open-mmlab/mmengine

[Feature] Serialize data list to torch.Tensor

Opened this issue · 0 comments

What is the feature?

The base dataset current serializes data lists to np.ndarray. As shown in this post https://ppwwyyxx.com/blog/2022/Demystify-RAM-Usage-in-Multiprocess-DataLoader/ and https://github.com/facebookresearch/detectron2/blob/main/detectron2/data/common.py#L173, serializing to torch.Tensor might be better. So I am wondering why the base dataset in mmengine does not serialize data lists to torch.Tensor?

Any other context?

No response