本文主要是介绍Assignment | 05-week1 -Improvise a Jazz Solo with an LSTM Network，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！
该系列仅在原课程基础上课后作业部分添加个人学习笔记，如有错误，还请批评指教。- ZJ
Coursera 课程 |deeplearning.ai |网易云课堂
CSDN：http://blog.csdn.net/JUNJUN_ZHAO/article/details/79420913
Welcome to your final programming assignment of this week! In this notebook, you will implement a model that uses an LSTM to generate music. You will even be able to listen to your own music at the end of the assignment.
欢迎来到本周的最终编程任务！ 在这次笔记中，您将实现一个使用 LSTM 生成音乐的模型。 你甚至可以在作业结束时听取自己的音乐。
You will learn to:
- Apply an LSTM to music generation.
- Generate your own jazz music with deep learning.
- 将LSTM应用于音乐生成。
- 深度学习生成自己的爵士音乐。
Please run the following cell to load all the packages required in this assignment. This may take a few minutes.
from __future__ import print_function
import IPython
import sys
from music21 import *
import numpy as np
from grammar import *
from qa import *
from preprocess import *
from music_utils import *
from data_utils import *
from keras.models import load_model, Model
from keras.layers import Dense, Activation, Dropout, Input, LSTM, Reshape, Lambda, RepeatVector
from keras.initializers import glorot_uniform
from keras.utils import to_categorical
from keras.optimizers import Adam
from keras import backend as Kd:\program files\python36\lib\site-packages\h5py\__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.from ._conv import register_converters as _register_converters
Using TensorFlow backend.
1 - Problem statement
You would like to create a jazz music piece specially for a friend’s birthday. However, you don’t know any instruments or music composition. Fortunately, you know deep learning and will solve this problem using an LSTM netwok.
You will train a network to generate novel jazz solos in a style representative of a body of performed work.
您想为朋友的生日专门制作爵士乐曲。 但是，你不知道任何乐器或音乐作品。 幸运的是，你知道深度学习，并将使用 LSTM 网络来解决这个问题。
您将训练一个网络，以演奏作品的代表作风格代表爵士乐独奏。
1.1 - Dataset
You will train your algorithm on a corpus of Jazz music. Run the cell below to listen to a snippet of the audio from the training set:
IPython.display.Audio('./data/30s_seq.mp3')音乐片段，无法上传。
We have taken care of the preprocessing of the musical data to render it in terms of musical “values.” You can informally think of each “value” as a note, which comprises a pitch and a duration. For example, if you press down a specific piano key for 0.5 seconds, then you have just played a note. In music theory, a “value” is actually more complicated than this–specifically, it also captures the information needed to play multiple notes at the same time. For example, when playing a music piece, you might press down two piano keys at the same time (playng multiple notes at the same time generates what’s called a “chord”和弦). But we don’t need to worry about the details of music theory for this assignment. For the purpose of this assignment, all you need to know is that we will obtain a dataset of values, and will learn an RNN model to generate sequences of values.
我们已经关注音乐数据的预处理，以音乐的“value”来表达它。你可以非正式地将每个“value”看作一个音符，它包含一个音高和一个持续时间。例如，如果您按下特定钢琴键0.5秒，那么您刚刚弹奏了一个音符。在音乐理论中，“value”实际上比这更复杂 - 具体来说，它还捕获了同时播放多个音符所需的信息。例如，在播放音乐作品时，可以同时按下两个钢琴键（同时播放多个音符生成所谓的“和弦”）。但是我们不需要担心这个任务的音乐理论的细节。为了这个任务的目的，你需要知道的是，我们将获得一个值的数据集，并将学习一个 RNN 模型来生成序列值。
Our music generation system will use 78 unique values. Run the following code to load the raw music data and preprocess it into values. This might take a few minutes.
我们的音乐生成系统将使用78个独特的值。运行以下代码以加载原始音乐数据并将其预处理为值。这可能需要几分钟的时间。
X, Y, n_values, indices_values = load_music_utils()
print('shape of X:', X.shape)
print('number of training examples:', X.shape[0])
print('Tx (length of sequence):', X.shape[1])
print('total # of unique values:', n_values)
print('Shape of Y:', Y.shape)
# 共 60 个训练样本，每个训练样本的 序列长度是 30 ，音符和弦相关的汇集表 共 78 shape of X: (60, 30, 78)
number of training examples: 60
Tx (length of sequence): 30
total # of unique values: 78
Shape of Y: (30, 60, 78)
You have just loaded the following:
X: This is an (m, Tx T x , 78) dimensional array. We have m training examples, each of which is a snippet of Tx=30 T x = 30 musical values. At each time step, the input is one of 78 different possible values, represented as a one-hot vector. Thus for example, X[i,t,:] is a one-hot vector representating the value of the i-th example at time t.X：这是一个（m, Tx T x ,78）维数组。 我们有 m 个训练样例，每个样例都是 Tx=30 T x = 30 音乐值的片段。 在每个时间步，输入是78个不同的可能值之一，表示为一个one-hot vector。 因此，例如， X[i,t,:]是表示第 i 个示例在时间 t 的值的 one-hot vector。Y: This is essentially the same asX, but shifted one step to the left (to the past). Similar to the dinosaurus assignment, we’re interested in the network using the previous values to predict the next value, so our sequence model will try to predict y⟨t⟩ y ⟨ t ⟩ given x⟨1⟩,…,x⟨t⟩ x ⟨ 1 ⟩ , … , x ⟨ t ⟩ . However, the data inYis reordered to be dimension (Ty,m,78) ( T y , m , 78 ) , where Ty=Tx T y = T x . This format makes it more convenient to feed to the LSTM later.Y：这与X基本相同，但向左移一步（到过去）。 与恐龙分配类似，我们对使用先前值预测下一个值的网络感兴趣，因此我们的序列模型将尝试预测 y⟨t⟩ y ⟨ t ⟩ 给出 x⟨1⟩,…,x⟨t⟩ x ⟨ 1 ⟩ , … , x ⟨ t ⟩ 。 然而，Y中的数据被重新排序为 (Ty,m,78) ( T y , m , 78 ) ，其中 Ty=Tx T y = T x 。 这种格式使得稍后进入 LSTM 更方便。n_values: The number of unique values in this dataset. This should be 78.indices_values: python dictionary mapping from 0-77 to musical values.
1.2 - Overview of our model
Here is the architecture of the model we will use. This is similar to the Dinosaurus model you had used in the previous notebook, except that in you will be implementing it in Keras. The architecture is as follows:
这是我们将使用的模型的架构。 这与您在前一个笔记中使用的 Dinosaurus 模型类似，只不过您将在 Keras 中实现它。 架构如下：
2 - Building the model
In this part you will build and train a model that will learn musical patterns. To do so, you will need to build a model that takes in X of shape (m,Tx,78) ( m , T x , 78 ) and Y of shape (Ty,m,78) ( T y , m , 78 ) . We will use an LSTM with 64 dimensional hidden states. Lets set n_a = 64.
在这部分你将建立和训练一个将学习音乐模式的模型。 为此，您需要构建一个模型，该模型需要形状为 (m,Tx,78) ( m , T x , 78 ) 和形状为 (Ty,m,78) ( T y , m , 78 ) 的Y。 我们将使用 64 维隐藏状态的LSTM。 让我们设置n_a = 64。
n_a = 64 Here’s how you can create a Keras model with multiple inputs and outputs. If you’re building an RNN where even at test time entire input sequence x⟨1⟩,x⟨2⟩,…,x⟨Tx⟩ x ⟨ 1 ⟩ , x ⟨ 2 ⟩ , … , x ⟨ T x ⟩ were given in advance, for example if the inputs were words and the output was a label, then Keras has simple built-in functions to build the model. However, for sequence generation, at test time we don’t know all the values of x⟨t⟩ x ⟨ t ⟩ in advance; instead we generate them one at a time using x⟨t⟩=y⟨t−1⟩ x ⟨ t ⟩ = y ⟨ t − 1 ⟩ . So the code will be a bit more complicated, and you’ll need to implement your own for-loop to iterate over the different time steps.
以下是如何创建具有多个输入和输出的 Keras 模型。 如果你正在建立一个 RNN，即使在测试时间，整个输入序列 x⟨1⟩,x⟨2⟩,…,x⟨Tx⟩ x ⟨ 1 ⟩ , x ⟨ 2 ⟩ , … , x ⟨ T x ⟩ 预先给定，例如，如果输入是单词并且输出是标签，则Keras具有简单的内置函数来构建模型。 但是，对于序列生成，在测试时我们并不知道 x⟨t⟩ x ⟨ t ⟩ 的所有值， 相反，我们使用 x⟨t⟩=y⟨t−1⟩ x ⟨ t ⟩ = y ⟨ t − 1 ⟩ 一次一个地生成它们。 所以代码会更复杂一点，你需要实现自己的 for-loop 来遍历不同的时间步骤。
The function djmodel() will call the LSTM layer Tx T x times using a for-loop, and it is important that all Tx T x copies have the same weights. I.e., it should not re-initiaiize the weights every time—the Tx T x steps should have shared weights. The key steps for implementing layers with shareable weights in Keras are:
- Define the layer objects (we will use global variables for this).
- Call these objects when propagating the input.
函数djmodel（）会使用 for 循环调用 LSTM 层 Tx T x 次，所有 Tx T x 拷贝具有相同的权重是很重要的。 即，每次都不应该重新初始化权重— Tx T x 步骤应该具有共享权重。 在 Keras 中实现可分享权重的关键步骤是：
1.定义图层对象（我们将使用全局变量）。
2.传播输入时调用这些对象。
We have defined the layers objects you need as global variables. Please run the next cell to create them. Please check the Keras documentation to make sure you understand what these layers are: Reshape(), LSTM(), Dense().
reshapor = Reshape((1, 78)) # Used in Step 2.B of djmodel(), below
LSTM_cell = LSTM(n_a, return_state = True) # Used in Step 2.C
densor = Dense(n_values, activation='softmax') # Used in Step 2.DEach of reshapor, LSTM_cell and densor are now layer objects, and you can use them to implement djmodel(). In order to propagate a Keras tensor object X through one of these layers, use layer_object(X) (or layer_object([X,Y]) if it requires multiple inputs.). For example, reshapor(X) will propagate X through the Reshape((1,78)) layer defined above.
