比较与torchaudio.transforms.SpectralCentroid的差异

torchaudio.transforms.SpectralCentroid

class torchaudio.transforms.SpectralCentroid(sample_rate: int, n_fft: int = 400, win_length: Optional[int] = None,
                                             hop_length: Optional[int] = None, pad: int = 0,
                                             window_fn: Callable[[...], torch.Tensor] = <built-in method hann_window of type object>,
                                             wkwargs: Optional[dict] = None)

更多内容详见torchaudio.transforms.SpectralCentroid。

mindspore.dataset.audio.SpectralCentroid

class mindspore.dataset.audio.SpectralCentroid(sample_rate, n_fft=400, win_length=None, hop_length=None,
                                               pad=0, window=WindowType.HANN)

更多内容详见mindspore.dataset.audio.SpectralCentroid。

差异对比

PyTorch：计算每个通道沿时间轴的频谱中心。支持自定义窗函数或对窗函数传入不同的配置参数。支持对STFT结果进行幅值规范化。

MindSpore：计算每个通道沿时间轴的频谱中心。

分类	子类	PyTorch	MindSpore	差异
参数	参数1	sample_rate	sample_rate	-
	参数2	n_fft	n_fft	-
	参数3	win_length	win_length	-
	参数4	hop_length	hop_length	-
	参数5	pad	pad
	参数6	window_fn	window	MindSpore仅支持5种窗函数
	参数7	wkwargs	-	自定义窗函数的入参，MindSpore不支持

代码示例

import numpy as np

fake_input = np.array([[[1, 1, 2, 2, 3, 3, 4]]]).astype(np.float32)

# PyTorch
import torch
import torchaudio.transforms as T

transformer = T.SpectralCentroid(sample_rate=44100, n_fft=8, window_fn=torch.hann_window)
torch_result = transformer(torch.from_numpy(fake_input))
print(torch_result)
# Out: tensor([[[4436.1182, 3768.7986]]])

# MindSpore
import mindspore.dataset.audio as audio

transformer = audio.SpectralCentroid(sample_rate=44100, n_fft=8, window=audio.WindowType.HANN)
ms_result = transformer(fake_input)
print(ms_result)
# Out: [[[[4436.117  3768.7979]]]]