search

QUVA-Lab / escnn

Equivariant Steerable CNNs Library for Pytorch https://quva-lab.github.io/escnn/
Other
345 stars 45 forks source link

check_equivariance test failed #71

Closed Guptajakala closed 1 year ago

Guptajakala commented 1 year ago 
gspace = gspaces.rot2dOnR2(4)
n_feat = 16
in_type = enn.FieldType(gspace, [gspace.trivial_repr]*3)
out_type = enn.FieldType(gspace, [gspace.regular_repr]*n_feat)

net = enn.R2Conv(in_type, out_type, kernel_size=3, stride=1, padding=1, dilation=1, bias=True, sigma=None)
net.check_equivariance()

I got error AssertionError: The error found during equivariance check with element "10[2pi/12]" is too high: max = 3.4398996829986572, mean = 0.3915270268917084 var =0.21077707409858704

Is this failure expected?

Gabri95 commented 1 year ago

Hi @Guptajakala

Are you sure this is the code you run to generate this error? The group you are choosing in the line gspace = gspaces.rot2dOnR2(4) only contains (multiples of) 90 degrees rotations but the error you report says the model failed at the rotation 5/6 * 2pi which doesn't belong to this group.

Best, Gabriele

Guptajakala commented 1 year ago

I found it was caused by a mistake. Now I run with the right code, it's all good.