Yi Ma

和scipy.fftpack.dct计算的也不一样

I used the latest version. It works well with make_voxceleb1_v2.pl in voxceleb/v2

I wrote a make_voxceleb1_v2.pl for sitw according to egs/voxceleb: ``` if (@ARGV != 2) { print STDERR "Usage: $0 \n"; print STDERR "e.g. $0 /export/voxceleb1 data/\n"; exit(1); } ($data_base, $out_dir)...

应该是np.sqrt(clean_mag ** 2 / (clean_mag**2 + noise_mag** 2))吧

I also meet same error on Ubuntu since I update pytorch to 1.7.0

Thanks for your response but what's UIO mode? And I think maybe it can be implemented in class DistributedSampler in dataset.py?

> > Thanks for your response but what's UIO mode? And I think maybe it can be implemented in class DistributedSampler in dataset.py? > > Check our [paper](https://arxiv.org/abs/2210.17016) for introduction...

I just noticed that I set the data type as 'raw', and the above code is not appropriate for 'shard'