LocalGConv shape mismatch

[kitbransby]

First of all thank you so much for writing this all in Pytorch. I was about to do the same thing myself and you saved me a lot of trouble.

I'm running into an error in the LocalGConv layer and I'm hoping you can help.

features = torch.rand((1, 2000, 43, 1500)) # bs=1, num_pts=2000, num_hyp=43, feature_vec = 1500
adj = torch.from_numpy(np.load('..')) # size (2000, 2000)
layer = LocalGConv(in_features=1500, out_features=128, adj=adj)
output = layer(features)

however, I get the following error due to incorrect shapes in the matmul operation as [2000,2000] @ [1,2000,43,128] is not allowed.

RuntimeError                              Traceback (most recent call last)
Input In [56], in <cell line: 1>()
----> 1 layer(features)

File ~/anaconda3/envs/torch_1.11/lib/python3.9/site-packages/torch/nn/modules/module.py:1110, in Module._call_impl(self, *input, **kwargs)
   1106 # If we don't have any hooks, we want to skip the rest of the logic in
   1107 # this function, and just call forward.
   1108 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1109         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1110     return forward_call(*input, **kwargs)
   1111 # Do not call functions when jit is used
   1112 full_backward_hooks, non_full_backward_hooks = [], []

Input In [53], in LocalGConv.forward(self, inputs)
     49 support_loop_reshape = support_loop.reshape(batch_size, num_pt, num_hypo, -1)
---> 50 output = (self.adj_mat @ support_reshape) + support_loop_reshape
     51 if self.bias is not None:
     52     ret = output + self.bias

RuntimeError: mat1 and mat2 shapes cannot be multiplied (256000x43 and 2000x2000)

Please let me know if you have any idea of what the fix is

[kitbransby]

Ah I see my error. Adj matrix is (43,43) for the hypothesis deformation not (num_pts, num_pts) like in Pixel2Mesh