keras：如何编写自定义的损失函数来聚合帧级预测到歌曲级预测

Posted 2023-02-25

【中文标题】keras：如何编写自定义的损失函数来聚合帧级预测到歌曲级预测

【英文标题】：keras: how to write customized loss function to aggregate over frame-level predictions to song-level prediction

【发布时间】：2019-03-21 01:29:03

【问题描述】：

我正在做歌曲流派分类（2 类）。对于每首歌曲，我将它们切成小帧 (5s) 以生成 MFCC 作为神经网络的输入特征，并且每一帧都有一个相关的歌曲流派标签。

数据如下所示：

 name         label   feature
 ....
 song_i_frame1 label   feature_vector_frame1
 song_i_frame2 label   feature_vector_frame2
 ...
 song_i_framek label   feature_vector_framek
 ...

我知道我可以随机选择 80% 的歌曲（它们的小帧）作为训练数据，其余的作为测试数据。但是现在我写 X_train 的方式是帧级别的帧，而 biney 交叉熵损失函数是在帧级别定义的。我想知道如何自定义损失函数，使其在帧级预测的聚合（例如歌曲的每一帧预测的多数票）上最小化。

目前，我拥有的是：

model_19mfcc = Model(input_shape = (X_train19.shape[1], X_train19.shape[2]))
model_19mfcc.compile(loss='binary_crossentropy', optimizer="RMSProp", metrics=["accuracy"])
history_fit = model_19mfcc.fit(X_train19, y_train,validation_split=0.25, batch_size = 1800/50, epochs= 200)

此外，当我将训练和测试数据输入 keras 时，数据的相应 ID（名称）会丢失，将数据（名称、等级和特征）保存在单独的 pandas 数据框中并匹配预测来自keras的一个好习惯？还是有其他好的选择？

提前致谢！

【参考方案1】：

类型分类通常不需要自定义损失函数。 可以使用Multiple Instance Learning (MIL) 设置将歌曲拆分为多个预测窗口的组合模型。

MIL 是一种监督学习方法，其中标签不是在每个独立样本（实例）上，而是在实例的“袋子”（无序集）上。 在您的情况下，实例是每 5 秒的 MFCC 功能窗口，而包就是整首歌曲。

在 Keras 中，我们使用TimeDistributed 层为所有窗口执行我们的模型。 然后我们使用GlobalAveragePooling1D 组合结果，有效地 跨窗口实施平均投票。这比多数投票更容易区分。

下面是一个可运行的例子：

import math

import keras
import librosa
import pandas
import numpy
import sklearn

def window_model(n_bands, n_frames, n_classes, hidden=32):
   from keras.layers import Input, Dense, Flatten, Conv2D, MaxPooling2D

   out_units = 1 if n_classes == 2 else n_classes
   out_activation = 'sigmoid' if n_classes == 2 else 'softmax'

   shape = (n_bands, n_frames, 1)

   # Basic CNN model
   # An MLP could also be used, but may need to reshape on input and output
   model = keras.Sequential([
       Conv2D(16, (3,3), input_shape=shape),
       MaxPooling2D((2,3)),
       Conv2D(16, (3,3)),
       MaxPooling2D((2,2)),
       Flatten(),
       Dense(hidden, activation='relu'),
       Dense(hidden, activation='relu'),
       Dense(out_units, activation=out_activation),
   ])
   return model

def song_model(n_bands, n_frames, n_windows, n_classes=3):
    from keras.layers import Input, TimeDistributed, GlobalAveragePooling1D

    # Create the frame-wise model, will be reused across all frames
    base = window_model(n_bands, n_frames, n_classes)
    # GlobalAveragePooling1D expects a 'channel' dimension at end
    shape = (n_windows, n_bands, n_frames, 1)

    print('Frame model')
    base.summary()

    model = keras.Sequential([
        TimeDistributed(base, input_shape=shape),
        GlobalAveragePooling1D(),
    ])

    print('Song model')
    model.summary()

    model.compile(loss='categorical_crossentropy', optimizer='SGD', metrics=['acc'])
    return model


def extract_features(path, sample_rate, n_bands, hop_length, n_frames, window_length, song_length):
    # melspectrogram might perform better with CNNs
    from librosa.feature import mfcc

