Our task here is to take some existing music data then train a model using this existing data. The model has to learn the patterns in music that we humans enjoy. Once it learns this, the model should be able to generate new music for us. It cannot simply copy-paste from the training data. It has to understand the patterns of music to generate new music. We here are not expecting our model to generate new music which is of professional quality, but we want it to generate a decent quality music which should be melodious and good to hear.
The Input data we are using for developing the model is from .mid file. Let's gain some domain Knowledge.
A MIDI file is not an audio recording. Rather, it is a set of instructions – for example, for pitch or tempo – and can use a thousand times less disk space than the equivalent recorded audio.
To process these files we use Music21
Music21 is a Python-based toolkit for computer-aided musicology.
People use music21 to answer questions from musicology using computers, to study large datasets of music, to generate musical examples, to teach fundamentals of music theory, to edit musical notation, study music and the brain, and to compose music (both algorithmically and directly).
pip install music21
Importing the necessary Libraries
from music21 import converter, instrument, note, chord
import glob
import pickle
import numpy as np
import pandas as pd
from keras.utils import np_utils
from keras.models import Sequential
from keras.layers import LSTM, Dense, Dropout, Activation
from keras.callbacks import ModelCheckpoint
from keras.utils import plot_model
import osThe .mid files are stored in music_notes folder and we are implementing the below code to extract the data from each file and store it in notes list. music21 library modules are utilized for parsing the files
notes=[]
for file in glob.glob('midi_songs/*.mid'):
midi=converter.parse(file)
print("Parsing %s" % file)
notes_to_parse=None
parts=instrument.partitionByInstrument(midi)
if parts:
notes_to_parse=parts.parts[0].recurse()
else:
notes_to_parse=midi.flat.notes
for element in notes_to_parse:
if isinstance(element,note.Note):
notes.append(str(element.pitch))
elif isinstance(element, chord.Chord):
notes.append('.'.join(str(n) for n in element.normalOrder))Parsing midi_songs\0fithos.mid
Parsing midi_songs\8.mid
Parsing midi_songs\ahead_on_our_way_piano.mid
Parsing midi_songs\AT.mid
Parsing midi_songs\balamb.mid
Parsing midi_songs\bcm.mid
Parsing midi_songs\BlueStone_LastDungeon.mid
Parsing midi_songs\braska.mid
Parsing midi_songs\caitsith.mid
Parsing midi_songs\Cids.mid
Parsing midi_songs\cosmo.mid
Parsing midi_songs\costadsol.mid
Parsing midi_songs\dayafter.mid
Parsing midi_songs\decisive.mid
Parsing midi_songs\dontbeafraid.mid
Parsing midi_songs\DOS.mid
Parsing midi_songs\electric_de_chocobo.mid
Parsing midi_songs\Eternal_Harvest.mid
Parsing midi_songs\EyesOnMePiano.mid
Parsing midi_songs\ff11_awakening_piano.mid
Parsing midi_songs\ff1battp.mid
Parsing midi_songs\FF3_Battle_(Piano).mid
Parsing midi_songs\FF3_Third_Phase_Final_(Piano).mid
Parsing midi_songs\ff4-airship.mid
Parsing midi_songs\Ff4-BattleLust.mid
Parsing midi_songs\ff4-fight1.mid
Parsing midi_songs\ff4-town.mid
Parsing midi_songs\FF4.mid
Parsing midi_songs\ff4pclov.mid
Parsing midi_songs\ff4_piano_collections-main_theme.mid
Parsing midi_songs\FF6epitaph_piano.mid
Parsing midi_songs\ff6shap.mid
Parsing midi_songs\Ff7-Cinco.mid
Parsing midi_songs\Ff7-Jenova_Absolute.mid
Parsing midi_songs\ff7-mainmidi.mid
Parsing midi_songs\Ff7-One_Winged.mid
Parsing midi_songs\ff7themep.mid
Parsing midi_songs\ff8-lfp.mid
Parsing midi_songs\FF8_Shuffle_or_boogie_pc.mid
Parsing midi_songs\FFIII_Edgar_And_Sabin_Piano.mid
Parsing midi_songs\FFIXQuMarshP.mid
Parsing midi_songs\FFIX_Piano.mid
Parsing midi_songs\FFVII_BATTLE.mid
Parsing midi_songs\FFX_-_Ending_Theme_(Piano_Version)_-_by_Angel_FF.mid
Parsing midi_songs\Fiend_Battle_(Piano).mid
Parsing midi_songs\Fierce_Battle_(Piano).mid
Parsing midi_songs\figaro.mid
Parsing midi_songs\Finalfantasy5gilgameshp.mid
Parsing midi_songs\Finalfantasy6fanfarecomplete.mid
