Fine tune birefnet

[agarwalma]

Hi,

I want to fine tune the birefnet model. I tried but couldn't find the model weights folder as shown in youtube video.

Shall i use model.safetensors model file from huggingface? I also created the dataset folders with training and testing data as described in video. Could you please let me know and clearly define the fine tuning of birefnet?

Thanks!

[ZhengPeng7]

You can read the README first. I've uploaded all weights to GitHub release and my google drive folder.

[agarwalma]

okay, Thanks for prompt reply. is that okay to just download this weight?
birefnet_general_epoch_244.pth zip file?

[ZhengPeng7]

Yeah, directly download them and put them in the right path. I didn't zip them.

[agarwalma]

How much training dataset or size of machine for training you would recommend (in cloud, A100-40G should be enough) for fine tuning, as described in video?

[ZhengPeng7]

It usually depends on the annotation's difficulty level and the amount of your data. If there is harder and more data, increasing epochs may still bring benefits.
However, if your task and annotation are very simple, some basic finetuning is already enough, where more epochs are not worthwhile.
For instance, for some simple seg tasks without complex shapes, 300-500 training samples for 30-100 epochs should be enough for basic performance.