For example, with respect to edge convolution, a linear translation + self.b_v and a non-linear operation F.relu() occur at line 50 in hypergraph.py before normalisation in lines 58 and 60. In fact, even the pooling of messages from neighbours (i.e., matrix multiplication the with incidence matrix A) occurs afterwards at line 59.
Similarly, with respect to vertex convolution at line 62 in the same function.
This is different to the order of computation in Algorithm 1 found in the paper, which states that normalisation is performed prior to linear translation and nonlinear operation.