Parsing midi_songs\Final_Fantasy_7_-_Judgement_Day_Piano.mid
Parsing midi_songs\Final_Fantasy_Matouyas_Cave_Piano.mid
Parsing midi_songs\fortresscondor.mid
Parsing midi_songs\Fyw_piano.mid
Parsing midi_songs\gerudo.mid
Parsing midi_songs\goldsaucer.mid
Parsing midi_songs\Gold_Silver_Rival_Battle.mid
Parsing midi_songs\great_war.mid
Parsing midi_songs\HighwindTakestotheSkies.mid
Parsing midi_songs\In_Zanarkand.mid
Parsing midi_songs\JENOVA.mid
Parsing midi_songs\Kingdom_Hearts_Dearly_Beloved.mid
Parsing midi_songs\Kingdom_Hearts_Traverse_Town.mid
Parsing midi_songs\Life_Stream.mid
Parsing midi_songs\lurk_in_dark.mid
Parsing midi_songs\mining.mid
Parsing midi_songs\Oppressed.mid
Parsing midi_songs\OTD5YA.mid
Parsing midi_songs\path_of_repentance.mid
Parsing midi_songs\pkelite4.mid
Parsing midi_songs\Rachel_Piano_tempofix.mid
Parsing midi_songs\redwings.mid
Parsing midi_songs\relmstheme-piano.mid
Parsing midi_songs\roseofmay-piano.mid
Parsing midi_songs\rufus.mid
Parsing midi_songs\Rydia_pc.mid
Parsing midi_songs\sandy.mid
Parsing midi_songs\sera_.mid
Parsing midi_songs\sobf.mid
Parsing midi_songs\Still_Alive-1.mid
Parsing midi_songs\Suteki_Da_Ne_(Piano_Version).mid
Parsing midi_songs\thenightmarebegins.mid
Parsing midi_songs\thoughts.mid
Parsing midi_songs\tifap.mid
Parsing midi_songs\tpirtsd-piano.mid
Parsing midi_songs\traitor.mid
Parsing midi_songs\ultimafro.mid
Parsing midi_songs\ultros.mid
Parsing midi_songs\VincentPiano.mid
Parsing midi_songs\ViviinAlexandria.mid
Parsing midi_songs\waltz_de_choco.mid
Parsing midi_songs\Zelda_Overworld.mid
Parsing midi_songs\z_aeristhemepiano.mid
notes[1:20]['E2',
'4.9',
'4.9',
'4.9',
'4.9',
'4.9',
'4.9',
'4.9',
'11.4',
'4.9',
'11.4',
'4.9',
'4.9',
'4.9',
'4.9',
'4.9',
'0.4',
'E2',
'4.9']
len(set(notes))358
saving the data in notes file for futher reuse
with open('notes', 'wb') as filepath:
pickle.dump(notes, filepath)The Neural network we are creating has LSTM Layers after Input Layer. we need to prepare our data as per it's requirement. At present our data is just a list of notes. We need to create a list of sequences as features and list of their next note as Target variable
Ex:
A sequence with increment 10
10,20,30,40,50,60
If we take 3 steps and and our data has single feature
| x | y |
|---|---|
| 10 20 30 | 40 |
| 20 30 40 | 50 |
| 30 40 50 | 60 |
and If we give 40,50,60 our model has to predict output as 70.
Our data example:
suppose we have only four notes. Let them be A, B, C, D
and input sequence is AABACCDB
we will create dictionary mapping them to integers
| A | B | C | D |
|---|---|---|---|
| 0 | 1 | 2 | 3 |
Now our input sequence became 00102231
Now we will create a list of sequences X
| x | y |
|---|---|
| 0 0 1 | 0 |
| 0 1 0 | 2 |
| 1 0 2 | 2 |
| 0 2 2 | 3 |
| 2 2 3 | 1 |
Now Y is one hot encoded.
| x | y |
|---|---|
| 0 0 1 | 0 0 0 0 |
| 0 1 0 | 0 0 1 0 |
| 1 0 2 | 0 0 1 0 |
| 0 2 3 | 0 0 0 1 |
| 2 3 4 | 0 1 0 0 |
seq_length=100
pitchnames=sorted(set(item for item in notes))Creating a dictionary mapping the pitched to integers
n_vocab=len(set(notes))
sequence_length = 100
# get all pitch names
pitchnames = sorted(set(item for item in notes))
# create a dictionary to map pitches to integers
note_to_int = dict((note, number) for number, note in enumerate(pitchnames))
network_input = []
network_output = []
# create input sequences and the corresponding outputs
for i in range(0, len(notes) - sequence_length, 1):
sequence_in = notes[i:i + sequence_length]
sequence_out = notes[i + sequence_length]
network_input.append([note_to_int[char] for char in sequence_in])
network_output.append(note_to_int[sequence_out])
n_patterns = len(network_input)
# reshape the input into a format compatible with LSTM layers
network_input = np.reshape(network_input, (n_patterns, sequence_length, 1))
# normalize input
network_input = network_input / float(n_vocab)
network_output = np_utils.to_categorical(network_output)network_input.shape(57077, 100, 1)
network_output.shape(57077, 358)
Our model will take 100 notes and predict the 101 one and and now 102 note is produced by feeding 2-101 notes and so on...
