From fede93c36ed9aabff39528a8aa7e61e973ab83b8 Mon Sep 17 00:00:00 2001
From: Cezary Pukownik <cezpuk@st.amu.edu.pl>
Date: Sat, 1 Jun 2019 17:05:38 +0200
Subject: [PATCH] get it working, on music21 and sequence style enoding

---
 project/generate.py | 57 +++++++++++++++++++++++--------
 project/midi.py     | 82 +++++++++++++++++++--------------------------
 project/train.py    | 53 ++++++++++++++++++++++-------
 3 files changed, 118 insertions(+), 74 deletions(-)

diff --git a/project/generate.py b/project/generate.py
index 9a5023b..14e78ff 100644
--- a/project/generate.py
+++ b/project/generate.py
@@ -1,7 +1,6 @@
 #!/usr/bin/env python3
 
 import numpy as np
-import midi
 import tensorflow as tf
 import pypianoroll as roll
 from keras.layers import Input, Dense, Conv2D
@@ -11,29 +10,59 @@ from keras.layers import Input, Dense, Conv2D, Flatten, LSTM, Dropout, TimeDistr
 from keras.models import Model, Sequential
 import matplotlib.pyplot as plt
 import settings
+import random
 import pickle
+from tqdm import trange, tqdm
 import sys
+from music21 import converter, instrument, note, chord, stream
 
 trained_model_path = sys.argv[1]
 output_path = sys.argv[2]
-treshold = float(sys.argv[3])
 
 # load and predict
+print('Loading... {}'.format(trained_model_path))
 model = pickle.load(open(trained_model_path, 'rb'))
+int_to_note = pickle.load(open('{}_dict'.format(trained_model_path), 'rb'))
 
-music = np.empty((4,96,128))
-for x in range(4):
-    generate_seed = np.random.randint(0, 127, 12288).reshape(1,96,128)
-    music[x] = model.predict(generate_seed).reshape(96,128)
+seed = [random.randint(0,50) for x in range(8)]
 
-generated_sample = music.reshape(4*96,128)
+music = []
 
-# binarize generated music
-generated_sample = generated_sample > treshold * generated_sample.max()
-# generated_sample = np.clip(generated_sample,0,1) * 128
+print('Generating...')
+for i in trange(500):
+    predicted_vector = model.predict(np.array(seed).reshape(1,8,1))
+    predicted_index = np.argmax(predicted_vector)
 
-# save to midi
-generated_midi = midi.to_midi(generated_sample, output_path='{}.mid'.format(output_path), is_drum=True, program=0,  )
+    music.append(int_to_note[predicted_index])
 
-#save plot for preview
-roll.plot(generated_midi, filename='{}.png'.format(output_path))
+    seed.append(predicted_index)
+    seed = seed[1:9]
+
+
+print('Saving...')
+offset = 0
+output_notes = []
+for event in tqdm(music):
+    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)
+    else:
+        new_note = note.Note(event)
+        new_note.offset = offset
+        new_note.storedInstrument = instrument.Piano()
+        output_notes.append(new_note)
+
+    offset += 0.5
+
+midi_stream = stream.Stream(output_notes)
+
+midi_stream.write('midi', fp='{}.mid'.format(output_path))
+
+print('Done!')
diff --git a/project/midi.py b/project/midi.py
index 447bbc2..e0a97bd 100644
--- a/project/midi.py
+++ b/project/midi.py
@@ -9,74 +9,62 @@ from math import floor
 import sys
 from collections import defaultdict
 import pickle
+from music21 import converter, instrument, note, chord, stream
+import music21
 
 midi_folder_path = sys.argv[1]
 output_path = sys.argv[2]
 
-def to_samples(multitrack, midi_res=settings.midi_resolution, how='by_group'):
+def to_sequence(midi_path):
+    seq_by_instrument = defaultdict( lambda : [] )
+    midi_file = music21.converter.parse(midi_path)
+    stream = music21.instrument.partitionByInstrument(midi_file)
+    for part in stream:
+        for event in part:
+            if part.partName != None:
+                # TODO: add note lenght as parameter
+                if isinstance(event, music21.note.Note):
+                    # to_export_event = (str(event.pitch), event.quarterLength)
+                    to_export_event = str(event.pitch)
+                    seq_by_instrument[part.partName].append(to_export_event)
+                elif isinstance(event, music21.chord.Chord):
+                    to_export_event = ' '.join(str(note) for note in event.pitches)
+                    # to_export_event = (' '.join(str(note) for note in event.pitches), event.quarterLength)
+                    seq_by_instrument[part.partName].append(to_export_event)
 
-    #how = 'by_group', 'by_instrument', 'merged',
+    X_train_by_instrument = defaultdict( lambda : [] )
+    y_train_by_instrument = defaultdict( lambda : [] )
 
