The question about imagenet feature

[stein-666]

Hi, thanks for your awesome work.
When I load the ImageNet feature downloaded from Hugging Face, an error is raised stating that the number of features does not equal 1,281,167. I wonder if I made a mistake when processing the data.

[YuchuanTian]

Hi, thank you for your interest in our work!
How many cards (subprocesses) have you inited for training? If the card number is less than 8, the error might be raised. A solution is to reallocate the batchsize dimension, e.g. loading several numpy in one process and concat them on the 0th dimension.