about a lightweight variant of the ExU-Net

[cjh-yan]

Hello, I saw that ExU-Net-L was mentioned in the article. But it seems that it is not given in the code you gave. Can you tell me what changes have been made based on ExU_Net?

[wngh1187]

Hello, thank you for your interest in my research.
I apologize for not providing information about the light version of ExU-Net mentioned in the paper.
In ExU-Net's argumet.py, I changed the parameters to the following and left the rest unchanged:
l_channel: [16, 32, 64, 128] -> [16, 32, 40, 64]
l_num_convblocks: [3, 4, 6, 3] -> [3, 3, 3, 3]
I will upload the parameters for the trained light version of ExU-Net as soon as possible.

[wngh1187]

We have uploaded the weights of the trained ExU-Net-L.
After defining the model and DDP, you can load the model weights based on the following code to get the same results as in the paper.

checkpoint = torch.load(args['path_bestmodel'], map_location=args['device'])
new_dict = {}
for o_key, n_key in zip(list(checkpoint.keys()), list(model.state_dict().keys())):
    new_dict[n_key] = checkpoint[o_key]
model.load_state_dict(new_dict)