Exercise: Implement djmodel(). You will need to carry out 2 steps:
- Create an empty list “outputs” to save the outputs of the LSTM Cell at every time step.
Loop for t∈1,…,Tx t ∈ 1 , … , T x :
A. Select the “t”th time-step vector from X. The shape of this selection should be (78,). To do so, create a custom Lambda layer in Keras by using this line of code:
x = Lambda(lambda x: X[:,t,:])(X)Look over the Keras documentation to figure out what this does. It is creating a “temporary” or “unnamed” function (that’s what Lambda functions are) that extracts out the appropriate one-hot vector, and making this function a Keras Layer object to apply to X.
B. Reshape x to be (1,78). You may find the `reshapor()` layer (defined below) helpful.C. Run x through one step of LSTM_cell. Remember to initialize the LSTM_cell with the previous step's hidden state $a$ and cell state $c$. Use the following formatting:
a, _, c = LSTM_cell(input_x, initial_state=[previous hidden state, previous cell state])D. Propagate the LSTM's output activation value through a dense+softmax layer using `densor`. E. Append the predicted value to the list of "outputs"
# GRADED FUNCTION: djmodeldef djmodel(Tx, n_a, n_values):"""Implement the modelArguments:Tx -- length of the sequence in a corpusn_a -- the number of activations used in our modeln_values -- number of unique values in the music data Returns:model -- a keras model with the """# Define the input of your model with a shape X --(m, Tx, n_values) n_values = 78 X = Input(shape=(Tx, n_values))# Define s0, initial hidden state for the decoder LSTM 隐藏状态 ： n_a = 64 a0 = Input(shape=(n_a,), name='a0')c0 = Input(shape=(n_a,), name='c0')a = a0c = c0### START CODE HERE ### # Step 1: Create empty list to append the outputs while you iterate (≈1 line)outputs = []# Step 2: Loopfor t in range(Tx):# Step 2.A: select the "t"th time step vector from X. (m, Tx, n_values)x = Lambda(lambda x: X[:, t, :])(X)# Step 2.B: Use reshapor to reshape x to be (1, n_values) (≈1 line)x = reshapor(x)# Step 2.C: Perform one step of the LSTM_cella, _, c = LSTM_cell(x, initial_state=[a, c])# Step 2.D: Apply densor to the hidden state output of LSTM_Cellout = densor(a)# Step 2.E: add the output to "outputs"outputs.append(out)# Step 3: Create model instancemodel = Model(input=[X, a0, c0], outputs=outputs)### END CODE HERE ###return modelRun the following cell to define your model. We will use Tx=30, n_a=64 (the dimension of the LSTM activations LSTM 激活函数的维度), and n_values=78. This cell may take a few seconds to run.
model = djmodel(Tx = 30 , n_a = 64, n_values = 78)C:\Users\qhtf\AppData\Roaming\Python\Python36\site-packages\ipykernel_launcher.py:44: UserWarning: Update your `Model` call to the Keras 2 API: `Model(outputs=[<tf.Tenso..., inputs=[<tf.Tenso...)`
You now need to compile your model to be trained. We will Adam and a categorical cross-entropy loss.
opt = Adam(lr=0.01, beta_1=0.9, beta_2=0.999, decay=0.01) # 超参数 learning_rate 学习率 ，beta_1, beta_2 的设定 decay:Learning rate decay over each update.model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
Finally, lets initialize a0 and c0 for the LSTM’s initial state to be zero.
m = 60
a0 = np.zeros((m, n_a))
c0 = np.zeros((m, n_a))Lets now fit the model! We will turn Y to a list before doing so, since the cost function expects Y to be provided in this format (one list item per time-step). So list(Y) is a list with 30 items, where each of the list items is of shape (60,78). Lets train for 100 epochs. This will take a few minutes.
让我们现在 fit 模型！ 在这样做之前，我们将把Y 变成一个列表，因为成本函数期望Y 以这种格式提供（每个时间步一个列表项）。 所以list（Y）是一个包含 30 个项目的列表，其中每个列表项目都是形状的（60,78）。 让训练100个epochs。 这将需要几分钟的时间。
model.fit([X, a0, c0], list(Y), epochs=100)
Epoch 1/100
60/60 [==============================] - 19s 316ms/step - loss: 125.9216 - dense_1_loss_1: 4.3546 - dense_1_loss_2: 4.3476 - dense_1_loss_3: 4.3439 - dense_1_loss_4: 4.3483 - dense_1_loss_5: 4.3439 - dense_1_loss_6: 4.3465 - dense_1_loss_7: 4.3428 - dense_1_loss_8: 4.3433 - dense_1_loss_9: 4.3390 - dense_1_loss_10: 4.3417 - dense_1_loss_11: 4.3352 - dense_1_loss_12: 4.3507 - dense_1_loss_13: 4.3469 - dense_1_loss_14: 4.3361 - dense_1_loss_15: 4.3412 - dense_1_loss_16: 4.3411 - dense_1_loss_17: 4.3493 - dense_1_loss_18: 4.3379 - dense_1_loss_19: 4.3318 - dense_1_loss_20: 4.3431 - dense_1_loss_21: 4.3416 - dense_1_loss_22: 4.3405 - dense_1_loss_23: 4.3349 - dense_1_loss_24: 4.3379 - dense_1_loss_25: 4.3429 - dense_1_loss_26: 4.3357 - dense_1_loss_27: 4.3455 - dense_1_loss_28: 4.3347 - dense_1_loss_29: 4.3432 - dense_1_loss_30: 0.0000e+00 - dense_1_acc_1: 0.0500 - dense_1_acc_2: 0.0333 - dense_1_acc_3: 0.0667 - dense_1_acc_4: 0.0333 - dense_1_acc_5: 0.0000e+00 - dense_1_acc_6: 0.0167 - dense_1_acc_7: 0.0667 - dense_1_acc_8: 0.0500 - dense_1_acc_9: 0.0667 - dense_1_acc_10: 0.0500 - dense_1_acc_11: 0.1000 - dense_1_acc_12: 0.0167 - dense_1_acc_13: 0.0167 - dense_1_acc_14: 0.0667 - dense_1_acc_15: 0.0000e+00 - dense_1_acc_16: 0.0500 - dense_1_acc_17: 0.0667 - dense_1_acc_18: 0.0333 - dense_1_acc_19: 0.1167 - dense_1_acc_20: 0.0000e+00 - dense_1_acc_21: 0.0500 - dense_1_acc_22: 0.0500 - dense_1_acc_23: 0.1167 - dense_1_acc_24: 0.1167 - dense_1_acc_25: 0.0333 - dense_1_acc_26: 0.0500 - dense_1_acc_27: 0.0500 - dense_1_acc_28: 0.0167 - dense_1_acc_29: 0.0667 - dense_1_acc_30: 0.0000e+00
Epoch 2/100
60/60 [==============================] - 0s 2ms/step - loss: 122.1951 - dense_1_loss_1: 4.3314 - dense_1_loss_2: 4.3034 - dense_1_loss_3: 4.2770 - dense_1_loss_4: 4.2792 - dense_1_loss_5: 4.2472 - dense_1_loss_6: 4.2584 - dense_1_loss_7: 4.2406 - dense_1_loss_8: 4.2140 - dense_1_loss_9: 4.2278 -
60/60 [==============================] - 0s 1ms/step - loss: 113.1482 - dense_1_loss_1: 4.2857 - dense_1_loss_2: 4.2005 - dense_1_loss_3: 4.0911 - dense_1_loss_4: 4.0643 - dense_1_loss_5: 3.9361 - dense_1_loss_6: 3.9924 - dense_1_loss_7: 3.8936 - dense_1_loss_8: 3.6965 - dense_1_loss_9: 3.8225 - dense_1_loss_10: 3.6392 - dense_1_loss_11: 3.7633 - dense_1_loss_12: 4.1325 - dense_1_acc_16: 0.0667 - dense_1_acc_17: 0.1833 - dense_1_acc_18: 0.0500 - dense_1_acc_19: 0.0667 - dense_1_acc_20: 0.1167 - dense_1_acc_21: 0.0833 - dense_1_acc_22: 0.0167 - dense_1_acc_23: 0.0333 - dense_1_acc_24: 0.0667 - dense_1_acc_25: 0.0500 - dense_1_acc_26: 0.1333 - dense_1_acc_27: 0.0833 - dense_1_acc_28: 0.0500 - dense_1_acc_29: 0.1000 - dense_1_acc_30: 0.0000e+00
Epoch 5/100
.
.
.
.