Key Layer for our model is LSTM. Let's know a little bit about it.
- Forget Gate
Bob is nice person but Alice is evil
As soon as the first full stop after “person” is encountered, the forget gate realizes that there may be a change of context in the next sentence. As a result of this, the subject of the sentence is forgotten and the place for the subject is vacated. And when we start speaking about “Dan” this position of the subject is allocated to “Dan”. This process of forgetting the subject is brought about by the forget gate.
- Input Gate
Bob knows swimming. He told me over the phone that he served for navy for 4 years
Now the important information here is that “Bob” knows swimming and that he has served the Navy for four years. This can be added to the cell state, however, the fact that he told all this over the phone is a less important fact and can be ignored. This process of adding some new information can be done via the input gate.
- Output Gate
Bob fought single handedly with the enemy and died for his country. For his contributions brave____________
In this phrase, there could be a number of options for the empty space. But we know that the current input of ‘brave’, is an adjective that is used to describe a noun. Thus, whatever word follows, has a strong tendency of being a noun. And thus, Bob could be an apt output.
This job of selecting useful information from the current cell state and showing it out as an output is done via the output gate.
model = Sequential()
model.add(LSTM(
512,
input_shape=(network_input.shape[1], network_input.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')WARNING:tensorflow:From C:\Users\SRINIVAS\Anaconda3\envs\innomatics\lib\site-packages\tensorflow\python\framework\op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.
Instructions for updating:
Colocations handled automatically by placer.
WARNING:tensorflow:From C:\Users\SRINIVAS\Anaconda3\envs\innomatics\lib\site-packages\keras\backend\tensorflow_backend.py:3445: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.
Instructions for updating:
Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.
filepath='my_weights-improvement={epoch:02d}-{loss:.4f}-bigger.hdf5'
checkpoint=ModelCheckpoint(filepath,monitor='loss',verbose=0,save_best_only=True,mode='min')
callbacks_list=[checkpoint]
model.fit(net_input,net_output,epochs=100,batch_size=128,callbacks=callbacks_list)WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.cast instead.
Epoch 1/100
57077/57077 [==============================] - 584s 10ms/step - loss: 4.7701
Epoch 2/100
57077/57077 [==============================] - 573s 10ms/step - loss: 4.7976
Epoch 3/100
57077/57077 [==============================] - 571s 10ms/step - loss: 4.7119
Epoch 4/100
57077/57077 [==============================] - 569s 10ms/step - loss: 4.7356
Epoch 5/100
57077/57077 [==============================] - 568s 10ms/step - loss: 4.7034
Epoch 6/100
57077/57077 [==============================] - 568s 10ms/step - loss: 4.7244
Epoch 7/100
57077/57077 [==============================] - 565s 10ms/step - loss: 4.7030
Epoch 8/100
57077/57077 [==============================] - 567s 10ms/step - loss: 4.6946
Epoch 9/100
57077/57077 [==============================] - 566s 10ms/step - loss: 4.6919
Epoch 10/100
57077/57077 [==============================] - 565s 10ms/step - loss: 4.6375
Epoch 11/100
57077/57077 [==============================] - 565s 10ms/step - loss: 4.5706
Epoch 12/100
57077/57077 [==============================] - 563s 10ms/step - loss: 4.5414
Epoch 13/100
38528/57077 [===================>..........] - ETA: 3:03 - loss: 4.6084
len(set(notes))358
model.load_weights('weights_folder/my_weights-improvement=168-0.1347-bigger.hdf5')
model.summary()_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
lstm_1 (LSTM) (None, 100, 512) 1052672
_________________________________________________________________
dropout_1 (Dropout) (None, 100, 512) 0
_________________________________________________________________
lstm_2 (LSTM) (None, 100, 512) 2099200
_________________________________________________________________
dropout_2 (Dropout) (None, 100, 512) 0
_________________________________________________________________
lstm_3 (LSTM) (None, 512) 2099200
_________________________________________________________________
dense_1 (Dense) (None, 256) 131328
_________________________________________________________________
dropout_3 (Dropout) (None, 256) 0
_________________________________________________________________
dense_2 (Dense) (None, 358) 92006
_________________________________________________________________
activation_1 (Activation) (None, 358) 0
=================================================================
Total params: 5,474,406
Trainable params: 5,474,406
Non-trainable params: 0
_________________________________________________________________
filepath='weights_folder/my_weights-improvement={epoch:02d}-{loss:.4f}-bigger.hdf5'
checkpoint=ModelCheckpoint(filepath,monitor='loss',verbose=0,save_best_only=True,mode='min')
callbacks_list=[checkpoint]
model.fit(net_input,net_output,epochs=200,batch_size=128,callbacks=callbacks_list)Epoch 1/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2689
Epoch 2/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2668
Epoch 3/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2543
Epoch 4/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2554
Epoch 5/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2508
Epoch 6/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.2518
Epoch 7/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2478
Epoch 8/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.2414
Epoch 9/200
57077/57077 [==============================] - 564s 10ms/step - loss: 0.2404
Epoch 10/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2396
Epoch 11/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2312
Epoch 12/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2267
