speedinghzl / alignseg Goto Github PK
View Code? Open in Web Editor NEWAlignSeg: Feature-Aligned Segmentation Networks (TPAMI 2021)
License: MIT License
AlignSeg: Feature-Aligned Segmentation Networks (TPAMI 2021)
License: MIT License
Dear author:
Thanks for your excellent work! When I tried to follow the AlignFA module, I found that the bilinear_interpolate_torch_gridsample
function didn't sample the features as I expected before. In one word, I think the normalization factors for the vertical and horizontal coordinates are reversed.
And I conduct a toy experiment to show this possible problem in the function:
For a tensor of shape (1, 1, 4, 8) like:
tensor([[[[ 0., 1., 2., 3., 4., 5., 6., 7.],
[ 8., 9., 10., 11., 12., 13., 14., 15.],
[16., 17., 18., 19., 20., 21., 22., 23.],
[24., 25., 26., 27., 28., 29., 30., 31.]]]])
when I want to sample the value from the pixel under the source pixel, I 'predict' a delta tensor (of shape (1, 2, 4, 8)) like:
tensor([[[[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0]],
[[1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1, 1],
[1, 1, 1, 1, 1, 1, 1, 1]]]])
And the expected output is:
tensor([[[[ 8.0000, 9.0000, 10.0000, 11.0000, 12.0000, 13.0000, 14.0000, 15.0000],
[16.0000, 17.0000, 18.0000, 19.0000, 20.0000, 21.0000, 22.0000, 23.0000],
[24.0000, 25.0000, 26.0000, 27.0000, 28.0000, 29.0000, 30.0000, 31.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]]]])
But when I fed these into the bilinear_interpolate_torch_gridsample
, I got:
tensor([[[[ 1.5000, 2.5000, 3.5000, 4.5000, 5.5000, 6.5000, 7.5000, 8.5000],
[ 9.5000, 10.5000, 11.5000, 12.5000, 13.5000, 14.5000, 15.5000, 16.5000],
[17.5000, 18.5000, 19.5000, 20.5000, 21.5000, 22.5000, 23.5000, 24.5000],
[19.5000, 20.3125, 21.1250, 21.9375, 22.7500, 23.5625, 24.3750, 25.1875]]]])
It is noted that I had set align_corners=True
, which is the default behavior up to Pytorch=1.2.0, since I use the Pytorch=1.7.0. I think it is because of a wrong normalization, specifically, the width of the tensor is used to normalize the vertical coordinates. The results can be explained by:
\delta_{v} = 1 / (w / s) * (h-1) / 2 = 1 / (8 / 1) * (4 - 1) / 2 = 1.5 / 8
So if we ignore the scale factor s
and the difference between (w, h) and (w-1, h-1), a more reasonable function is like:
def bilinear_interpolate_torch_gridsample_new(input, size, delta=0):
out_h, out_w = size
n, c, h, w = input.shape
s = 1.0
norm = torch.tensor([[[[w/s, h/s]]]]).type_as(input).to(input.device) # not [h/s, w/s]
w_list = torch.linspace(-1.0, 1.0, out_h).view(-1, 1).repeat(1, out_w)
h_list = torch.linspace(-1.0, 1.0, out_w).repeat(out_h, 1)
grid = torch.cat((h_list.unsqueeze(2), w_list.unsqueeze(2)), 2)
grid = grid.repeat(n, 1, 1, 1).type_as(input).to(input.device)
grid = grid + delta.permute(0, 2, 3, 1) / norm
output = F.grid_sample(input, grid, align_corners=True)
return output
and the output of this function is:
tensor([[[[ 3.0000, 4.0000, 5.0000, 6.0000, 7.0000, 8.0000, 9.0000, 10.0000],
[11.0000, 12.0000, 13.0000, 14.0000, 15.0000, 16.0000, 17.0000, 18.0000],
[19.0000, 20.0000, 21.0000, 22.0000, 23.0000, 24.0000, 25.0000, 26.0000],
[15.0000, 15.6250, 16.2500, 16.8750, 17.5000, 18.1250, 18.7500, 19.3750]]]])
So is it a special design (because the final prediction results seem good using your raw code), a problem caused by the version, or a real problem?
Dear author:
Your work is excellent! I wonder when to release the code. Looking forward to your work!
Hi
Could you please tell me which Pytorch version you use.
Thanks
in alignseg.py
I meet
AttributeError: module 'torch.distributed' has no attribute '_all_gather_base'
which is possibly caused by the torch version problem.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.