59 : 'Tuba',
60 : 'Muted Trumpet',
61 : 'French Horn',
62 : 'Brass Section',
63 : 'SynthBrass 1',
64 : 'SynthBrass 2',
# Reed
65 : 'Soprano Sax',
66 : 'Alto Sax',
67 : 'Tenor Sax',
68 : 'Baritone Sax',
69 : 'Oboe',
70 : 'English Horn',
71 : 'Bassoon',
72 : 'Clarinet',
# Pipe
73 : 'Piccolo',
74 : 'Flute',
75 : 'Recorder',
76 : 'Pan Flute',
77 : 'Blown Bottle',
78 : 'Shakuhachi',
79 : 'Whistle',
80 : 'Ocarina',
# Synth Lead
81 : 'Lead 1 (square)',
82 : 'Lead 2 (sawtooth)',
83 : 'Lead 3 (calliope)',
84 : 'Lead 4 (chiff)',
85 : 'Lead 5 (charang)',
86 : 'Lead 6 (voice)',
87 : 'Lead 7 (fifths)',
88 : 'Lead 8 (bass + lead)',
# Synth Pad
89 : 'Pad 1 (new age)',
90 : 'Pad 2 (warm)',
91 : 'Pad 3 (polysynth)',
92 : 'Pad 4 (choir)',
93 : 'Pad 5 (bowed)',
94 : 'Pad 6 (metallic)',
95 : 'Pad 7 (halo)',
96 : 'Pad 8 (sweep)',
# Synth Effects
97 : 'FX 1 (rain)',
98 : 'FX 2 (soundtrack)',
99 : 'FX 3 (crystal)',
100 : 'FX 4 (atmosphere)',
101 : 'FX 5 (brightness)',
102 : 'FX 6 (goblins)',
103 : 'FX 7 (echoes)',
104 : 'FX 8 (sci-fi)',
# Ethnic
105 : 'Sitar',
106 : 'Banjo',
107 : 'Shamisen',
108 : 'Koto',
109 : 'Kalimba',
110 : 'Bag pipe',
111 : 'Fiddle',
112 : 'Shanai',
# Percussive
113 : 'Tinkle Bell',
114 : 'Agogo',
115 : 'Steel Drums',
116 : 'Woodblock',
117 : 'Taiko Drum',
118 : 'Melodic Tom',
119 : 'Synth Drum',
120 : 'Reverse Cymbal',
# Sound Effects
121 : 'Guitar Fret Noise',
122 : 'Breath Noise',
123 : 'Seashore',
124 : 'Bird Tweet',
125 : 'Telephone Ring',
126 : 'Helicopter',
127 : 'Applause',
128 : 'Gunshot'
}
midi_group = {
# Piano
1 : 'Piano',
2 : 'Piano',
3 : 'Piano',
4 : 'Piano',
5 : 'Piano',
6 : 'Piano',
7 : 'Piano',
8 : 'Piano',
# Chromatic Percussion
9 : 'Chromatic_Percussion',
10 : 'Chromatic_Percussion',
11 : 'Chromatic_Percussion',
12 : 'Chromatic_Percussion',
13 : 'Chromatic_Percussion',
14 : 'Chromatic_Percussion',
15 : 'Chromatic_Percussion',
16 : 'Chromatic_Percussion',
# Organ
17 : 'Organ',
18 : 'Organ',
19 : 'Organ',
20 : 'Organ',
21 : 'Organ',
22 : 'Organ',
