Hello, I was looking over the Tensorflow and Pytorch versions of the CPD model and realized they are inconsistent.
In Tensorflow version when we are converting input to the features there is an extra GVP layer. Here is the node embedding and edge embedding functions defined under StructuralFeatures which is called here. Right after this we apply another set of GVP layers for both edge and vertex embeddings.
Meanwhile in Pytorch version we are missing the GVP layers under StructuralFeatures. Instead edge and vertex features only normalize while converting the inputs to features under ProteinGraphDataset. Then just like in Tensorflow version GVP layers applied to normalized vertex and edge features.
Can you please clarify this issue?
Hello, I was looking over the Tensorflow and Pytorch versions of the CPD model and realized they are inconsistent.
In Tensorflow version when we are converting input to the features there is an extra GVP layer. Here is the node embedding and edge embedding functions defined under StructuralFeatures which is called here. Right after this we apply another set of GVP layers for both edge and vertex embeddings.
Meanwhile in Pytorch version we are missing the GVP layers under StructuralFeatures. Instead edge and vertex features only normalize while converting the inputs to features under ProteinGraphDataset. Then just like in Tensorflow version GVP layers applied to normalized vertex and edge features.
Can you please clarify this issue?