如何在 sound_classification_ml_production 中正确生成 .npy 特征？

Question

我运行 github 回购代码 sound_classification_ml_production 并获得了很好的准确性，但发现 .npy 功能是作者预先制作的。所以我想生成自己的 .npy 并添加代码：

# FeatureExtractor class including librosa audio processing functions
class FeatureExtractor:

...


def compute_save_features(self, 
                    mode='mfcc', 
                    sample_rate=22050,
                    n_fft=2048,
                    hop_length=512,
                    n_mfcc=40,
                    output_path='features',
                    deltas=False
                    ):
    dataset_features = []
    max_pad = self._compute_max_pad_length(self.max_audio_duration, 
                                           sample_rate=sample_rate, 
                                           n_fft=n_fft,
                                           hop_length=hop_length)
    print('Max Padding = ', max_pad)
    
    if not os.path.exists(output_path):
        print('Creating output folder: ', output_path)
        os.makedirs(output_path)
    else:
        print('Output folder already existed')
        
    print('Saving features in ', output_path)
    i = 0
    t = time.time()
    
    features_path = []
    for relative_filepath in self.dataset_df['filepath']:
        filepath = base_path + relative_filepath;
        print('compute_save_features, filepath = ' + str(filepath))

        if i % 100 == 0:
            print('{} files processed in {}s'.format(i, time.time() - t))

        print('compute_save_features, librosa.load, filepath = ' + str(filepath))
        audio_file, sample_rate = librosa.load(filepath, sr=sample_rate, res_type='kaiser_fast')
        if mode == 'mfcc':
            audio_features = self.compute_mfcc(audio_file, sample_rate, n_fft, hop_length, n_mfcc, deltas)  
        elif mode == 'stft':
            audio_features = self.compute_stft(audio_file, sample_rate, n_fft, hop_length)
        elif mode == 'mel-spectogram':
            audio_features = self.compute_mel_spectogram(audio_file, sample_rate, n_fft, hop_length)
        
        audio_features = np.pad(audio_features, pad_width=((0, 0), (0, max_pad - audio_features.shape[1])))
        print('compute_save_features, audio_features = ' + str(type(audio_features)) + ', ' + str(audio_features))

        # here I add code and save audio_features to my .npy
        npy_path = os.path.join(output_path, filepath.split('/')[-1].replace('wav', 'npy'))
        print('compute_save_features, npy_path = ' + str(npy_path))
        np.save(npy_path, audio_features)
        
        ...

fe = FeatureExtractor(base_path + 'UrbanSound8K/metadata/UrbanSound8K.csv')

dataset_df = fe.compute_save_features(mode='mfcc', n_mfcc=13, output_path=base_path + 'my_features_mfcc', deltas=True)

...

与作者原版.npy对比后，有很多不同：

任何人都可以告诉我如何生成与 repo 代码显示中完全相同的 .npy 功能，在此先感谢。

Answer 1

Keras 层是随机初始化的。要重现相同的结果，您必须设置相同的种子。如果你没有种子价值 - 你就没有机会重现。

如何在 sound_classification_ml_production 中正确生成 .npy 特征？

How to produce .npy features correctly in sound_classification_ml_production?

python

tensorflow

keras

librosa