23 : 'Organ',
24 : 'Organ',
# Guitar
25 : 'Guitar',
26 : 'Guitar',
27 : 'Guitar',
28 : 'Guitar',
29 : 'Guitar',
30 : 'Guitar',
31 : 'Guitar',
32 : 'Guitar',
# Bass
33 : 'Bass',
34 : 'Bass',
35 : 'Bass',
36 : 'Bass',
37 : 'Bass',
38 : 'Bass',
39 : 'Bass',
40 : 'Bass',
# Strings
41 : 'Strings',
42 : 'Strings',
43 : 'Strings',
44 : 'Strings',
45 : 'Strings',
46 : 'Strings',
47 : 'Strings',
48 : 'Strings',
# Ensemble
49 : 'Ensemble',
50 : 'Ensemble',
51 : 'Ensemble',
52 : 'Ensemble',
53 : 'Ensemble',
54 : 'Ensemble',
55 : 'Ensemblee',
56 : 'Ensemble',
# Brass
57 : 'Brass',
58 : 'Brass',
59 : 'Brass',
60 : 'Brass',
61 : 'Brass',
62 : 'Brass',
63 : 'Brass',
64 : 'Brass',
# Reed
65 : 'Reed',
66 : 'Reed',
67 : 'Reed',
68 : 'Reed',
69 : 'Reed',
70 : 'Reed',
71 : 'Reed',
72 : 'Reed',
# Pipe
73 : 'Pipe',
74 : 'Pipe',
75 : 'Pipe',
76 : 'Pipe',
77 : 'Pipe',
78 : 'Pipe',
79 : 'Pipe',
80 : 'Pipe',
# Synth Lead
81 : 'Synth_Lead',
82 : 'Synth_Lead',
83 : 'Synth_Lead',
84 : 'Synth_Lead',
85 : 'Synth_Lead',
86 : 'Synth_Lead',
87 : 'Synth_Lead',
88 : 'Synth_Lead',
# Synth Pad
89 : 'Synth_Pad',
90 : 'Synth_Pad',
91 : 'Synth_Pad',
92 : 'Synth_Pad',
93 : 'Synth_Pad',
94 : 'Synth_Pad',
95 : 'Synth_Pad',
96 : 'Synth_Pad',
# Synth Effects
97 : 'Synth_Effects',
98 : 'Synth_Effects',
99 : 'Synth_Effects',
100 : 'Synth_Effects',
101 : 'Synth_Effects',
102 : 'Synth_Effects',
103 : 'Synth_Effects',
104 : 'Synth_Effects',
# Ethnic
105 : 'Ethnic',
106 : 'Ethnic',
107 : 'Ethnic',
108 : 'Ethnic',
109 : 'Ethnic',
110 : 'Ethnic',
111 : 'Ethnic',
112 : 'Ethnic',
# Percussive
113 : 'Percussive',
114 : 'Percussive',
115 : 'Percussive',
116 : 'Percussive',
117 : 'Percussive',
118 : 'Percussive',
119 : 'Percussive',
120 : 'Percussive',
# Sound Effects
121 : 'Sound_Effects',
122 : 'Sound_Effects',
123 : 'Sound_Effects',
124 : 'Sound_Effects',
125 : 'Sound_Effects',
126 : 'Sound_Effects',
127 : 'Sound_Effects',
128 : 'Sound_Effects'
}

BIN
project/settings.pyc
View File

Binary file not shown.

