cannot re-initialize CUDA in forked subproess

[chumingqian]

Is any friend meet this situation,

i already solve this problem :
by using

import torch.multiprocessing as mp

def get_mean_and_std_4channel(dataset):
    '''Compute the mean and std value of dataset.'''
    mp.set_start_method('spawn') # set multiprocessing context to 'spawn'
    dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True, num_workers=2, multiprocessing_context='spawn')
    mean = torch.zeros(4)

However , the cuda also dumped because out of memory, for the 380*400 resolution, batch size =2, channel=3, nfft =2048, but i set trainable Mel and stft both are False.

It takes 23 G cuda memory,

so, is there any method reduce the cuda memory cost

[KinWaiCheuk]

I am not sure how is your dataset written.
Have you tried putting the nnAudio after the dataloader instead of inside it? What is the memory required if you do it like this?

i.e. you load a batch of waveforms first, then use nnAudio to convert that batch of waveforms into spectrograms.