Small question about Xnet vs Nestnet difference

[CozyDoomer]

The only architectural difference between the two seems to be this:

• nestnet
  skip=interm[(n_upsample_blocks+1)*i+j]
• xnet
  skip=interm[(n_upsample_blocks+1)*i: (n_upsample_blocks+1)*i+j+1]

in the upblock parameters.

can someone confirm this and maybe quickly explain how this affects the model?
would they both count as Unet++ architecture according to the paper but with different skip connections?

[zsk-tech]

I think xnet is the real UNet++. nestnet doesn't connect current layer to all previous layers . Therefore，the accuracy of the nestnet is not as high as xnet.
I think nestnet is wrong. And xnet is a modified version of nestnet.

And I think there is also something wrong with xnet,even if I'm not sure.
For example,In xnet , the author reverse its order when using decoder_filters=[256 128 64 32 16].
So the number of filters the author wants to use should be [16 32 64 128 256].
But the number of filters the author actually uses is [32 64 128 256 16].
And I also found other things（I think it‘s wrong） that I don’t understand very well.

[CozyDoomer]

Hey @zsk-tech thank you for your input.

I agree that the connections are different..
but that does not mean it is necessarily 'wrong' it could just be a version of Unet++ with less skip connections but still working correctly right?

What we can say is it makes a difference in the number of trainable parameters.
E.g. for resnet18:

# xnet resnet18
Total params: 18,273,898
Trainable params: 18,261,156
Non-trainable params: 12,742

# nestnet resnet18
Total params: 17,462,890
Trainable params: 17,450,148
Non-trainable params: 12,742

I was also confused by the filters of the additional decoder skip connections.

Where did you see the author using [32 64 128 256 16] number of filters?
For me when I look up the filters of the layers that are skip connections it looks like this for resnet18:

stage4_unit1_relu1 (Activation) (None, 16, 16, 256)
stage3_unit1_relu1 (Activation) (None, 32, 32, 128)
stage2_unit1_relu1 (Activation) (None, 64, 64, 64)
relu0 (Activation)              (None, 128, 128, 64)
relu1 (Activation)              (None, 8, 8, 512)
stage3_unit2_relu1 (Activation) (None, 16, 16, 256)
stage2_unit2_relu1 (Activation) (None, 32, 32, 128)
stage1_unit2_relu1 (Activation) (None, 64, 64, 64)

But maybe you looked into what actually happens in build_xnet and that's why I'm confused about that statement.

[zsk-tech]

Thank you for your explanation.
I think what you said is right. Xnet and nestnet are just two different versions.

Q: Where did you see the author using [32 64 128 256 16] number of filters?

A: I derived it from the code in build_xnet.But it is also possible that my understanding of the code is incorrect.

[CozyDoomer]

Okay that makes sense, I'll try to look at it when I find the time and report back!

The build_xnet function it's not easy to unpack sadly, maybe the authors find time to answer in the meantime :)

[zsk-tech]

OKay