How to prefetch data?

[speedcell4]

Currently, the bottleneck of my model training is on the data loading part, is there any example about how to prefetch data? Like the pin_memory and num_workers arguments of torch.utils.data.DataLoader

[zhangguanheng66]

which dataset are you using now? We have re-written language modeling and text classification datasets to make them more compatible with torch.utils.data.DataLoader.

[Nayef211]

@speedcell4 lmk if you're still seeing these bottlenecks. There's also dataloader v2 coming out soon which should have some improvements on data loading perf.

[speedcell4]

a "coming out soon" without reference + closing issue without confirming
really good manner! @Nayef211

[Nayef211]

a "coming out soon" without reference + closing issue without confirming really good manner! @Nayef211

@speedcell4 we are currently going through the backlog of issues and responding to/closing out issues that we think are legacy and may no longer be relevant as we've made a lot of updates to our library over the last couple of years. Feel free to keep the issue open if you're still running into the performance bottlenecks.