123
project/train.py
View File

@ -1,68 +1,79 @@
#!python3
#!/usr/bin/env python3
import os
import sys
import pickle
import settings
import keras
import argparse
import warnings
from model import Seq2SeqModel
from extract import make_folder_if_not_exist
import numpy as np
from keras.layers import Input, Dense, Conv2D, Flatten, LSTM, Dropout, TimeDistributed, RepeatVector, Activation, Bidirectional, Reshape
from keras.models import Model, Sequential
from keras.utils.np_utils import to_categorical
# TODO:
# FIXME:
def parse_argv():
parser = argparse.ArgumentParser()
parser.add_argument('n', help='name for experiment', type=str)
parser.add_argument('--b', help='batch_size', type=int)
parser.add_argument('--l', help='latent_dim', type=int)
parser.add_argument('--e', help='epochs', type=int)
parser.add_argument('--i', help='refrance to instrument to train, if you want to train only one instument')
parser.add_argument('-r', help='reset, use when you want to reset waights and train from scratch', action='store_true')
args = parser.parse_args()
return args
def load_data(samples_path):
print('Loading... {}'.format(train_data_path))
train_X = np.load(train_data_path, allow_pickle=True)['arr_0']
train_y = np.load(train_data_path, allow_pickle=True)['arr_1']
return train_X, train_y
def load_workflow():
workflow_path = os.path.join('training_sets', EXPERIMENT_NAME, 'workflow.pkl')
if os.path.isfile(workflow_path):
model_workflow = pickle.load(open(workflow_path,'rb'))
else:
raise FileNotFoundError(f'There is no workflow.pkl file in trainig_sets/{EXPERIMENT_NAME}/ folder')
return model_workflow
# TODO: make transformer class with fit, transform and reverse definitions
def preprocess_samples(train_X, train_y):
vocab_X = np.unique(train_X)
vocab_y = np.unique(train_y)
vocab = np.concatenate([vocab_X, vocab_y])
n_vocab = vocab.shape[0]
note_to_int = dict((note, number) for number, note in enumerate(vocab))
int_to_note = dict((number, note) for number, note in enumerate(vocab))
_train_X = []
_train_y = []
for sample in train_X:
# TODO: add normalizasion
_train_X.append([note_to_int[note] for note in sample])
def train_models(model_workflow):
instruments = [instrument if how == 'melody' else instrument[1] for key, (instrument, how) in model_workflow.items()]
# make_folder_if_not_exist(os.mkdir(os.path.join('models',EXPERIMENT_NAME)))
found = False
for instrument in instruments:
train_X = np.array(_train_X).reshape(train_X.shape[0], train_X.shape[1], 1)
train_y = np.array([note_to_int[note] for note in train_y]).reshape(-1,1)
train_y = to_categorical(train_y)
if INSTRUMENT == None or INSTRUMENT == instrument:
data_path = os.path.join('training_sets', EXPERIMENT_NAME, instrument.lower() + '_data.pkl')
model_path = os.path.join('models', EXPERIMENT_NAME, f'{instrument.lower()}_model.h5')
return train_X, train_y, n_vocab, int_to_note
x_train, y_train, _ = pickle.load(open(data_path,'rb'))
model = Seq2SeqModel(LATENT_DIM, x_train, y_train)
if os.path.isfile(model_path) and not RESET:
model.load(model_path)
train_data_path = sys.argv[1]
print(f'Training: {instrument}')
model.fit(BATCH_SIZE, EPOCHS, callbacks=[])
model.save(model_path)
found = True
train_X, train_y = load_data(train_data_path)
train_X, train_y, n_vocab, int_to_note = preprocess_samples(train_X, train_y)
if not found:
raise ValueError(f'Instrument not found. Use one of the {instruments}')
if __name__ == '__main__':
save_model_path = sys.argv[2]
epochs = int(sys.argv[3])
warnings.filterwarnings("ignore")
args = parse_argv()
EXPERIMENT_NAME = args.n
BATCH_SIZE = args.b
LATENT_DIM = args.l
EPOCHS = args.e
RESET = args.r
INSTRUMENT = args.i
model = Sequential()
model.add(LSTM(512, input_shape=(train_X.shape[1], train_X.shape[2]), return_sequences=True))
model.add(Dropout(0.3))
model.add(LSTM(512, return_sequences=True))
model.add(Dropout(0.3))
model.add(LSTM(512))
model.add(Dense(256))
model.add(Dropout(0.3))
model.add(Dense(n_vocab))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
# This code will train our model, with given by parameter number of epochs
print('Training...')
model.fit(train_X, train_y, epochs=epochs, batch_size=64)
# it saves model, and additional informations of model
# that is needed to generate music from it
pickle.dump(model, open(save_model_path,'wb'))
pickle.dump((int_to_note, n_vocab, train_X.shape[1]), open('{}_dict'.format(save_model_path),'wb'))
print('Done!')
print("Model saved to: {}".format(save_model_path))
# default settings if not args passed
if not BATCH_SIZE:
BATCH_SIZE = 32
if not LATENT_DIM:
LATENT_DIM = 256
if not EPOCHS:
EPOCHS = 1
if not RESET:
RESET = False
train_models(load_workflow())