Running MedSAM on instance segmentation dataset

[sheldonsebastianNIH]

Hi

I am trying to train the SAM model for instance segmentation for multiple cells in image. I have the ground truth masks and bounding boxes.

I have prepared the training data by creating binary masks for each cell along with bounding boxes

• Source Image 1
• Ground Truth 1
  • Binary Cell_1 Mask + bounding box co-ordinates
  • Binary Cell_2 Mask + bounding box co-ordinates
• Source Image 2
• Ground Truth 2
  • Binary Cell_1 Mask + bounding box co-ordinates
  • Binary Cell_2 Mask + bounding box co-ordinates

But when training the model I get the following error:
RuntimeError: stack expects each tensor to be equal size, but got [340, 4] at entry 0 and [276, 4] at entry 1

Should I create binary mask combining all the cell masks into single image?

[[Edit]]:

• I see that we need to pass single label at a time:
  • https://github.com/bowang-lab/MedSAM/blob/main/train_one_gpu.py#L95C25-L95C25