60/60 [==============================] - 0s 1ms/step - loss: 6.2689 - dense_1_loss_1: 3.7845 - dense_1_loss_2: 1.2346 - dense_1_loss_3: 0.3671 - dense_1_loss_4: 0.0967 - dense_1_loss_5: 0.0647 - dense_1_loss_6: 0.0501 - dense_1_loss_7: 0.0355 - dense_1_loss_8: 0.0351 - dense_1_loss_9: 0.0329 - dense_1_loss_10: 0.0281 - dense_1_loss_11: 0.0288 - dense_1_loss_12: 0.0283 - dense_1_loss_13: 0.0255 - dense_1_loss_14: 0.0270 - dense_1_loss_15: 0.0291 - dense_1_loss_16: 0.0289 - dense_1_loss_17: 0.0265 - dense_1_loss_18: 0.0271 - dense_1_loss_19: 0.0261 - dense_1_loss_20: 0.0293 - dense_1_loss_21: 0.0304 - 56 - dense_1_loss_1: 3.7809 - dense_1_loss_2: 1.2235 - dense_1_loss_3: 0.3603 - dense_1_loss_4: 0.0947 - dense_1_loss_5: 0.0633 - dense_1_loss_6: 0.0489 - dense_1_loss_7: 0.0347 - dense_1_loss_8: 0.0342 - dense_1_loss_9: 0.0321 - dense_1_loss_10: 0.0274 - dense_1_loss_11: 0.0279 - dense_1_loss_12: 0.0276 - dense_1_loss_13: 0.0249 - dense_1_loss_14: 0.0262 - dense_1_loss_15: 0.0282 - dense_1_loss_16: 0.0282 - dense_1_loss_17: 0.0259 - dense_1_loss_18: 0.0264 - dense_1_loss_19: 0.0254 - dense_1_loss_20: 0.0285 - dense_1_loss_21: 0.0296 -
60/60 [==============================] - 0s 1ms/step - loss: 6.1823 - dense_1_loss_1: 3.7776 - dense_1_loss_2: 1.2114 - dense_1_loss_3: 0.3527 - dense_1_loss_4: 0.0929 - dense_1_loss_5: 0.0619 - dense_1_loss_6: 0.0477 - dense_1_loss_7: 0.0338 - dense_1_loss_8: 0.0334 - dense_1_loss_9: 0.0314 - dense_1_loss_10: 0.0268 - dense_1_loss_11: 0.0272 - dense_1_loss_12: 0.0271 - dense_1_loss_13: 0.0244 - dense_1_loss_14: 0.0256 - dense_1_loss_15: 0.0275 - dense_1_loss_16: 0.0276 - dense_1_loss_17: 0.0253 - dense_1_loss_18: 0.0257 - dense_1_loss_19: 0.0247 - dense_1_loss_20: 0.0279 - dense_1_loss_21: 0.0289 - dense_1_loss_22: 0.0262 - dense_1_loss_23: 0.0255 - dense_1_loss_24: 0.0244 - dense_1_loss_25: 0.0262 - dense_1_loss_26: 0.0248 - dense_1_loss_27: 0.0280 - dense_1_loss_28: 0.0313 - dense_1_loss_29: 0.0346 - dense_1_loss_30: 0.0000e+00 - dense_1_acc_1: 0.1000 - dense_1_acc_2: 0.6667 - dense_1_acc_3: 0.9333 - dense_1_acc_4: 1.0000 - dense_1_acc_5: 1.0000 - dense_1_acc_6: 1.0000 - dense_1_acc_7: 1.0000 - dense_1_acc_8: 1.0000 - dense_1_acc_9: 1.0000 - dense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
Epoch 89/100
60/60 [==============================] - 0s 1ms/step - loss: 6.1403 - dense_1_loss_1: 3.7743 - dense_1_loss_2: 1.1999 - dense_1_loss_3: 0.3456 - dense_1_loss_4: 0.0906 - dense_1_loss_5: 0.0605 - dense_1_loss_6: 0.0464 - dense_1_loss_7: 0.0331 - dense_1_loss_8: 0.0326 - dense_1_loss_9: 0.0305 - dense_1_loss_10: 0.0262 - dense_1_loss_11: 0.0265 - dense_1_loss_12: 0.0264 - dense_1_loss_13: 0.0238 - dense_1_loss_14: 0.0250 - dense_1_loss_15: 0.0269 - dense_1_loss_16: 0.0269 - dense_1_loss_17: 0.0248 - dense_1_loss_18: 0.0250 - dense_1_loss_19: 0.0241 - dense_1_loss_20: 0.0272 - dense_1_loss_21: 0.0282 - dense_1_loss_22: 0.0257 - dense_1_loss_23: 0.0249 - dense_1_loss_24: 0.0238 - dense_1_loss_25: 0.0257 - dense_1_loss_26: 0.0242 - dense_1_loss_27: 0.0272 - dense_1_loss_28: 0.0308 - dense_1_loss_29: 0.0335 - dense_1_loss_30: 0.0000e+00 - dense_1_acc_1: 0.1000 - dense_1_acc_2: 0.6667 - dense_1_acc_3: 0.9333 - dense_1_acc_4: 1.0000 - dense_1_acc_5: 1.0000 - dense_1_acc_6: 1.0000 - dense_1_acc_7: 1.0000 - dense_1_acc_8: 1.0000 - dense_1_acc_9: 1.0000 - dense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
Epoch 90/100
60/60 [==============================] - 0s 1ms/step - loss: 6.1011 - dense_1_loss_1: 3.7711 - dense_1_loss_2: 1.1893 - dense_1_loss_3: 0.3391 - dense_1_loss_4: 0.0885 - dense_1_loss_5: 0.0592 - dense_1_loss_6: 0.0454 - dense_1_loss_7: 0.0324 - dense_1_loss_8: 0.0319 - dense_1_loss_9: 0.0298 - dense_1_loss_10: 0.0256 - dense_1_loss_11: 0.0260 - dense_1_loss_12: 0.0258 - dense_1_loss_13: 0.0231 - dense_1_loss_14: 0.0244 - dense_1_loss_15: 0.0265 - dense_1_loss_16: 0.0262 - dense_1_loss_17: 0.0240 - dense_1_loss_18: 0.0245 - dense_1_loss_19: 0.0236 - dense_1_loss_20: 0.0265 - dense_1_loss_21: 0.0275 - dense_1_loss_22: 0.0251 - dense_1_loss_23: 0.0244 - dense_1_loss_24: 0.0232 - dense_1_loss_25: 0.0252 - dense_1_loss_26: 0.0239 - dense_1_loss_27: 0.0260 - dense_1_loss_28: 0.0304 - dense_1_loss_29: 0.0323 - dense_1_loss_30: 0.0000e+00 - dense_1_acc_1: 0.1000 - dense_1_acc_2: 0.6667 - dense_1_acc_3: 0.9333 - dense_1_acc_4: 1.0000 - dense_1_acc_5: 1.0000 - dense_1_acc_6: 1.0000 - dense_1_acc_7: 1.0000 - dense_1_acc_8: 1.0000 - dense_1_acc_9: 1.0000 - dense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
dense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
Epoch 92/100
60/60 [==============================] - 0s 1ms/step - loss: 6.0243 - dense_1_loss_1: 3.7647 - dense_1_loss_2: 1.1673 - dense_1_loss_3: 0.3269 - dense_1_loss_4: 0.0846 - dense_1_loss_5: 0.0567 - dense_1_loss_6: 0.0433 - dense_1_loss_7: 0.0310 - dense_1_loss_8: 0.0305 - dense_1_loss_9: 0.0285 - dense_1_loss_10: 0.0245 - dense_1_loss_11: 0.0249 - dense_1_loss_12: 0.0246 - dense_1_loss_13: 0.0221 - dense_1_loss_14: 0.0233 - dense_1_loss_15: 0.0253 - dense_1_loss_16: 0.0250 - dense_1_loss_17: 0.0229 - dense_1_loss_18: 0.0235 - dense_1_loss_19: 0.0226 - dense_1_loss_20: 0.0253 - dense_1_loss_21: 0.0263 - dense_1_loss_22: 0.0240 - dense_1_loss_23: 0.0233 - dense_1_loss_24: 0.0223 - dense_1_loss_25: 0.0240 - dense_1_loss_26: 0.0227 - dense_1_loss_27: 0.0250 - dense_1_loss_28: 0.0288 - dense_1_loss_29: 0.0305 - dense_1_loss_30: 0.0000e+00 - dense_1_acc_1: 0.1000 - dense_1_acc_2: 0.6667 - dense_1_acc_3: 0.9333 - dense_1_acc_4: 1.0000 - dense_1_acc_5: 1.0000 - dense_1_acc_6: 1.0000 - dense_1_acc_7: 1.0000 - dense_1_acc_8: 1.0000 - dense_1_acc_9: 1.0000 - dense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
Epoch 93/100
60/60 [==============================] - 0s 1ms/step - loss: 5.9890 - dense_1_loss_1: 3.7615 - dense_1_loss_2: 1.1573 - dense_1_loss_3: 0.3211 - dense_1_loss_4: 0.0832 - dense_1_loss_5: 0.0555 - dense_1_loss_6: 0.0424 - dense_1_loss_7: 0.0302 - dense_1_loss_8: 0.0298 - dense_1_loss_9: 0.0279 - dense_1_loss_10: 0.0239 - dense_1_loss_11: 0.0243 - dense_1_loss_12: 0.0241 - dense_1_loss_13: 0.0216 - dense_1_loss_14: 0.0227 - dense_1_loss_15: 0.0246 - dense_1_loss_16: 0.0245 - dense_1_loss_17: 0.0225 - dense_1_loss_18: 0.0229 - dense_1_loss_19: 0.0220 - dense_1_loss_20: 0.0248 - dense_1_loss_21: 0.0257 - dense_1_loss_22: 0.0234 - dense_1_loss_23: 0.0228 - dense_1_loss_24: 0.0219 - dense_1_loss_25: 0.0234 - dense_1_loss_26: 0.0220 - dense_1_loss_27: 0.0249 - dense_1_loss_28: 0.0279 - dense_1_loss_29: 0.0301 - dense_1_loss_30: 0.0000e+00 - dense_1_acc_1: 0.1000 - dense_1_acc_2: 0.6833 - dense_1_acc_3: 0.9333 - dense_1_acc_4: 1.0000 - dense_1_acc_5: 1.0000 - dense_1_acc_6: 1.0000 - - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
Epoch 95/100
