Confusion about Property Optimization

[Zhoushun2021]

Hello, I have a question in train a time-dependent energy prediction model. In the file energy_model.py, the loss_energy is about a property in the dataset, i.e. batch.y, for example it represents the formation energy. But why is the input to self.decoder batch.batch?

[jiaor17]

Hi,
In Pytorch Geometric, batch.batch is a tensor that stores the batch indices for each node and helps keep track of which nodes belong to which original graph, you can see the source codes here for more details. Usually, we use this tensor for scatter-based operations in GNNs, like a final pooling to aggregate graph features from node features.

[Zhoushun2021]

Thank you, does this loss_energy represent the loss of a property, such as formation energy. And does pred_e stand for the formation energy obtained from the prediction? Also, I have encountered the following problem during the training process, what should I do to solve @jiaor17