David Shi

you can download the problem dataset from: http://www.smallworld-network.com.cn/img/upsidedownhouse.tar.gz

btw, if I rescale the photos down to 1024x768, the problem goes away, but the model loses most of its detail due to low resolution of the images.

I manually rescaled the data set down to 1024x768, and used that to run the whole process, the problem goes away. After that, I tried 2048x1536, the refined model broke...

I am training it on CelebA Dataset. I cropped all the faces from the dataset to size 128x128 and the LR images are 32x32 upscaled to 128x128 using bicubic filter....

用了最新的版本PEFT还是同样的错误。 ```bash (textgen) [shiqi@nqs-server Chinese-Vicuna]$ ./generate.sh bin /home/shiqi/anaconda3/envs/textgen/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cuda117.so Namespace(model_path='decapoda-research/llama-7b-hf', lora_path='Chinese-Vicuna/Chinese-Vicuna-lora-7b-belle-and-guanaco', use_typewriter=0, use_local=0) The tokenizer class you load from this checkpoint is not the same type as the class this function...

> @mrbemani 但是不应该是最新版吧，这个问题就是因为PEFT一直在更新导致的问题。所以我们现在是固定了以前的一个版本。 是最新版的Chinese-Vicuna，标点符号忘记打了，哈哈