Unoptimal training results

[ty625911724]

Dear authors:

Thanks a lot your excellent works!

However, when I try to run the training code, I get strange memory consuming and unoptimal training results.
I just run the traing stage with the following cmd:
python scripts/train.py

Then, only about 8GB GPU memory is used instead of 80 GB for A100.
After the training, all the checkpoint trained by myself generate obviously worse results compared with "ffhq-sg2-510.ckpt".

Could you please give me some suggestions?

Thanks you! : )