-    # TODO: add transpositions of every sample to every possible key transposition
-    # np.roll(sample, pitch_interval, axis=1) for transposition
-    # np.roll(sample, time_steps, axis=0) for time shifting
+    for instrument, sequence in seq_by_instrument.items():
+        for i in range(len(sequence)-8) :
+            X_train_by_instrument[instrument].append(np.array(sequence[i: i + 8])) # <seq lenth
+            y_train_by_instrument[instrument].append(np.array(sequence[i + 8]))
+            # TODO: Notes to integers
 
-    samples_by_instrument = defaultdict( lambda : [] )
-
-    for track in multitrack.tracks:
-
-        key = settings.midi_group[track.program + 1] if not track.is_drum else 'Drums'
-
-        # this makes pack of samples of N x 96 x 128 shape
-        number_of_beats = floor(track.pianoroll.shape[0] / midi_res)
-        track_pianoroll = track.pianoroll[: number_of_beats * midi_res]
-        track_beats = track_pianoroll.reshape(number_of_beats, midi_res, 128)
-
-        # save collected pack of data to dictionary with samples packs for groups of instruments
-        for sample in track_beats:
-            if sample.sum() != 0:
-                samples_by_instrument[key].append(sample)
-
-        # TODO: add posibility of choosing between saving samples to groups of instrument, or to every instrument separatly or with no differance
-        # TODO: add option, for looking only for one instrument/group
-        # TODO: add option for colecting, more than one beat per sample (min 4)
-
-    return samples_by_instrument
-
-def to_midi(samples, output_path=settings.generated_midi_path, program=0, tempo=120, is_drum=False, beat_resolution=settings.beat_resolution):
-    tracks = [roll.Track(samples, program=program, is_drum=is_drum)]
-    return_midi = roll.Multitrack(tracks=tracks, tempo=tempo, downbeat=[0, 96, 192, 288], beat_resolution=beat_resolution)
-    roll.write(return_midi, output_path)
-    return return_midi
-
-# TODO: Make optial function to erase information of note lenth - ??
-def ignore_note_lenght():
-    pass
+    return X_train_by_instrument, y_train_by_instrument
 
 def main():
     print('Exporting...')
 
-    samples_pack_by_instrument = defaultdict( lambda : list() )
+    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)
-            #load midi ro pypianoroll - Multirack
-            try:
-                multitrack = roll.parse(midi_file_path)
-            except:
-                # IDEA: Log errors, and save to file?
-                continue
+            _X_train, _y_train = to_sequence(midi_file_path)
 
-            for key, value in to_samples(multitrack).items():
-                samples_pack_by_instrument[key].extend(value)
+            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))
 
     # this is for intrument separation
     print('Saving...')
     if not os.path.exists(output_path):
         os.makedirs(output_path)
-    for key, value in tqdm(samples_pack_by_instrument.items()):
-        np.savez_compressed('{}/{}.npz'.format(output_path, key), np.array(value))
+    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))
 
     print('Done!')
 
diff --git a/project/train.py b/project/train.py
index a5a8ac3..8bf03c3 100644
--- a/project/train.py
+++ b/project/train.py
@@ -5,35 +5,62 @@ import settings
 from tensorflow.keras import layers
 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
 import numpy as np
 import sys
 import pickle
 
+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
+
+# TODO: make transformer class with fit, transform and reverse definitions
+def preprocess_samples(train_X, train_y):
+    vocab = np.unique(train_X)
+    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])
+
+    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)
+
+    return train_X, train_y, n_vocab, int_to_note
+
 train_data_path = sys.argv[1]
+
+train_X, train_y = load_data(train_data_path)
+train_X, train_y, n_vocab, int_to_note = preprocess_samples(train_X, train_y)
+
 save_model_path = sys.argv[2]
 epochs = int(sys.argv[3])
 
 model = Sequential()
-model.add(LSTM(128,input_shape=(96, 128),return_sequences=True))
+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(128))
-model.add(Dense(128))
+model.add(LSTM(512))
+model.add(Dense(256))
 model.add(Dropout(0.3))
-model.add(Dense(128*96))
+model.add(Dense(n_vocab))
 model.add(Activation('softmax'))
-model.add(Reshape((96, 128)))
 model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
 
-# load training data
-print('Traing Samples: {}'.format(train_data_path))
-train_X = np.load(train_data_path)['arr_0']
-
 # model training
-model.fit(train_X, train_X, epochs=epochs, batch_size=32)
+print('Training...')
+model.fit(train_X, train_y, epochs=epochs, batch_size=64)
 
 # save trained model
-pickle_path = '{}.pickle'.format(save_model_path)
-pickle.dump(model, open(pickle_path,'wb'))
-print("Model save to {}".format(pickle_path))
+pickle.dump(model, open(save_model_path,'wb'))
+pickle.dump(int_to_note, open('{}_dict'.format(save_model_path),'wb'))
+print('Done!')
+print("Model saved to: {}".format(save_model_path))