I’ve recently started to experiment with Pytorch Geometric Temporal. I’m trying to train a model consisting of recurrent graph neural network layers (e.g. GConvGRU or GConvLSTM) on the MTM Dataset with the purpose of eventually performing sequence continuation. I’m now stuck with what seems my incomprehension how to properly configure these layers for specific shapes of the input features.
At the moment, the constructor of the layer is called with an input channel count of 3 (corresponding to the dimension of the joint positions), an arbitrary output channel count such as 16, and a similarily arbitrary filter size of 2.
It seems that no matter how a prepare the shape of the input features, whenever I call the forward function of this layer, I receive an error message that either points to a mismatch between matrices or an index that is out of bounds.
When using a feature shape of 3, 21, 128 (which is what the dataset returns by default when obtained for 128 frames), the error message is as follows:
File “C:\Users\dbisig\anaconda3\envs\pytorch112_geomtemp\lib\site-packages\torch_geometric\nn\dense\linear.py”, line 118, in forward
return F.linear(x, self.weight, self.bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (63x128 and 3x16)
When using a feature shape of 21, 3, 128 (by moving axis 0 and 1) the error message is as follows:
File “C:\Users\dbisig\anaconda3\envs\pytorch112_geomtemp\lib\site-packages\torch_scatter\scatter.py”, line 21, in scatter_sum
return out.scatter_add_(dim, index, src)
RuntimeError: index 3 is out of bounds for dimension 0 with size 3
Can somebody explain to me how to do this properly? Or point me to example code that works with the MTM Dataset?
Thanks a lot and best regards
Daniel