60/60 [==============================] - 0s 1ms/step - loss: 5.9199 - dense_1_loss_1: 3.7554 - dense_1_loss_2: 1.1372 - dense_1_loss_3: 0.3089 - dense_1_loss_4: 0.0802 - dense_1_loss_5: 0.0533 - dense_1_loss_6: 0.0405 - dense_1_loss_7: 0.0291 - dense_1_loss_8: 0.0286 - dense_1_loss_9: 0.0267 - dense_1_loss_10: 0.0230 - dense_1_loss_11: 0.0232 - dense_1_loss_12: 0.0231 - dense_1_loss_13: 0.0208 - dense_1_loss_14: 0.0218 - dense_1_loss_15: 0.0236 - dense_1_loss_16: 0.0235 - dense_1_loss_17: 0.0215 - dense_1_loss_18: 0.0220 - dense_1_loss_19: 0.0210 - dense_1_loss_20: 0.0237 - dense_1_loss_21: 0.0246 - dense_1_loss_22: 0.0224 - dense_1_loss_23: 0.0218 - dense_1_loss_24: 0.0209 - dense_1_loss_25: 0.0224 - dense_1_loss_26: 0.0213 - dense_1_loss_27: 0.0235 - ddense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_loss_4: 0.0757 - dense_1_loss_5: 0.0501 - dense_1_loss_6: 0.0383 - dense_1_loss_7: 0.0274 - dense_1_loss_8: 0.0268 - dense_1_loss_9: 0.0252 - dense_1_loss_10: 0.0216 - dense_1_loss_11: 0.0219 - dense_1_loss_12: 0.0217 - dense_1_loss_13: 0.0194 - dense_1_loss_14: 0.0205 - dense_1_loss_15: 0.0223 - dense_1_loss_16: 0.0220 - dense_1_loss_17: 0.0202 - dense_1_loss_18: 0.0207 - dense_1_loss_19: 0.0198 - dense_1_loss_20: 0.0223 - dense_1_loss_21: 0.0232 - dense_1_loss_22: 0.0211 - dense_1_loss_23: 0.0205 - dense_1_loss_24: 0.0197 - dense_1_loss_25: 0.0211 - dense_1_loss_26: 0.0201 - dense_1_loss_27: 0.0218 - dense_1_acc_7: 1.0000 - dense_1_acc_8: 1.0000 - dense_1_acc_9: 1.0000 - dense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
Epoch 100/100
60/60 [==============================] - 0s 1ms/step - loss: 5.7645 - dense_1_loss_1: 3.7402 - dense_1_loss_2: 1.0904 - dense_1_loss_3: 0.2835 - dense_1_loss_4: 0.0733 - dense_1_loss_5: 0.0484 - dense_1_loss_6: 0.0368 - dense_1_loss_7: 0.0264 - dense_1_loss_8: 0.0258 - dense_1_loss_9: 0.0243 - dense_1_loss_10: 0.0209 - dense_1_loss_11: 0.0209 - dense_1_loss_12: 0.0209 - dense_1_loss_13: 0.0187 - dense_1_loss_14: 0.0197 - dense_1_loss_15: 0.0214 - dense_1_loss_16: 0.0212 - dense_1_loss_17: 0.0195 - dense_1_loss_18: 0.0199 - dense_1_loss_19: 0.0190 - dense_1_loss_20: 0.0215 - dense_1_loss_21: 0.0223 - dense_1_loss_22: 0.0203 - dense_1_loss_23: 0.0197 - dense_1_loss_24: 0.0190 - dense_1_loss_25: 0.0203 - dense_1_loss_26: 0.0192 - dense_1_loss_27: 0.0212 - dense_1_loss_28: 0.0239 - dense_1_loss_29: 0.0259 - dense_1_loss_30: 0.0000e+00 - dense_1_acc_1: 0.1000 - dense_1_acc_2: 0.6833 - dense_1_acc_3: 0.9333 - dense_1_acc_4: 1.0000 - dense_1_acc_5: 1.0000 - dense_1_acc_6: 1.0000 - dense_1_acc_7: 1.0000 - dense_1_acc_8: 1.0000 - dense_1_acc_9: 1.0000 - dense_1_acc_10: 1.0000 - dense_1_acc_11: 1.0000 - dense_1_acc_12: 1.0000 - dense_1_acc_13: 1.0000 - dense_1_acc_14: 1.0000 - dense_1_acc_15: 1.0000 - dense_1_acc_16: 1.0000 - dense_1_acc_17: 1.0000 - dense_1_acc_18: 1.0000 - dense_1_acc_19: 1.0000 - dense_1_acc_20: 1.0000 - dense_1_acc_21: 1.0000 - dense_1_acc_22: 1.0000 - dense_1_acc_23: 1.0000 - dense_1_acc_24: 1.0000 - dense_1_acc_25: 1.0000 - dense_1_acc_26: 1.0000 - dense_1_acc_27: 1.0000 - dense_1_acc_28: 1.0000 - dense_1_acc_29: 1.0000 - dense_1_acc_30: 0.0333
<keras.callbacks.History at 0x235dc340a20>
You should see the model loss going down. Now that you have trained a model, lets go on the the final section to implement an inference algorithm, and generate some music!
3 - Generating music 音乐生成
You now have a trained model which has learned the patterns of the jazz soloist. Lets now use this model to synthesize new music.
3.1 - Predicting & Sampling
At each step of sampling, you will take as input the activation a and cell state c from the previous state of the LSTM, forward propagate by one step, and get a new output activation as well as cell state. The new activation a can then be used to generate the output, using densor as before.
To start off the model, we will initialize x0 as well as the LSTM activation and and cell value a0 and c0 to be zeros.
Exercise: Implement the function below to sample a sequence of musical values. Here are some of the key steps you’ll need to implement inside the for-loop that generates the Ty T y output characters:
Step 2.A: Use LSTM_Cell, which inputs the previous step’s c and a to generate the current step’s c and a.
Step 2.B: Use densor (defined previously) to compute a softmax on a to get the output for the current step.
Step 2.C: Save the output you have just generated by appending it to outputs.
Step 2.D: Sample x to the be “out“‘s one-hot version (the prediction) so that you can pass it to the next LSTM’s step. We have already provided this line of code, which uses a Lambda function.
x = Lambda(one_hot)(out) [Minor technical note: Rather than sampling a value at random according to the probabilities in out, this line of code actually chooses the single most likely note at each step using an argmax.]
# GRADED FUNCTION: music_inference_modeldef music_inference_model(LSTM_cell, densor, n_values = 78, n_a = 64, Ty = 100):"""Uses the trained "LSTM_cell" and "densor" from model() to generate a sequence of values.Arguments:LSTM_cell -- the trained "LSTM_cell" from model(), Keras layer objectdensor -- the trained "densor" from model(), Keras layer objectn_values -- integer, umber of unique valuesn_a -- number of units in the LSTM_cellTy -- integer, number of time steps to generateReturns:inference_model -- Keras model instance"""# Define the input of your model with a shape x0 = Input(shape=(1, n_values))# Define s0, initial hidden state for the decoder LSTMa0 = Input(shape=(n_a,), name='a0')c0 = Input(shape=(n_a,), name='c0')a = a0c = c0x = x0### START CODE HERE #### Step 1: Create an empty list of "outputs" to later store your predicted values (≈1 line)outputs = []# Step 2: Loop over Ty and generate a value at every time stepfor t in range(Ty):# Step 2.A: Perform one step of LSTM_cell (≈1 line)a, _, c = LSTM_cell(x, initial_state=[a, c])# Step 2.B: Apply Dense layer to the hidden state output of the LSTM_cell (≈1 line)out = densor(a)# Step 2.C: Append the prediction "out" to "outputs". out.shape = (None, 78) (≈1 line)outputs.append(out)# Step 2.D: Select the next value according to "out", and set "x" to be the one-hot representation of the# selected value, which will be passed as the input to LSTM_cell on the next step. We have provided # the line of code you need to do this. x = Lambda(one_hot)(out)# Step 3: Create model instance with the correct "inputs" and "outputs" (≈1 line)inference_model = Model(inputs=[x0, a0, c0], outputs=outputs)### END CODE HERE ###return inference_modelRun the cell below to define your inference model. This model is hard coded to generate 50 values. 硬编码生成 50 个值
inference_model = music_inference_model(LSTM_cell, densor, n_values = 78, n_a = 64, Ty = 50)Finally, this creates the zero-valued vectors you will use to initialize x and the LSTM state variables a and c.
x_initializer = np.zeros((1, 1, 78))
a_initializer = np.zeros((1, n_a))
c_initializer = np.zeros((1, n_a))Exercise: Implement predict_and_sample(). This function takes many arguments including the inputs [x_initializer, a_initializer, c_initializer]. In order to predict the output corresponding to this input, you will need to carry-out 3 steps:
1. Use your inference model to predict an output given your set of inputs. The output pred should be a list of length 20 where each element is a numpy-array of shape ( Ty T y , n_values)
2. Convert pred into a numpy array of Ty T y indices. Each index corresponds is computed by taking the argmax of an element of the pred list. Hint.
3. Convert the indices into their one-hot vector representations. Hint.
实现predict_and_sample（）。 这个函数有很多参数，包括输入[x_initializer，a_initializer，c_initializer]。 为了预测与此输入相对应的输出，您需要执行3个步骤：
- 使用您的推理模型来预测给定输入集的输出。 输出
pred应该是一个长度为20的列表，其中每个元素是形状的数组（ Ty T y ，n_values） - 将
pred转换为 Ty T y 索引的numpy数组。 每个索引对应的是通过采用pred列表元素的argmax来计算的。 Hint. - 将这些索引转换为它们的一个热点向量表示。 Hint.
