LSTM机器学习生成音乐

import keras
from keras.callbacks import ModelCheckpoint
from keras.models import Input, Model
from keras.layers import  Dropout, Dense,LSTM 
from keras.optimizers import Adam
from keras.utils import plot_model
# X_one_hot.shape[1:] = (100, 308)
input_tensor = Input(shape=X_one_hot.shape[1:])
lstm = LSTM(512,return_sequences=True)(input_tensor)
dropout = Dropout(0.3)(lstm)

lstm = LSTM(256)(dropout)
dropout = Dropout(0.3)(lstm)
# Y_one_hot.shape[-1] = 308
dense = Dense(Y_one_hot.shape[-1], activation='softmax')(dropout)

model = Model(inputs=input_tensor, outputs=dense)
# 画图
# plot_model(model, to_file='model.png', show_shapes=True, expand_nested=True, dpi=500)
optimizer = Adam(lr=0.001)
model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])
model.summary()import keras
from keras.callbacks import ModelCheckpoint
from keras.models import Input, Model
from keras.layers import  Dropout, Dense,LSTM 
from keras.optimizers import Adam
from keras.utils import plot_model
# X_one_hot.shape[1:] = (100, 308)
input_tensor = Input(shape=X_one_hot.shape[1:])
lstm = LSTM(512,return_sequences=True)(input_tensor)
dropout = Dropout(0.3)(lstm)

lstm = LSTM(256)(dropout)
dropout = Dropout(0.3)(lstm)
# Y_one_hot.shape[-1] = 308
dense = Dense(Y_one_hot.shape[-1], activation='softmax')(dropout)

model = Model(inputs=input_tensor, outputs=dense)
# 画图
# plot_model(model, to_file='model.png', show_shapes=True, expand_nested=True, dpi=500)
optimizer = Adam(lr=0.001)
model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])
model.summary()

训练

? 相比较于上一次的RNN写诗，这一次，我们可以将数据集全部放到内存中进行训练，因为此次数据集比较小，可以将其全部放到内存中。不过，还是建议将数据集放到GPU比较好的电脑上面跑（比如说，白嫖kaggle，hhh）。

filepath = "./{epoch}--weights{loss:.2f}.hdf5"
checkpoint = ModelCheckpoint(
    filepath,
    monitor='loss',
    verbose=0,
    save_best_only=True,
    mode='min'
)
callbacks_list = [checkpoint]
model.fit(X_one_hot, Y_one_hot, epochs=100, batch_size=2048,callbacks=callbacks_list)

def get_data(filename):
    """从文件中获取音符

    :param filename: [文件名]
    :type filename: [str]
    :return: [返回音符]
    :rtype: [list]
    """
    with open(filename) as f:
       all_notes = f.readlines()
    return [ note[:len(note)-1]  for note in all_notes]
# 从保存的数据集中获得数据
all_notes = get_data("data.txt")

# 加载模型
from keras.models import load_model
model = load_model("weights-804-0.01.hdf5")

from collections import Counter
from keras.utils import to_categorical

counter = Counter(all_notes)
note_count = sorted(counter.items(),key=lambda x : -x[1])
notes,_ = zip(*note_count)
# note到id的映射
note_to_id = {note:id for id,note in enumerate(notes)}
# id到note的映射
id_to_note = {id:note for id,note in enumerate(notes)}
# 构建X_train，目的是为了实现随机从X_one_hot选择一个数据，然后进行predict 
X_train = []
sequence_batch = 100
for i in range(len(all_notes)-sequence_batch):
    X_pre = all_notes[i:i+sequence_batch]
    X_train.append([note_to_id[note] for note in X_pre])
X_one_hot = to_categorical(X_train)

def predict_next(X_predict):
    """通过前100个音符，预测下一个音符

    :param X_predict: [前100个音符]
    :type X_predict: [list]
    :return: [下一个音符的id]
    :rtype: [int]
    """
    prediction = model.predict(X_predict)
    index = np.argmax(prediction)
    return index

import numpy as np
from music21 import converter, instrument, note, chord, stream
def generate_notes():
    """随机从X_one_hot抽取一个数据（长为100），然后进行predict，最后生成音乐

    :return: [note数组（['D5', '2.6', 'F#5', 'D3', ……]）]
    :rtype: [list]
    """
    # 随机从X_one_hot选择一个数据进行predict
    randindex = np.random.randint(0, len(X_one_hot) - 1)
    predict_input = X_one_hot[randindex]
    # music_output里面是一个数组，如['D5', '2.6', 'F#5', 'D3', 'E5', '2.6', 'G5', 'F#5']
    music_output = [id_to_note[id] for id in X_train[randindex]]
    # 产生长度为1000的音符序列
    for note_index in range(1000):
        prediction_input = np.reshape(predict_input, (1,X_one_hot.shape[1],X_one_hot.shape[2]))
        # 预测下一个音符id
        predict_index = predict_next(prediction_input)
        # 将id转换成音符
        music_note = id_to_note[predict_index]
        music_output.append(music_note)
        # X_one_hot.shape[-1] = 308
        one_hot_note = np.zeros(X_one_hot.shape[-1])
        one_hot_note[predict_index] = 1
        one_hot_note = np.reshape(one_hot_note,(1,X_one_hot.shape[-1]))
        # 重新构建LSTM的输入
        predict_input = np.concatenate((predict_input[1:],one_hot_note))
    return music_output

predict_notes = generate_notes()

create_music(predict_notes,"auto_music")