Normalization of features, batch-wise training, feature extraction

[giannispic]

Hello Thomas,
My question is about the normalization of the features of the nodes. In your paper and in your official site you only mention the normalization of the adjacency matrix A but not that of the feature array X. Thanks in advance.

Also about batchwise training. I want to feed the network with OpenPose landmarks of shape (25, 2). Can i perform regular batchwise training just by giving the input X of shape (B, 25, 2) and adj matrix (B, 25, 25) where B is the batch size, or do i necessarily have to follow the sparse block diagonal approach?

Also, in your implementation you classify each node. Can i use the output features of the last gcn layer (flatten it) and feed it for insrtance to a MLP to get a complete instance feature represantation based on which i can perform my tasks??

Best regards,
John.