实现 预测采样 函数 1. 使用模型，预测出结果，并作为输出，2. 输出结果 pred 转化为 数组，3.将 indicies 转化为 one -hot 向量
# GRADED FUNCTION: predict_and_sampledef predict_and_sample(inference_model, x_initializer = x_initializer, a_initializer = a_initializer, c_initializer = c_initializer):"""Predicts the next value of values using the inference model.Arguments:inference_model -- Keras model instance for inference timex_initializer -- numpy array of shape (1, 1, 78), one-hot vector initializing the values generationa_initializer -- numpy array of shape (1, n_a), initializing the hidden state of the LSTM_cellc_initializer -- numpy array of shape (1, n_a), initializing the cell state of the LSTM_celReturns:results -- numpy-array of shape (Ty, 78), matrix of one-hot vectors representing the values generatedindices -- numpy-array of shape (Ty, 1), matrix of indices representing the values generated"""### START CODE HERE #### Step 1: Use your inference model to predict an output sequence given x_initializer, a_initializer and c_initializer.pred = inference_model.predict([x_initializer, a_initializer, c_initializer])# Step 2: Convert "pred" into an np.array() of indices with the maximum probabilities ---argmax返回的是最大数的索引indices = np.argmax(np.array(pred), axis=-1)# Step 3: Convert indices to one-hot vectors, the shape of the results should be (1, )results = to_categorical(indices, num_classes=x_initializer.shape[-1])### END CODE HERE ###return results, indicesresults, indices = predict_and_sample(inference_model, x_initializer, a_initializer, c_initializer)
print("np.argmax(results[12]) =", np.argmax(results[12]))
print("np.argmax(results[17]) =", np.argmax(results[17]))
print("list(indices[12:18]) =", list(indices[12:18]))np.argmax(results[12]) = 29
np.argmax(results[17]) = 31
list(indices[12:18]) = [array([29], dtype=int64), array([31], dtype=int64), array([55], dtype=int64), array([8], dtype=int64), array([29], dtype=int64), array([31], dtype=int64)]
Expected Output: Your results may differ because Keras’ results are not completely predictable. However, if you have trained your LSTM_cell with model.fit() for exactly 100 epochs as described above, you should very likely observe a sequence of indices that are not all identical. Moreover, you should observe that: np.argmax(results[12]) is the first element of list(indices[12:18]) and np.argmax(results[17]) is the last element of list(indices[12:18]).
| **np.argmax(results[12])** = | 1 |
| **np.argmax(results[12])** = | 42 |
| **list(indices[12:18])** = | [array([1]), array([42]), array([54]), array([17]), array([1]), array([42])] |
3.3 - Generate music
Finally, you are ready to generate music. Your RNN generates a sequence of values. The following code generates music by first calling your predict_and_sample() function. These values are then post-processed into musical chords (meaning that multiple values or notes can be played at the same time).
Most computational music algorithms use some post-processing because it is difficult to generate music that sounds good without such post-processing. The post-processing does things such as clean up the generated audio by making sure the same sound is not repeated too many times, that two successive notes are not too far from each other in pitch, and so on. One could argue that a lot of these post-processing steps are hacks; also, a lot the music generation literature has also focused on hand-crafting post-processors, and a lot of the output quality depends on the quality of the post-processing and not just the quality of the RNN. But this post-processing does make a huge difference, so lets use it in our implementation as well.
最后，您已准备好制作音乐。您的 RNN 生成一系列值。下面的代码首先调用你的predict_and_sample（）函数来生成音乐。这些值然后被后处理成音乐和弦（意味着可以同时播放多个值或音符）。
大多数计算音乐算法使用一些后期处理，因为如果没有这样的后期处理，很难生成听起来很好的音乐。后期处理通过确保相同的声音不重复多次，两个连续的音符在音高中彼此相距不太远等来清理所生成的音频。有人可能会说，很多这些后处理步骤都是黑客行为;另外，很多音乐创作文献也专注于手工制作后期处理器，很多输出质量取决于后期处理的质量，而不仅仅取决于 RNN 的质量。但是这个后期处理确实有很大的不同，所以我们也可以在我们的实现中使用它。
Lets make some music!
Run the following cell to generate music and record it into your out_stream. This can take a couple of minutes.
out_stream = generate_music(inference_model)Predicting new values for different set of chords.
Generated 51 sounds using the predicted values for the set of chords ("1") and after pruning
Generated 51 sounds using the predicted values for the set of chords ("2") and after pruning
Generated 51 sounds using the predicted values for the set of chords ("3") and after pruning
Generated 50 sounds using the predicted values for the set of chords ("4") and after pruning
Generated 50 sounds using the predicted values for the set of chords ("5") and after pruning
Your generated music is saved in output/my_music.midi
To listen to your music, click File->Open… Then go to “output/” and download “my_music.midi”. Either play it on your computer with an application that can read midi files if you have one, or use one of the free online “MIDI to mp3” conversion tools to convert this to mp3.
As reference, here also is a 30sec audio clip we generated using this algorithm.
IPython.display.Audio('./data/30s_trained_model.mp3')音乐片段，无法上传。
Congratulations!
You have come to the end of the notebook.
Here’s what you should remember:
- A sequence model can be used to generate musical values, which are then post-processed into midi music. 序列模型可用于生成音乐值，然后将其后处理为MIDI音乐。
- Fairly similar models can be used to generate dinosaur names or to generate music, with the major difference being the input fed to the model. 相当相似的模型可用于生成恐龙名称或生成音乐，主要区别在于输入给模型的输入
- In Keras, sequence generation involves defining layers with shared weights, which are then repeated for the different time steps 1,…,Tx 1 , … , T x . 在Keras中，序列生成涉及定义具有共享权重的图层，然后针对不同的时间步骤重复这些图层
Congratulations on completing this assignment and generating a jazz solo!
作业中用到的 utils 中的代码。
'''
data_utils.py
'''from music_utils import *
from preprocess import *
from keras.utils import to_categoricalchords, abstract_grammars = get_musical_data('data/original_metheny.mid')
corpus, tones, tones_indices, indices_tones = get_corpus_data(abstract_grammars)
N_tones = len(set(corpus))
n_a = 64
x_initializer = np.zeros((1, 1, 78))
a_initializer = np.zeros((1, n_a))
c_initializer = np.zeros((1, n_a))def load_music_utils():chords, abstract_grammars = get_musical_data('data/original_metheny.mid')corpus, tones, tones_indices, indices_tones = get_corpus_data(abstract_grammars)N_tones = len(set(corpus))X, Y, N_tones = data_processing(corpus, tones_indices, 60, 30) return (X, Y, N_tones, indices_tones)def generate_music(inference_model, corpus = corpus, abstract_grammars = abstract_grammars, tones = tones, tones_indices = tones_indices, indices_tones = indices_tones, T_y = 10, max_tries = 1000, diversity = 0.5):"""Generates music using a model trained to learn musical patterns of a jazz soloist. Creates an audio streamto save the music and play it.Arguments:model -- Keras model Instance, output of djmodel()corpus -- musical corpus, list of 193 tones as strings (ex: 'C,0.333,<P1,d-5>')abstract_grammars -- list of grammars, on element can be: 'S,0.250,<m2,P-4> C,0.250,<P4,m-2> A,0.250,<P4,m-2>'tones -- set of unique tones, ex: 'A,0.250,<M2,d-4>' is one element of the set.tones_indices -- a python dictionary mapping unique tone (ex: A,0.250,< m2,P-4 >) into their corresponding indices (0-77)indices_tones -- a python dictionary mapping indices (0-77) into their corresponding unique tone (ex: A,0.250,< m2,P-4 >)Tx -- integer, number of time-steps used at training timetemperature -- scalar value, defines how conservative/creative the model is when generating musicReturns:predicted_tones -- python list containing predicted tones"""# set up audio streamout_stream = stream.Stream()# Initialize chord variablescurr_offset = 0.0 # variable used to write sounds to the Stream.num_chords = int(len(chords) / 3) # number of different set of chordsprint("Predicting new values for different set of chords.")# Loop over all 18 set of chords. At each iteration generate a sequence of tones# and use the current chords to convert it into actual sounds for i in range(1, num_chords):# Retrieve current chord from streamcurr_chords = stream.Voice()# Loop over the chords of the current set of chordsfor j in chords[i]:# Add chord to the current chords with the adequate offset, no need to understand thiscurr_chords.insert((j.offset % 4), j)# Generate a sequence of tones using the model_, indices = predict_and_sample(inference_model)indices = list(indices.squeeze())pred = [indices_tones[p] for p in indices]predicted_tones = 'C,0.25 'for k in range(len(pred) - 1):predicted_tones += pred[k] + ' ' predicted_tones += pred[-1]#### POST PROCESSING OF THE PREDICTED TONES ##### We will consider "A" and "X" as "C" tones. It is a common choice.predicted_tones = predicted_tones.replace(' A',' C').replace(' X',' C')# Pruning #1: smoothing measurepredicted_tones = prune_grammar(predicted_tones)# Use predicted tones and current chords to generate soundssounds = unparse_grammar(predicted_tones, curr_chords)# Pruning #2: removing repeated and too close together soundssounds = prune_notes(sounds)# Quality assurance: clean up soundssounds = clean_up_notes(sounds)# Print number of tones/notes in soundsprint('Generated %s sounds using the predicted values for the set of chords ("%s") and after pruning' % (len([k for k in sounds if isinstance(k, note.Note)]), i))# Insert sounds into the output streamfor m in sounds:out_stream.insert(curr_offset + m.offset, m)for mc in curr_chords:out_stream.insert(curr_offset + mc.offset, mc)curr_offset += 4.0# Initialize tempo of the output stream with 130 bit per minuteout_stream.insert(0.0, tempo.MetronomeMark(number=130))# Save audio stream to finemf = midi.translate.streamToMidiFile(out_stream)mf.open("output/my_music.midi", 'wb')mf.write()print("Your generated music is saved in output/my_music.midi")mf.close()# Play the final stream through output (see 'play' lambda function above)# play = lambda x: midi.realtime.StreamPlayer(x).play()# play(out_stream)return out_streamdef predict_and_sample(inference_model, x_initializer = x_initializer, a_initializer = a_initializer, c_initializer = c_initializer):"""Predicts the next value of values using the inference model.Arguments:inference_model -- Keras model instance for inference timex_initializer -- numpy array of shape (1, 1, 78), one-hot vector initializing the values generationa_initializer -- numpy array of shape (1, n_a), initializing the hidden state of the LSTM_cellc_initializer -- numpy array of shape (1, n_a), initializing the cell state of the LSTM_celTy -- length of the sequence you'd like to generate.Returns:results -- numpy-array of shape (Ty, 78), matrix of one-hot vectors representing the values generatedindices -- numpy-array of shape (Ty, 1), matrix of indices representing the values generated"""### START CODE HERE ###pred = inference_model.predict([x_initializer, a_initializer, c_initializer])indices = np.argmax(pred, axis = -1)results = to_categorical(indices, num_classes=78)### END CODE HERE ###return results, indices'''
grammar.py''''''
Author: Ji-Sung Kim, Evan Chow
Project: jazzml / (used in) deepjazz
Purpose: Extract, manipulate, process musical grammarDirectly taken then cleaned up from Evan Chow's jazzml,
https://github.com/evancchow/jazzml,with permission.
