Tencent / ncnn

ncnn is a high-performance neural network inference framework optimized for the mobile platform
Other
20.55k stars 4.18k forks source link

请求支持torch.std 、torch.stft和torch.istft #5077

Open Baiyuetribe opened 1 year ago

Baiyuetribe commented 1 year ago

error log | 日志或报错信息 | ログ

现有条件下,pnnx转换报错

ignore torch.std torch.std_1 param dim=(2,3)
ignore torch.std torch.std_1 param keepdim=False
ignore torch.std torch.std_1 param unbiased=True
ignore torch.istft torch.istft_298 param center=True
ignore torch.istft torch.istft_298 param hop_length=1024
ignore torch.istft torch.istft_298 param length=347136
ignore torch.istft torch.istft_298 param n_fft=4096
ignore torch.istft torch.istft_298 param normalized=True
ignore torch.istft torch.istft_298 param onesided=None
ignore torch.istft torch.istft_298 param return_complex=False
ignore torch.istft torch.istft_298 param win_length=4096
ignore torch.stft torch.stft_297 param center=False
ignore torch.stft torch.stft_297 param hop_length=1024
ignore torch.stft torch.stft_297 param n_fft=4096
ignore torch.stft torch.stft_297 param normalized=True
ignore torch.stft torch.stft_297 param onesided=None
ignore torch.stft torch.stft_297 param return_complex=True
ignore torch.stft torch.stft_297 param win_length=4096

stft在图像、音频领域非常有效,比如sd绘画模型中的free u;音频分离模型decums;语音降噪模型mp-senet、类HiFiGAN的声码器vocos等

nihui commented 1 week ago

https://github.com/Tencent/ncnn/pull/5779