Open jasonrute opened 1 year ago
The issue with TF-GNN is related to https://github.com/tensorflow/gnn/pull/96 , which removed tfgnn.keras.layers.TotalSize
and introduced tfgnn.keras.layers.MakeEmptyFeature
. Now it should be possible (and actually easier) to create the empty hidden_state
feature for the tfgnn.GraphTensor
context, where the node hidden states are later pooled at the end of message-passing.
Currently, our code only works with tensorflow < 2.10 and tensorflow_gnn <0.3. Here are the specific issues with upgrading.
tensorflow_gnn==0.3
Causes an issue in prediction with the tf_gnn model. In particular, to reproduce:
pip install tensorflow_gnn==0.3
g2t-server --arch tfgnn ...
(no need to interact with Coq)in line https://github.com/IBM/graph2tac/blob/1e394a76e805ba866a60a757c774da4715789d5c/graph2tac/tfgnn/models.py#L258
tensorflow==2.10
Causes an issue when training the TF-GNN model. To reproduce:
pip install tensorflow==2.10
at line https://github.com/IBM/graph2tac/blob/1e394a76e805ba866a60a757c774da4715789d5c/graph2tac/tfgnn/train.py#L409-L411