'''from collections import OrderedDict, defaultdict
from itertools import groupby
from music21 import *
import copy, random, pdb#from preprocess import *''' Helper function to determine if a note is a scale tone. '''
def __is_scale_tone(chord, note):# Method: generate all scales that have the chord notes th check if note is# in names# Derive major or minor scales (minor if 'other') based on the quality# of the chord.scaleType = scale.DorianScale() # i.e. minor pentatonicif chord.quality == 'major':scaleType = scale.MajorScale()# Can change later to deriveAll() for flexibility. If so then use list# comprehension of form [x for a in b for x in a].scales = scaleType.derive(chord) # use deriveAll() later for flexibilityallPitches = list(set([pitch for pitch in scales.getPitches()]))allNoteNames = [i.name for i in allPitches] # octaves don't matter# Get note name. Return true if in the list of note names.noteName = note.namereturn (noteName in allNoteNames)''' Helper function to determine if a note is an approach tone. '''
def __is_approach_tone(chord, note):# Method: see if note is +/- 1 a chord tone.for chordPitch in chord.pitches:stepUp = chordPitch.transpose(1)stepDown = chordPitch.transpose(-1)if (note.name == stepDown.name or note.name == stepDown.getEnharmonic().name ornote.name == stepUp.name ornote.name == stepUp.getEnharmonic().name):return Truereturn False''' Helper function to determine if a note is a chord tone. '''
def __is_chord_tone(lastChord, note):return (note.name in (p.name for p in lastChord.pitches))''' Helper function to generate a chord tone. '''
def __generate_chord_tone(lastChord):lastChordNoteNames = [p.nameWithOctave for p in lastChord.pitches]return note.Note(random.choice(lastChordNoteNames))''' Helper function to generate a scale tone. '''
def __generate_scale_tone(lastChord):# Derive major or minor scales (minor if 'other') based on the quality# of the lastChord.scaleType = scale.WeightedHexatonicBlues() # minor pentatonicif lastChord.quality == 'major':scaleType = scale.MajorScale()# Can change later to deriveAll() for flexibility. If so then use list# comprehension of form [x for a in b for x in a].scales = scaleType.derive(lastChord) # use deriveAll() later for flexibilityallPitches = list(set([pitch for pitch in scales.getPitches()]))allNoteNames = [i.name for i in allPitches] # octaves don't matter# Return a note (no octave here) in a scale that matches the lastChord.sNoteName = random.choice(allNoteNames)lastChordSort = lastChord.sortAscending()sNoteOctave = random.choice([i.octave for i in lastChordSort.pitches])sNote = note.Note(("%s%s" % (sNoteName, sNoteOctave)))return sNote''' Helper function to generate an approach tone. '''
def __generate_approach_tone(lastChord):sNote = __generate_scale_tone(lastChord)aNote = sNote.transpose(random.choice([1, -1]))return aNote''' Helper function to generate a random tone. '''
def __generate_arbitrary_tone(lastChord):return __generate_scale_tone(lastChord) # fix later, make random note.''' Given the notes in a measure ('measure') and the chords in that measure('chords'), generate a list of abstract grammatical symbols to represent that measure as described in GTK's "Learning Jazz Grammars" (2009). Inputs: 1) "measure" : a stream.Voice object where each element is anote.Note or note.Rest object.>>> m1<music21.stream.Voice 328482572>>>> m1[0]<music21.note.Rest rest>>>> m1[1]<music21.note.Note C>Can have instruments and other elements, removes them here.2) "chords" : a stream.Voice object where each element is a chord.Chord.>>> c1<music21.stream.Voice 328497548>>>> c1[0]<music21.chord.Chord E-4 G4 C4 B-3 G#2>>>> c1[1]<music21.chord.Chord B-3 F4 D4 A3>Can have instruments and other elements, removes them here. Outputs:1) "fullGrammar" : a string that holds the abstract grammar for measure.Format: (Remember, these are DURATIONS not offsets!)"R,0.125" : a rest element of (1/32) length, or 1/8 quarter note. "C,0.125<M-2,m-6>" : chord note of (1/32) length, generatedanywhere from minor 6th down to major 2nd down.(interval <a,b> is not ordered). '''def parse_melody(fullMeasureNotes, fullMeasureChords):# Remove extraneous elements.xmeasure = copy.deepcopy(fullMeasureNotes)chords = copy.deepcopy(fullMeasureChords)measure.removeByNotOfClass([note.Note, note.Rest])chords.removeByNotOfClass([chord.Chord])# Information for the start of the measure.# 1) measureStartTime: the offset for measure's start, e.g. 476.0.# 2) measureStartOffset: how long from the measure start to the first element.measureStartTime = measure[0].offset - (measure[0].offset % 4)measureStartOffset = measure[0].offset - measureStartTime# Iterate over the notes and rests in measure, finding the grammar for each# note in the measure and adding an abstract grammatical string for it. fullGrammar = ""prevNote = None # Store previous note. Need for interval.numNonRests = 0 # Number of non-rest elements. Need for updating prevNote.for ix, nr in enumerate(measure):# Get the last chord. If no last chord, then (assuming chords is of length# >0) shift first chord in chords to the beginning of the measure.try: lastChord = [n for n in chords if n.offset <= nr.offset][-1]except IndexError:chords[0].offset = measureStartTimelastChord = [n for n in chords if n.offset <= nr.offset][-1]# FIRST, get type of note, e.g. R for Rest, C for Chord, etc.# Dealing with solo notes here. If unexpected chord: still call 'C'.elementType = ' '# R: First, check if it's a rest. Clearly a rest --> only one possibility.if isinstance(nr, note.Rest):elementType = 'R'# C: Next, check to see if note pitch is in the last chord.elif nr.name in lastChord.pitchNames or isinstance(nr, chord.Chord):elementType = 'C'# L: (Complement tone) Skip this for now.# S: Check if it's a scale tone.elif __is_scale_tone(lastChord, nr):elementType = 'S'# A: Check if it's an approach tone, i.e. +-1 halfstep chord tone.elif __is_approach_tone(lastChord, nr):elementType = 'A'# X: Otherwise, it's an arbitrary tone. Generate random note.else:elementType = 'X'# SECOND, get the length for each element. e.g. 8th note = R8, but# to simplify things you'll use the direct num, e.g. R,0.125if (ix == (len(measure)-1)):# formula for a in "a - b": start of measure (e.g. 476) + 4diff = measureStartTime + 4.0 - nr.offsetelse:diff = measure[ix + 1].offset - nr.offset# Combine into the note info.noteInfo = "%s,%.3f" % (elementType, nr.quarterLength) # back to diff# THIRD, get the deltas (max range up, max range down) based on where# the previous note was, +- minor 3. Skip rests (don't affect deltas).intervalInfo = ""if isinstance(nr, note.Note):numNonRests += 1if numNonRests == 1:prevNote = nrelse:noteDist = interval.Interval(noteStart=prevNote, noteEnd=nr)noteDistUpper = interval.add([noteDist, "m3"])noteDistLower = interval.subtract([noteDist, "m3"])intervalInfo = ",<%s,%s>" % (noteDistUpper.directedName, noteDistLower.directedName)# print "Upper, lower: %s, %s" % (noteDistUpper,# noteDistLower)# print "Upper, lower dnames: %s, %s" % (# noteDistUpper.directedName,# noteDistLower.directedName)# print "The interval: %s" % (intervalInfo)prevNote = nr# Return. Do lazy evaluation for real-time performance.grammarTerm = noteInfo + intervalInfo fullGrammar += (grammarTerm + " ")return fullGrammar.rstrip()''' Given a grammar string and chords for a measure, returns measure notes. '''
def unparse_grammar(m1_grammar, m1_chords):m1_elements = stream.Voice()currOffset = 0.0 # for recalculate last chord.prevElement = Nonefor ix, grammarElement in enumerate(m1_grammar.split(' ')):terms = grammarElement.split(',')currOffset += float(terms[1]) # works just fine# Case 1: it's a rest. Just appendif terms[0] == 'R':rNote = note.Rest(quarterLength = float(terms[1]))m1_elements.insert(currOffset, rNote)continue# Get the last chord first so you can find chord note, scale note, etc.try: lastChord = [n for n in m1_chords if n.offset <= currOffset][-1]except IndexError:m1_chords[0].offset = 0.0lastChord = [n for n in m1_chords if n.offset <= currOffset][-1]# Case: no < > (should just be the first note) so generate from range# of lowest chord note to highest chord note (if not a chord note, else# just generate one of the actual chord notes). # Case #1: if no < > to indicate next note range. Usually this lack of < ># is for the first note (no precedent), or for rests.if (len(terms) == 2): # Case 1: if no < >.insertNote = note.Note() # default is C# Case C: chord note.if terms[0] == 'C':insertNote = __generate_chord_tone(lastChord)# Case S: scale note.elif terms[0] == 'S':insertNote = __generate_scale_tone(lastChord)# Case A: approach note.# Handle both A and X notes here for now.else:insertNote = __generate_approach_tone(lastChord)# Update the stream of generated notesinsertNote.quarterLength = float(terms[1])if insertNote.octave < 4:insertNote.octave = 4m1_elements.insert(currOffset, insertNote)prevElement = insertNote# Case #2: if < > for the increment. Usually for notes after the first one.else:# Get lower, upper intervals and notes.interval1 = interval.Interval(terms[2].replace("<",''))interval2 = interval.Interval(terms[3].replace(">",''))if interval1.cents > interval2.cents:upperInterval, lowerInterval = interval1, interval2else:upperInterval, lowerInterval = interval2, interval1lowPitch = interval.transposePitch(prevElement.pitch, lowerInterval)highPitch = interval.transposePitch(prevElement.pitch, upperInterval)numNotes = int(highPitch.ps - lowPitch.ps + 1) # for range(s, e)# Case C: chord note, must be within increment (terms[2]).