search

MASILab / 3DUX-Net

237 stars 33 forks source link

Training my own data, the loss does not drop #40

Open MOMOANNIE opened 1 year ago

MOMOANNIE commented 1 year ago

Hello, when I train my own data, I set dataset == 'flare', out_classes = 2, when the training reaches 10,000 steps, why the loss does not drop? image The loss curve is as follows: image

leeh43 commented 1 year ago

Hi, thank you for your interest. The problem occurred here is related to the cropping. You can see there is a warning "N forground 0, background #######, unable to generate class balanced samples". Please double check your label input and the foreground region within the label. Seems like there is no foreground within the label region for them to crop.

MOMOANNIE commented 1 year ago

Hi, thank you for your interest. The problem occurred here is related to the cropping. You can see there is a warning "N forground 0, background #######, unable to generate class balanced samples". Please double check your label input and the foreground region within the label. Seems like there is no foreground within the label region for them to crop.

Thank you for answering my question in time, I checked the image and label of the training data. The following two pictures are the corresponding image and label. There will be multiple foregrounds in the label image, but some foregrounds will be very small. Will this affect it?

image image

s-j-y1 commented 10 months ago

您好，想请教一下如何换成自己的数据集

ghost commented 6 months ago

Hello, the dataset I use is also small target binary segmentation, does this model work reasonably on your dataset?