Where and how does prefetch_factor work?

[yzholdfish]

a simple trainning

During the training, i found that there will be a long wait every other period of time, which corresponds to the value of num_workers.In dataloader, prefetch_factor is 2, i think the cycle should be prefetch_factor * num_workers

I commented out the calculation process in the picture 1, and the phenomenon is more obvious

i find the part of source code of dataloader

im confused
where does prefetch_factor work?
work in the whole time or just in the initial of the dataloader as the picture above shows

[eqy]

What is the number of workers here and have you tried increasing it? I’m curious if the number of workers isn’t enough to keep up with the device’s training throughput; note that if the selected number of workers cannot keep up in terms of throughput, prefetching alone cannot address the underlying issue as the workers will always be busy.