# First, transpose note with lowerInterval to get note that is# the lower bound. Then iterate over, and find valid notes. Then# choose randomly from those.if terms[0] == 'C':relevantChordTones = []for i in range(0, numNotes):currNote = note.Note(lowPitch.transpose(i).simplifyEnharmonic())if __is_chord_tone(lastChord, currNote):relevantChordTones.append(currNote)if len(relevantChordTones) > 1:insertNote = random.choice([i for i in relevantChordTonesif i.nameWithOctave != prevElement.nameWithOctave])elif len(relevantChordTones) == 1:insertNote = relevantChordTones[0]else: # if no choices, set to prev element +-1 whole stepinsertNote = prevElement.transpose(random.choice([-2,2]))if insertNote.octave < 3:insertNote.octave = 3insertNote.quarterLength = float(terms[1])m1_elements.insert(currOffset, insertNote)# Case S: scale note, must be within increment.elif terms[0] == 'S':relevantScaleTones = []for i in range(0, numNotes):currNote = note.Note(lowPitch.transpose(i).simplifyEnharmonic())if __is_scale_tone(lastChord, currNote):relevantScaleTones.append(currNote)if len(relevantScaleTones) > 1:insertNote = random.choice([i for i in relevantScaleTonesif i.nameWithOctave != prevElement.nameWithOctave])elif len(relevantScaleTones) == 1:insertNote = relevantScaleTones[0]else: # if no choices, set to prev element +-1 whole stepinsertNote = prevElement.transpose(random.choice([-2,2]))if insertNote.octave < 3:insertNote.octave = 3insertNote.quarterLength = float(terms[1])m1_elements.insert(currOffset, insertNote)# Case A: approach tone, must be within increment.# For now: handle both A and X cases.else:relevantApproachTones = []for i in range(0, numNotes):currNote = note.Note(lowPitch.transpose(i).simplifyEnharmonic())if __is_approach_tone(lastChord, currNote):relevantApproachTones.append(currNote)if len(relevantApproachTones) > 1:insertNote = random.choice([i for i in relevantApproachTonesif i.nameWithOctave != prevElement.nameWithOctave])elif len(relevantApproachTones) == 1:insertNote = relevantApproachTones[0]else: # if no choices, set to prev element +-1 whole stepinsertNote = prevElement.transpose(random.choice([-2,2]))if insertNote.octave < 3:insertNote.octave = 3insertNote.quarterLength = float(terms[1])m1_elements.insert(currOffset, insertNote)# update the previous element.prevElement = insertNotereturn m1_elements '''
inference_code.py
'''def inference_model(LSTM_cell, densor, n_x = 78, n_a = 64, Ty = 100):"""Uses the trained "LSTM_cell" and "densor" from model() to generate a sequence of values.Arguments:LSTM_cell -- the trained "LSTM_cell" from model(), Keras layer objectdensor -- the trained "densor" from model(), Keras layer objectn_x -- number of unique valuesn_a -- number of units in the LSTM_cellTy -- number of time steps to generateReturns:inference_model -- Keras model instance"""# Define the input of your model with a shape x0 = Input(shape=(1, n_x))# Define s0, initial hidden state for the decoder LSTMa0 = Input(shape=(n_a,), name='a0')c0 = Input(shape=(n_a,), name='c0')a = a0c = c0x = x0### START CODE HERE #### Step 1: Create an empty list of "outputs" to later store your predicted values (≈1 line)outputs = []# Step 2: Loop over Ty and generate a value at every time stepfor t in range(Ty):# Step 2.A: Perform one step of LSTM_cell (≈1 line)a, _, c = LSTM_cell(x, initial_state=[a, c])# Step 2.B: Apply Dense layer to the hidden state output of the LSTM_cell (≈1 line)out = densor(a)# Step 2.C: Append the prediction "out" to "outputs" (≈1 line)outputs.append(out)# Step 2.D: Set the prediction "out" to be the next input "x". You will need to use RepeatVector(1). (≈1 line)x = RepeatVector(1)(out)# Step 3: Create model instance with the correct "inputs" and "outputs" (≈1 line)inference_model = Model(inputs=[x0, a0, c0], outputs=outputs)### END CODE HERE ###return inference_modelinference_model = inference_model(LSTM_cell, densor)x1 = np.zeros((1, 1, 78))
x1[:,:,35] = 1
a1 = np.zeros((1, n_a))
c1 = np.zeros((1, n_a))
predicting = inference_model.predict([x1, a1, c1])indices = np.argmax(predicting, axis = -1)
results = to_categorical(indices, num_classes=78)'''midi.py'''"""
File: midi.py
Author: Addy771
Description:
A script which converts MIDI files to WAV and optionally to MP3 using ffmpeg.
Works by playing each file and using the stereo mix device to record at the same time
"""import pyaudio # audio recording
import wave # file saving
import pygame # midi playback
import fnmatch # name matching
import os # file listing#### CONFIGURATION ####do_ffmpeg_convert = True # Uses FFmpeg to convert WAV files to MP3. Requires ffmpeg.exe in the script folder or PATH
do_wav_cleanup = True # Deletes WAV files after conversion to MP3
sample_rate = 44100 # Sample rate used for WAV/MP3
channels = 2 # Audio channels (1 = mono, 2 = stereo)
buffer = 1024 # Audio buffer size
mp3_bitrate = 128 # Bitrate to save MP3 with in kbps (CBR)
input_device = 1 # Which recording device to use. On my system Stereo Mix = 1# Begins playback of a MIDI file
def play_music(music_file):try:pygame.mixer.music.load(music_file)except pygame.error:print ("Couldn't play %s! (%s)" % (music_file, pygame.get_error()))returnpygame.mixer.music.play()# Init pygame playback
bitsize = -16 # unsigned 16 bit
pygame.mixer.init(sample_rate, bitsize, channels, buffer)# optional volume 0 to 1.0
pygame.mixer.music.set_volume(1.0)# Init pyAudio
format = pyaudio.paInt16
audio = pyaudio.PyAudio()try:# Make a list of .mid files in the current directory and all subdirectoriesmatches = []for root, dirnames, filenames in os.walk("./"):for filename in fnmatch.filter(filenames, '*.mid'):matches.append(os.path.join(root, filename))# Play each song in the listfor song in matches:# Create a filename with a .wav extensionfile_name = os.path.splitext(os.path.basename(song))[0]new_file = file_name + '.wav'# Open the stream and start recordingstream = audio.open(format=format, channels=channels, rate=sample_rate, input=True, input_device_index=input_device, frames_per_buffer=buffer)# Playback the songprint("Playing " + file_name + ".mid\n")play_music(song)frames = []# Record frames while the song is playingwhile pygame.mixer.music.get_busy():frames.append(stream.read(buffer))# Stop recordingstream.stop_stream()stream.close()# Configure wave file settingswave_file = wave.open(new_file, 'wb')wave_file.setnchannels(channels)wave_file.setsampwidth(audio.get_sample_size(format))wave_file.setframerate(sample_rate)print("Saving " + new_file) # Write the frames to the wave filewave_file.writeframes(b''.join(frames))wave_file.close()# Call FFmpeg to handle the MP3 conversion if desiredif do_ffmpeg_convert:os.system('ffmpeg -i ' + new_file + ' -y -f mp3 -ab ' + str(mp3_bitrate) + 'k -ac ' + str(channels) + ' -ar ' + str(sample_rate) + ' -vn ' + file_name + '.mp3')# Delete the WAV file if desiredif do_wav_cleanup: os.remove(new_file)# End PyAudio audio.terminate() except KeyboardInterrupt:# if user hits Ctrl/C then exit# (works only in console mode)pygame.mixer.music.fadeout(1000)pygame.mixer.music.stop()raise SystemExit'''
music_utils.py
'''from __future__ import print_function
import tensorflow as tf
import keras.backend as K
from keras.layers import RepeatVector
import sys
from music21 import *
import numpy as np
from grammar import *
from preprocess import *
from qa import *def data_processing(corpus, values_indices, m = 60, Tx = 30):# cut the corpus into semi-redundant sequences of Tx valuesTx = Tx N_values = len(set(corpus))np.random.seed(0)X = np.zeros((m, Tx, N_values), dtype=np.bool)Y = np.zeros((m, Tx, N_values), dtype=np.bool)for i in range(m):
# for t in range(1, Tx):random_idx = np.random.choice(len(corpus) - Tx)corp_data = corpus[random_idx:(random_idx + Tx)]for j in range(Tx):idx = values_indices[corp_data[j]]if j != 0:X[i, j, idx] = 1Y[i, j-1, idx] = 1Y = np.swapaxes(Y,0,1)Y = Y.tolist()return np.asarray(X), np.asarray(Y), N_values def next_value_processing(model, next_value, x, predict_and_sample, indices_values, abstract_grammars, duration, max_tries = 1000, temperature = 0.5):"""Helper function to fix the first value.Arguments:next_value -- predicted and sampled value, index between 0 and 77x -- numpy-array, one-hot encoding of next_valuepredict_and_sample -- predict functionindices_values -- a python dictionary mapping indices (0-77) into their corresponding unique value (ex: A,0.250,< m2,P-4 >)abstract_grammars -- list of grammars, on element can be: 'S,0.250,<m2,P-4> C,0.250,<P4,m-2> A,0.250,<P4,m-2>'duration -- scalar, index of the loop in the parent functionmax_tries -- Maximum numbers of time trying to fix the valueReturns:next_value -- process predicted value"""# fix first note: must not have < > and not be a restif (duration < 0.00001):tries = 0while (next_value.split(',')[0] == 'R' or len(next_value.split(',')) != 2):# give up after 1000 tries; random from input's first notesif tries >= max_tries:#print('Gave up on first note generation after', max_tries, 'tries')# np.random is exclusive to highrand = np.random.randint(0, len(abstract_grammars))next_value = abstract_grammars[rand].split(' ')[0]else:next_value = predict_and_sample(model, x, indices_values, temperature)tries += 1return next_valuedef sequence_to_matrix(sequence, values_indices):"""Convert a sequence (slice of the corpus) into a matrix (numpy) of one-hot vectors corresponding to indices in values_indicesArguments:sequence -- python listReturns:x -- numpy-array of one-hot vectors """sequence_len = len(sequence)x = np.zeros((1, sequence_len, len(values_indices)))for t, value in enumerate(sequence):if (not value in values_indices): print(value)x[0, t, values_indices[value]] = 1.return xdef one_hot(x):x = K.argmax(x)x = tf.one_hot(x, 78) x = RepeatVector(1)(x)return x'''
preprocess.py
''''''
Author: Ji-Sung Kim
Project: deepjazz
Purpose: Parse, cleanup and process data.Code adapted from Evan Chow's jazzml, https://github.com/evancchow/jazzml with
express permission.
