dydxdt

I am working on the TF Lite model on Android, and I have got into trouble for a while. I wonder this problem: except some classification, detection, pose estimation and...

about ckpt model

1

您好，请问下为什么提供的训练好的模型是ckpt文件，不是pth文件？是用tf训练的吗？并且在轻量化上，LightViT可能主要对比的结构是transformer一类，而不是mobilenet等常用的轻量级网络？因为自己训练出的LightViT-small模型有200多M。谢谢～

How to prepare the tfrecords? Can you give a valid link? The link mentioned is invalid now.

1

Thank you!

Is 'normalized' classifier necessary?

3

Thanks for your code. I just wonder the reason why we need to normalize the 'classifier' before the 'flatten' op. Does it perform bettern than that without normalizing? Thank you...

About freeze and unfreeze?

2

您好，想问下训练代码里先只freeze最后一层训练30个epoch，然后freeze所有层训练50epoch，这样做是一般比直接加载预训练模型的权重效果会好吗？还是出于什么样的考虑呀

Question about training

1

Your job is great! I have some questions about training epochs. I want to train BrushNet on my own data, and I see the default training epoch is 10000. And...

Thanks for you great work. I can't find the whole tag list file, like "tags.txt". I can find the tags.txt in the DeepDanbooru project, so can you tell me where...

Thanks for your good work. I use the offered SDXL weights to finetune with my own data, but it seems the loss doesn't converge and I wonder whether the offered...

Question about full-parameter finetuning

2

Thx for your great work! I have a question about training arguments, i.e. is the parameter max_steps=10000 proper for full-parameter finetuning? I use my own train datset for full-parameter finetuning...