About the speed at which pseudo-tags are generated

When I use the maskcut.py to generate pseudo-labels, it takes 3 hours for each set of data in ImageNet, and there are 1000 sets of such data, is this normal, and what can I do to improve the running speed
Hope to get your answer, I will be very grateful,

```
Loading weight from https://dl.fbaipublicfiles.com/dino/dino_vitbase8_pretrain/dino_vitbase8_pretrain.pth
Load base pre-trained feature...
0%|▊ | 6/1300 [01:22<4:51:28, 13.52s/it
```