Questions about ThreadDataloader

[chaoscls]

I think ThreadDataloader is just use another thread to fetch data, and put it into a buffer queue. I don't know if it really makes sense to do that, or if it is really faster to use ThreadDataloader? Because I found there is a similar queue to buffer data in pytorch's Dataloader (https://github.com/pytorch/pytorch/blob/f4228e7037c614ed10b4d5c6fbd1467999e8d3e9/torch/utils/data/dataloader.py#LL1015C4-L1015C4).
So I'm really confused when should I use ThreadDataloader?

[KumoLiu]

Hi @chaoscls, typical usage is to accelerate light-weight preprocessing (usually cached all the deterministic transforms and no IO operations), because it leverages the separate thread to execute preprocessing to avoid unnecessary IPC between multiple workers of DataLoader. And as CUDA may not work well with the multi-processing of DataLoader, ThreadDataLoader can be useful for GPU transforms.
For more details, you can refer to https://github.com/Project-MONAI/tutorials/blob/main/acceleration/fast_model_training_guide.md#4-threaddataloader-vs-dataloader.

Hope it can help you, thanks!