search

yccyenchicheng / SDFusion

MIT License
404 stars 34 forks source link

stride=(1, 2, 2) in Downsample() and Upsample()? #30

Open TrepangCat opened 1 year ago

TrepangCat commented 1 year ago

Appreciate your great work! When I try to run your text2shape model, I notice that stride=(1, 2, 2) in Downsample() and Upsample(). Could you tell me more reasons why you chose stride=(1, 2, 2)? Why not stride=(2, 2, 2)?

liopeer commented 5 months ago

I just realized that as well and came here to see if someone already raised the issue. I will probably switch this to (2, 2, 2) since I expect the (1, 2, 2) stride to bias the network towards not treating the axes all the same, which is not what I want in a shift-equivariant network like a CNN.

weitunglin commented 1 week ago

Same here. Is there anyone who knows why SDFusion sets stride to (1, 2, 2) and why it works?

liopeer commented 1 week ago

Same here. Is there anyone who knows why SDFusion sets stride to (1, 2, 2) and why it works?

@weitunglin I am 95% sure that the code is simply buggy and this is undesired. There is no reason why they would want to have a non-cube-like latent.