Open erfanloghmani opened 3 months ago
I performed a workaround on the pandas requirement issue by forking the torch_geometric_temporal
repo and removing the requirement because it seems the code does not use pandas directly, as mentioned here.
I also added other benchmarks from the Durendal repo. These models are supported:
Only one method needs to be added if we want, which is TNTComplEx
To run the code:
python snapshot_based.py --seed {seed} --model {model_name}
, where you can use one of the model names listed above
Add DURENDAL (8c2176b commit) method with Update-Then-Aggregate schema.
There is only one challenge with the requirements. The DURENDAL code requires
torch_geometric_temporal==0.54.0
which needspandas==1.3.5
, however, our pandas version in TGB ispandas==1.5.3
. Downgrading to 1.3.5 will cause problem with reading pickle fileml_{data}.pkl
because the file is created using pandas 1.5.3. My workaround to test the code was to export the file using pandas 1.3.5. We can also try using pd.read_pickle as mentioned here. I didn't change the code in the base to avoid any issues with other parts.Now that I have the code structure, I think adopting other methods and baselines from DURENDAL code base should be relatively easier. I will try to work on it if I got time.