change_summary.txt

----- Bugs and critical changes -----

1. Summary node. For a graph with N nodes {0, ... N-1}, add a new summary node N and connect all nodes to it. 
1.1 Edges used to be added to N-1 rather than N. Fixed
1.2 Edges used to be single directional, so no message passing happened. Need to add both edge (i, j) and (j, i). Fixed.
1.3 Generate two different datasets for graphs with/without summary nodes. Otherwise, when args.summary = False, the training step didn't work for dataset with summary node. Fixed.

2. get_device()
When device == 'cpu' Tensor.get_device() returns -1 and causes device issue, changed to Tensor.device instead.

----- Useful changes -----

3. RelEncoding -> MultiDimRelEncoding (more general)
Each RelEncoding now can be multidimensional, e.g. hierarchical shortest distance (hsd) has several hierarchies.

4. pre_process
4.1. Rewrote function to include both original and augmented edge_idex/edge_attr. Could be useful for visualization and debugging.
4.2 Split it two versions as mentioned in 1.3 

5. Weighted sampler
Added another option to use weighted data sampler when creating dataloader. Maybe helpful for unbalanced tasks.

----- Non-essential changes -----
6. Overfitting happend too quickly. Reduce num_epochs from 100 to 60. Increase it if no longer see overfitting.

7. Add a chuck for plotting at the end of training, so no need to create a new Tensorboard record each time if you don't want to. 

8. Moved get_optimizer to utils 

9. Move ModifiedAtomEncoder & ModifiedBondEncoder from utils.py to graph_transformer.py, so changes in jupyter notebook can be made more easily.

10. Cleaned up code a little bit in all files.