作为一个练习，我试图在多个短音频序列上训练LSTM，我的目标是能够复制被训练的音频样本的类型。首先，也许不是那么困难，我决定在一堆脚鼓样本上对其进行训练。用于测试目的的小型数据集here on my github
现在，我遇到的问题是如何拆分我的数据进行训练：
还有另一个问题，传递给fit函数的参数应该是什么？使用python生成器在keras中将音频样本切分到.fit_generator()函数中，这是我到目前为止尝试过的，并且不起作用。它可以编译并开始训练，但是我很快就耗尽了内存。
我试图使我的代码尽可能容易理解，但是可能有点冗长：
# Audio LSTMv5
from __future__ import print_function
import tensorflow as tf
import numpy as np
import librosa
# Target log path
logs_path = '/tmp/tensorflow/audio-rnn'
writer = tf.summary.FileWriter(logs_path)
'''
Audio Loader Module | Librosa + Tensorflow wrapper | Returns Tensor of shape (#of Audio Files, #of Samples, #of Channels)
Load Audio | Preprocessing
'''
from audio_loader_and_encoder import lqe
quantization_channels = 256
loader = lqe()
# Returns Audio File given the path to said file
def load_audios(path, loader):
# code that imports an audio file as an array of values
# signature -> (filepath, sample_rate, duration of audio (1 second), mono, plot graph that's to be loaded)
audio = loader.load_audio(path, 8000, 1, True, False)
print("Loaded Audio")
min_audio = np.min(audio)
max_audio = np.max(audio)
# normalize
audio = (audio - min_audio) / (max_audio - min_audio)
print(audio.dtype, min_audio, max_audio)
return audio, min_audio, max_audio
'''__________________________________________________________________________________________________________________________________________________________'''
'''
Splits Audios into smaller slices to train on
'''
from tqdm import tqdm
def sample_generator(audio):
# try to estimate next_sample (0 -255) based on 256 previous samples
step = 5
next_sample = []
samples = []
for j in tqdm(range(0, audio.shape[0] - maxlen, step)):
seq = audio[j: j + maxlen + 1]
seq_matrix = np.zeros((maxlen, nb_output), dtype = bool)
for i,s in enumerate(seq):
sample_ = int(s * (nb_output - 1)) # 0-255
if i < maxlen:
seq_matrix[i, sample_] = True
else:
seq_vec = np.zeros(nb_output, dtype=bool)
seq_vec[sample_] = True
next_sample.append(seq_vec)
samples.append(seq_matrix)
samples = np.array(samples, dtype=bool)
next_sample = np.array(next_sample, dtype=bool)
#print(samples.shape, next_sample.shape)
return samples, next_sample
'''
Don't call fit in a loop, you're just retraining last part, read the issue here:
https://stackoverflow.com/questions/51373088/is-there-a-difference-between-calling-fit-in-a-loop-vs-fit-with-batch-size/51373594
Read the function descriptions here:
https://keras.io/models/model/
'''
'''__________________________________________________________________________________________________________________________________________________________'''
'''
Data generator | Yields batches to train on
'''
directory_name = "Kick"
audio, min_audio, max_audio = 0, 0, 0
def audio_generator(list):
for filename in list:
print(directory_name + "\\" + str(filename))
audio, min_audio, max_audio = load_audios(directory_name + "\\" + str(filename), loader)
samples, next_sample = sample_generator(audio)
yield samples, next_sample
'''__________________________________________________________________________________________________________________________________________________________'''
#This is just to organize the files on my pc
def list_sorter(audio_list):
sorted = []
for i, audio in enumerate(audio_list):
sorted.append(str(i+1) + ".wav")
return sorted
def generator_test():
for generated in audio_generator(list_sorter(os.listdir(directory_name))):
print("Generating", generated)
input("Press Enter to continue...")
'''
Parameters | Model Setup
'''
import os
os.environ["KERAS_BACKEND"] = "tensorflow"
from keras.models import Sequential, Model
from keras.layers import Dense, Activation, Dropout
from keras.layers import LSTM
from keras.layers import Input
from keras.optimizers import RMSprop
# so try to estimate next sample afte given (maxlen) samples
maxlen = 256 # 128 / sr = 0.016 sec
nb_output = 256 # resolution - 8bit encoding
latent_dim = 256
inputs = Input(shape=(maxlen, nb_output))
x = LSTM(latent_dim, return_sequences=True)(inputs)
x = Dropout(0.4)(x)
x = LSTM(latent_dim)(x)
x = Dropout(0.4)(x)
output = Dense(nb_output, activation='softmax')(x)
model = Model(inputs, output)
#optimizer = Adam(lr=0.005)
optimizer = RMSprop(lr=0.01)
model.compile(loss='categorical_crossentropy', optimizer=optimizer)
from keras.callbacks import EarlyStopping, CSVLogger, ModelCheckpoint
csv_logger = CSVLogger('training_audio.log')
escb = EarlyStopping(monitor='val_loss', patience=20, verbose=1)
# had error here https://keras.io/callbacks/
filepath = "models/audio-{epoch:.1f}-{val_loss:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, period=2)
'''
Helpers and Plotters
'''
# Plot loss function
import matplotlib.pyplot as plt
def plot_history():
print("Training history")
fig = plt.figure(figsize=(10,4))
ax1 = fig.add_subplot(1, 2, 1)
plt.plot(model.history.history['loss'])
ax1.set_title('loss')
ax2 = fig.add_subplot(1, 2, 2)
plt.plot(model.history.history['val_loss'])
ax2.set_title('validation loss')
plt.show()
# load array to audio buffer and play!!
from IPython.display import Audio, display
# predict and plot audio waveform
def test_model():
def sample(preds, temperature=1.0, min_value=0, max_value=1):
# helper function to sample an index from a probability array
preds = np.asarray(preds).astype('float64')
preds = np.log(preds) / temperature
exp_preds = np.exp(preds)
preds = exp_preds / np.sum(exp_preds)
probas = np.random.multinomial(1, preds, 1)
v = np.argmax(probas)/float(probas.shape[1])
return v * (max_value - min_value) + min_value
for start in range(5000,15000,10000):
seq = audio[start: maxlen]
seq_matrix = np.zeros((maxlen, nb_output), dtype=bool)
for i,s in enumerate(seq):
sample_ = int(s * (nb_output - 1)) # 0-255
seq_matrix[i, sample_] = True
for i in tqdm(range(5000)):
z = model.predict(seq_matrix.reshape((1,maxlen,nb_output)))
s = sample(z[0], 1.0)
seq = np.append(seq, s)
sample_ = int(s * (nb_output - 1))
seq_vec = np.zeros(nb_output, dtype=bool)
seq_vec[sample_] = True
seq_matrix = np.vstack((seq_matrix, seq_vec)) # added generated note info
seq_matrix = seq_matrix[1:]
# scale back
seq = seq * (max_audio - min_audio) + min_audio
print("Saving audio")
librosa.output.write_wav("audio_sample.wav", seq, 16000)
# plot
plt.figure(figsize=(30,5))
plt.plot(seq.transpose())
plt.show()
display(Audio(seq, rate=16000))
'''__________________________________________________________________________________________________________________________________________________________'''
# run model
def run():
audio_list = list_sorter(os.listdir(directory_name))
model.fit_generator(audio_generator(audio_list), initial_epoch = 0, max_q_size = 2, steps_per_epoch = 64, shuffle = True, verbose = 1, epochs = 20, callbacks=[csv_logger, escb, checkpoint])
plot_history()
test_model()
run()
'''_________________________________________________________________________'''