ZhengPeng7 / BiRefNet

[CAAI AIR'24] Bilateral Reference for High-Resolution Dichotomous Image Segmentation
https://www.birefnet.top
MIT License
322 stars 28 forks source link

the training stalls #9

Closed rendenzvous closed 3 months ago

rendenzvous commented 3 months ago

the training stalls, always in the first iteration of the first epoch,my gpu is 16G,my dataset just 39 pictures。 ![Uploading 2222.png…]()

ZhengPeng7 commented 3 months ago

You picture was not uploaded successfully...

rendenzvous commented 3 months ago

2222 here is

ZhengPeng7 commented 3 months ago

But, what is the problem here?... It seems no trouble in training as the logs are being printed.

rendenzvous commented 3 months ago

You picture was not uploaded successfully...

it don't move. don't have any reaction. don't move to the next Interation and the next epoch

ZhengPeng7 commented 3 months ago

Really? you can decrease the resolution of input for training, e.g., 512x512, to test if it's just slow. I remember Windows has some mechanism to use CPU memory as a supplement for GPU memory, which though causes much much delay. BTW, as I mentioned in README, 36.5G GPU memory is needed to use my default setting.

rendenzvous commented 3 months ago

Really? you can decrease the resolution of input for training, e.g., 512x512, to test if it's just slow. I remember Windows has some mechanism to use CPU memory as a supplement for GPU memory, which though causes much much delay. BTW, as I mentioned in README, 36.5G GPU memory is needed to use my default setting.

thank you so much ,it has been sloved,best wishes to you!

ZhengPeng7 commented 3 months ago

Glad to hear that. Good luck to you too :)