Quetion for multilabel UNET

[jinxsfe]

Hi, thanks for your suggest and support ,

I am tried to use UNET multilabel to segment the online data you provided, I had notice that there are label amount button for notebook provided. I open the semantic segmentation label for image platform, and which shows

like multiple color, so the color(images color>8)amount equal label amount? for example, the color is around 11 so label should be 11?

[jinxsfe]

the label amount that I choose is 1 so validation procedure did not improve

[jinxsfe]

also whether I input online data as you provided(using default training parameter) and my own data, training progress is failed due to


also when using online data

sees algorithms can not catch the signal

[jinxsfe]

@IvanHCenalmor appreciate for your help, If it's parameter issue , I will adjust it via my self

[IvanHCenalmor]

Hi @jinxsfe,

With online data, are you refering to the DeepBacs - Multi-label U-Net training dataset?

If that it's the case, as it is explained on the Notes.txt:

To train the multilabel-U-Net, use the `target_boundaries` images as GT. 

Also if that it's the case, you will need to set the label number to 3. The data, as it is explained on zenodo has 2 labels:

... manually annotated instance segmentation mask and the corresponding 2-label semantic segmentation mask used for model training.

So then, as it is explained on the label section on the notebook, you need to set the label number to 3, because it's the 2 labels plus the background:

labels: The number of different labels that the network needs to learn, which also includes the background. For example: to segment two different kind of objects in an image (cats and dogs), labels = 3 (2 labels for the two kinds and one more label for the background).

If you set the parameters like that it should work:

Best regards,
Iván