'''from __future__ import print_functionfrom music21 import *
from collections import defaultdict, OrderedDict
from itertools import groupby, zip_longestfrom grammar import *from grammar import parse_melody
from music_utils import *#----------------------------HELPER FUNCTIONS----------------------------------#''' Helper function to parse a MIDI file into its measures and chords '''
def __parse_midi(data_fn):# Parse the MIDI data for separate melody and accompaniment parts.midi_data = converter.parse(data_fn)# Get melody part, compress into single voice.melody_stream = midi_data[5] # For Metheny piece, Melody is Part #5.melody1, melody2 = melody_stream.getElementsByClass(stream.Voice)for j in melody2:melody1.insert(j.offset, j)melody_voice = melody1for i in melody_voice:if i.quarterLength == 0.0:i.quarterLength = 0.25# Change key signature to adhere to comp_stream (1 sharp, mode = major).# Also add Electric Guitar. melody_voice.insert(0, instrument.ElectricGuitar())melody_voice.insert(0, key.KeySignature(sharps=1))# The accompaniment parts. Take only the best subset of parts from# the original data. Maybe add more parts, hand-add valid instruments.# Should add least add a string part (for sparse solos).# Verified are good parts: 0, 1, 6, 7 '''partIndices = [0, 1, 6, 7]comp_stream = stream.Voice()comp_stream.append([j.flat for i, j in enumerate(midi_data) if i in partIndices])# Full stream containing both the melody and the accompaniment. # All parts are flattened. full_stream = stream.Voice()for i in range(len(comp_stream)):full_stream.append(comp_stream[i])full_stream.append(melody_voice)# Extract solo stream, assuming you know the positions ..ByOffset(i, j).# Note that for different instruments (with stream.flat), you NEED to use# stream.Part(), not stream.Voice().# Accompanied solo is in range [478, 548)solo_stream = stream.Voice()for part in full_stream:curr_part = stream.Part()curr_part.append(part.getElementsByClass(instrument.Instrument))curr_part.append(part.getElementsByClass(tempo.MetronomeMark))curr_part.append(part.getElementsByClass(key.KeySignature))curr_part.append(part.getElementsByClass(meter.TimeSignature))curr_part.append(part.getElementsByOffset(476, 548, includeEndBoundary=True))cp = curr_part.flatsolo_stream.insert(cp)# Group by measure so you can classify. # Note that measure 0 is for the time signature, metronome, etc. which have# an offset of 0.0.melody_stream = solo_stream[-1]measures = OrderedDict()offsetTuples = [(int(n.offset / 4), n) for n in melody_stream]measureNum = 0 # for now, don't use real m. nums (119, 120)for key_x, group in groupby(offsetTuples, lambda x: x[0]):measures[measureNum] = [n[1] for n in group]measureNum += 1# Get the stream of chords.# offsetTuples_chords: group chords by measure number.chordStream = solo_stream[0]chordStream.removeByClass(note.Rest)chordStream.removeByClass(note.Note)offsetTuples_chords = [(int(n.offset / 4), n) for n in chordStream]# Generate the chord structure. Use just track 1 (piano) since it is# the only instrument that has chords. # Group into 4s, just like before. chords = OrderedDict()measureNum = 0for key_x, group in groupby(offsetTuples_chords, lambda x: x[0]):chords[measureNum] = [n[1] for n in group]measureNum += 1# Fix for the below problem.# 1) Find out why len(measures) != len(chords).# ANSWER: resolves at end but melody ends 1/16 before last measure so doesn't# actually show up, while the accompaniment's beat 1 right after does.# Actually on second thought: melody/comp start on Ab, and resolve to# the same key (Ab) so could actually just cut out last measure to loop.# Decided: just cut out the last measure. del chords[len(chords) - 1]assert len(chords) == len(measures)return measures, chords''' Helper function to get the grammatical data from given musical data. '''
def __get_abstract_grammars(measures, chords):# extract grammarsabstract_grammars = []for ix in range(1, len(measures)):m = stream.Voice()for i in measures[ix]:m.insert(i.offset, i)c = stream.Voice()for j in chords[ix]:c.insert(j.offset, j)parsed = parse_melody(m, c)abstract_grammars.append(parsed)return abstract_grammars#----------------------------PUBLIC FUNCTIONS----------------------------------#''' Get musical data from a MIDI file '''
def get_musical_data(data_fn):measures, chords = __parse_midi(data_fn)abstract_grammars = __get_abstract_grammars(measures, chords)return chords, abstract_grammars''' Get corpus data from grammatical data '''
def get_corpus_data(abstract_grammars):corpus = [x for sublist in abstract_grammars for x in sublist.split(' ')]values = set(corpus)val_indices = dict((v, i) for i, v in enumerate(values))indices_val = dict((i, v) for i, v in enumerate(values))return corpus, values, val_indices, indices_val'''
def load_music_utils():chord_data, raw_music_data = get_musical_data('data/original_metheny.mid')music_data, values, values_indices, indices_values = get_corpus_data(raw_music_data)X, Y = data_processing(music_data, values_indices, Tx = 20, step = 3)return (X, Y)
'''
'''
qa.py
''''''
Author: Ji-Sung Kim, Evan Chow
Project: deepjazz
Purpose: Provide pruning and cleanup functions.Code adapted from Evan Chow's jazzml, https://github.com/evancchow/jazzml
with express permission.
'''
from itertools import zip_longest
import randomfrom music21 import *#----------------------------HELPER FUNCTIONS----------------------------------#''' Helper function to down num to the nearest multiple of mult. '''
def __roundDown(num, mult):return (float(num) - (float(num) % mult))''' Helper function to round up num to nearest multiple of mult. '''
def __roundUp(num, mult):return __roundDown(num, mult) + mult''' Helper function that, based on if upDown < 0 or upDown >= 0, rounds number down or up respectively to nearest multiple of mult. '''
def __roundUpDown(num, mult, upDown):if upDown < 0:return __roundDown(num, mult)else:return __roundUp(num, mult)''' Helper function, from recipes, to iterate over list in chunks of n length. '''
def __grouper(iterable, n, fillvalue=None):args = [iter(iterable)] * nreturn zip_longest(*args, fillvalue=fillvalue)#----------------------------PUBLIC FUNCTIONS----------------------------------#''' Smooth the measure, ensuring that everything is in standard note lengths (e.g., 0.125, 0.250, 0.333 ... ). '''
def prune_grammar(curr_grammar):pruned_grammar = curr_grammar.split(' ')for ix, gram in enumerate(pruned_grammar):terms = gram.split(',')terms[1] = str(__roundUpDown(float(terms[1]), 0.250, random.choice([-1, 1])))pruned_grammar[ix] = ','.join(terms)pruned_grammar = ' '.join(pruned_grammar)return pruned_grammar''' Remove repeated notes, and notes that are too close together. '''
def prune_notes(curr_notes):for n1, n2 in __grouper(curr_notes, n=2):if n2 == None: # corner case: odd-length listcontinueif isinstance(n1, note.Note) and isinstance(n2, note.Note):if n1.nameWithOctave == n2.nameWithOctave:curr_notes.remove(n2)return curr_notes''' Perform quality assurance on notes '''
def clean_up_notes(curr_notes):removeIxs = []for ix, m in enumerate(curr_notes):# QA1: ensure nothing is of 0 quarter note len, if so changes its lenif (m.quarterLength == 0.0):m.quarterLength = 0.250# QA2: ensure no two melody notes have same offset, i.e. form a chord.# Sorted, so same offset would be consecutive notes.if (ix < (len(curr_notes) - 1)):if (m.offset == curr_notes[ix + 1].offset andisinstance(curr_notes[ix + 1], note.Note)):removeIxs.append((ix + 1))curr_notes = [i for ix, i in enumerate(curr_notes) if ix not in removeIxs]return curr_notesReferences
The ideas presented in this notebook came primarily from three computational music papers cited below. The implementation here also took significant inspiration and used many components from Ji-Sung Kim’s github repository.
- Ji-Sung Kim, 2016, deepjazz
- Jon Gillick, Kevin Tang and Robert Keller, 2009. Learning Jazz Grammars
- Robert Keller and David Morrison, 2007, A Grammatical Approach to Automatic Improvisation
- François Pachet, 1999, Surprising Harmonies
We’re also grateful to François Germain for valuable feedback.
PS: 欢迎扫码关注公众号：「SelfImprovementLab」！专注「深度学习」，「机器学习」，「人工智能」。以及 「早起」，「阅读」，「运动」，「英语 」「其他」不定期建群 打卡互助活动。
这篇关于Assignment | 05-week1 -Improvise a Jazz Solo with an LSTM Network的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！
