conversion.md

Some ideas behind the conversion from TensorFlow to pytorch:

• We replace Sequential of Tensorflow by nn.Module in pytorch, where the layers are coordinated in forward
• We replace Spektral's GCNConv with torch_geometric's GCNConv for graphs
• We replace TensorFlow's GlobalAveragePooling1D by torch.mean over the time dimension

Some other logics have been updated (to document)