Hed about hed-unet HOT 7 CLOSED

Kindly refer to Fig. 3 in the paper (https://arxiv.org/pdf/2103.01849.pdf). The attention is not in the channel dimension, but in the "resolution" dimension. There is only one query (the output), and one key/value for each resolution level.

In order to use it for your segmentation project, you're going to need to adapt data_loading.py for your own dataset and then simply call python train.py.

from hed-unet.

from hed-unet.

Thanks for your reply! I had met another question about how to visualize the effective receptive field like fig9 in the article. Can you show me the code? This can help me a lot. Thanks again!

from hed-unet.

The general idea is to backpropagate the output of a central pixel, and average that over your validation dataset:

erfs = []

for img, mask in val_loader:
  img = img.to(device)
  img.requires_grad = True
  out = model(img)
  _, _, H, W = out.shape
  center = out[:, :, H//2, W//2].mean()
  center.backward()
  erfs.append(img.grad.abs().mean(dim=[1,2]))

erf = torch.mean(torch.stack(erfs, dim=0), dim=0)

Written from memory - I am not on the server to double-check the accuracy of the above code. Let me know if you have any additional questions.

from hed-unet.

Thanks for your reply! I got this idea and understood the idea of ERF after your teaching. I have another question about how to get the figure of edge like fig 5. In other words, how to detect the edge from the segmentation results? Can you show me the code?

from hed-unet.

The model directly generates an edge detection map, there is no need to derive the edges from the segmentation results. I.e. model(inputs)[batch_idx, 0] will be the segmentation prediction and model(inputs)[batch_idx, 1] is the edge prediction.

The function used for the plots in fig. 5 is literally the first function in train.py:

from hed-unet.

from hed-unet.