    # Load a fixed length section of sound
    # Might need to pad if some songs are too short
    y, sr = librosa.load(path, sr=sample_rate, offset=0, duration=song_length)
    assert sr == sample_rate, sr
    _song_length = len(y)/sample_rate

    assert _song_length == song_length, _song_length

    # Split into windows
    window_samples = int(sample_rate * window_length)
    window_hop = window_samples//2 # use 50% overlap
    windows = librosa.util.frame(y, frame_length=window_samples, hop_length=window_hop)

    # Calculate features for each window
    features = []
    for w in range(windows.shape[1]):
        win = windows[:, w]
        f = mfcc(y=win, sr=sample_rate, n_mfcc=n_bands,
                 hop_length=hop_length, n_fft=2*hop_length)
        f = numpy.expand_dims(f, -1) # add channels dimension 
        features.append(f)

    features = numpy.stack(features)
    return features

def main():

    # Settings for our model
    n_bands = 13 # MFCCs
    sample_rate = 22050
    hop_length = 512
    window_length = 5.0
    song_length_max = 1.0*60
    n_frames = math.ceil(window_length / (hop_length/sample_rate))
    n_windows = math.floor(song_length_max / (window_length/2))-1

    model = song_model(n_bands, n_frames, n_windows)

    # Generate some example data
    ex =  librosa.util.example_audio_file()
    examples = 8
    numpy.random.seed(2)
    songs = pandas.DataFrame(
        'path': [ex] * examples,
        'genre': numpy.random.choice([ 'rock', 'metal', 'blues' ], size=examples),
    )
    assert len(songs.genre.unique() == 3) 

    print('Song data')
    print(songs)

    def get_features(path):
        f = extract_features(path, sample_rate, n_bands,
                    hop_length, n_frames, window_length, song_length_max)
        return f

    from sklearn.preprocessing import LabelBinarizer

    binarizer = LabelBinarizer()
    y = binarizer.fit_transform(songs.genre.values)
    print('y', y.shape, y)

    features = numpy.stack([ get_features(p) for p in songs.path ])
    print('features', features.shape)

    model.fit(features, y) 


if __name__ == '__main__':
    main()

示例输出内部和组合模型摘要：

Frame model
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 11, 214, 16)       160       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 5, 71, 16)         0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 3, 69, 16)         2320      
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 1, 34, 16)         0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 544)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 32)                17440     
_________________________________________________________________
dense_2 (Dense)              (None, 32)                1056      
_________________________________________________________________
dense_3 (Dense)              (None, 3)                 99        
=================================================================
Total params: 21,075
Trainable params: 21,075
Non-trainable params: 0

_________________________________________________________________
Song model
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
time_distributed_1 (TimeDist (None, 23, 3)             21075     
_________________________________________________________________
global_average_pooling1d_1 ( (None, 3)                 0         
=================================================================
Total params: 21,075
Trainable params: 21,075
Non-trainable params: 0
_________________________________________________________________

以及输入模型的特征向量的形状：

features (8, 23, 13, 216, 1)

8 首歌曲，每首 23 个窗口，13 个 MFCC 频段，每个窗口 216 帧。 还有一个大小为 1 的第五维度，让 Keras 开心……

【讨论】：

谢谢！我只是想知道，在我只有大约 150 首歌曲进行分类的情况下，我应该使用一些预训练模型（如果是的话，对预训练模型有什么建议），因为这个数据集的大小在聚合到歌曲级别？

如果你能找到一个在某些音乐分类任务上预训练的模型，这将是对其进行微调的理想选择。但是没有那么多预训练的音乐网。

您可以使用数据增强来人为地增加数据集的大小。使用许多不同的时移可以有很大帮助，音高变换和时间拉伸（少量）也可以。