Terry Zhang comments

Results 31 comments of


                                            Terry Zhang

如何控制一张图中文本的个数？

我尝试了一下，调整long_side参数可以增加图像中文本数量，或者用更大的背景图像也可以在一张图中生成更多。但long_side调大会显著降低程序运行速度。

Why does the 'loss=nan' appear when finetuning model using fp16 or bf16 mixed precision in 'fine_tune.py' ?

Try disabling xformers, or changing another version. I 've never got NAN since disabling xformers.

Code for dataset split

market_sct.txt can be easily generated as below: ``` for person A in person ids: collect camera ids of all images of person A; randomly select a camera id C from...

During training, loss=nan and a broken LoRA is generated

I solve this issue by disabling xformers during training.

Cannot train LoRAs while offline?

The training scripts load models from huggingface (.from_pretrained). If these models haven't been cached in your local storage, you may get error when training offline.

The prompt appears when the initializing checkpoint appears "This is not exten if you are initializing cliptextModel ..."

I also meet the same issue, saying "vision_model.xx not used". I believe it will not affect the training, as vision model in CLIP is never used in the text encoder....

Terry Zhang

如何控制一张图中文本的个数？

Why does the 'loss=nan' appear when finetuning model using fp16 or bf16 mixed precision in 'fine_tune.py' ?

Code for dataset split

During training, loss=nan and a broken LoRA is generated

Cannot train LoRAs while offline?

The prompt appears when the initializing checkpoint appears "This is not exten if you are initializing cliptextModel ..."

您好

lsvt四点坐标顺序

pickle protocal

make subdir configurable