Epoch 13/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2219
Epoch 14/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2239
Epoch 15/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2198
Epoch 16/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2158
Epoch 17/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2112
Epoch 18/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2125
Epoch 19/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2071
Epoch 20/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.2042
Epoch 21/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2061
Epoch 22/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2031
Epoch 23/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.2009
Epoch 24/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.2001
Epoch 25/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1974
Epoch 26/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1915
Epoch 27/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1932
Epoch 28/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1932
Epoch 29/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1868
Epoch 30/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1876
Epoch 31/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1874
Epoch 32/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1885
Epoch 33/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1859
Epoch 34/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1798
Epoch 35/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1852
Epoch 36/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1822
Epoch 37/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1816
Epoch 38/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1803
Epoch 39/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1771
Epoch 40/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1728
Epoch 41/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1737
Epoch 42/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1758
Epoch 43/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1714
Epoch 44/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1721
Epoch 45/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1713
Epoch 46/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1700
Epoch 47/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1653
Epoch 48/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1686
Epoch 49/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1721
Epoch 50/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1642
Epoch 51/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1642
Epoch 52/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1648
Epoch 53/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1574
Epoch 54/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1603
Epoch 55/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1658
Epoch 56/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1582
Epoch 57/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1589
Epoch 58/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1620
Epoch 59/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1550
Epoch 60/200
57077/57077 [==============================] - 559s 10ms/step - loss: 0.1588
Epoch 61/200
57077/57077 [==============================] - 554s 10ms/step - loss: 0.1584
Epoch 62/200
57077/57077 [==============================] - 570s 10ms/step - loss: 0.1537
Epoch 63/200
57077/57077 [==============================] - 574s 10ms/step - loss: 0.1556
Epoch 64/200
57077/57077 [==============================] - 569s 10ms/step - loss: 0.1540
Epoch 65/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1529
Epoch 66/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1564
Epoch 67/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1526
Epoch 68/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1574
Epoch 69/200
57077/57077 [==============================] - 563s 10ms/step - loss: 0.1565
Epoch 70/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1510
Epoch 71/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1495
Epoch 72/200
57077/57077 [==============================] - 561s 10ms/step - loss: 0.1489
Epoch 73/200
57077/57077 [==============================] - 562s 10ms/step - loss: 0.1484
Epoch 74/200
39936/57077 [===================>..........] - ETA: 2:48 - loss: 0.1467
# Save the model
model.save('music_generator.h5')
plot_model(model, to_file='model.png')# pick a random sequence from the input as a starting point for the prediction
start = np.random.randint(0, len(network_input)-1)
int_to_note = dict((number, note) for number, note in enumerate(pitchnames))
pattern = network_input[start]
prediction_output = []
pattern=list(pattern)
for note_index in range(500):
prediction_input=np.reshape(pattern,(1,len(pattern),1))
prediction_input=prediction_input/float(n_vocab)
prediction=model.predict(prediction_input,verbose=0)
index=np.argmax(prediction)
result=int_to_note[index]
prediction_output.append(result)
pattern.append(index)
pattern=pattern[1:len(pattern)]
prediction_output[:20]['2.5',
'0.5',
'2.5',
'3.8',
'0.5',
'0.5',
'0.5',
'2.5',
'0.5',
'2.5',
'0.5',
'0.5',
'0.5',
'2.5',
'0.5',
'2.5',
'3.8',
'0.5',
'0.5',
'0.5']
from music21 import instrument, note, stream, chord
offset = 0
output_notes = []
# create note and chord objects based on the values generated by the model
for pattern in prediction_output:
# pattern is a chord
if ('.' in pattern) or pattern.isdigit():
notes_in_chord = pattern.split('.')
notes = []
for current_note in notes_in_chord:
new_note = note.Note(int(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)
# pattern is a note
else:
new_note = note.Note(pattern)
new_note.offset = offset
new_note.storedInstrument = instrument.Piano()
output_notes.append(new_note)
# increase offset each iteration so that notes do not stack
offset += 0.5
midi_stream = stream.Stream(output_notes)
midi_stream.write('midi', fp='new_test_output_168.mid')'new_test_output_168.mid'