DeepGraphLearning
CUDA index out of bounds
When I executed the 'inference.py' script, I encountered a CUDA error. The error message is displayed in the following image:
This error occurred immediately after the script completed executing line 121 and began processing line 24. Any ideas for this bug?
Hi @BL-Lac149597870, this error seems to indicate the it is attempting to access an element of a tensor or array at an index that does not exist. Have you checked the the maximum value of protein.atom2graph and the shape of sigma?
The error occured before the script executing line 124. It appears there may be an issue with the graph_construction_model. Removing the SpatialEdge and KNNEdge classes from the train.yaml file resolves the previously mentioned error.
Hi @BL-Lac149597870, can you locate which pdb cause this issue and share the file?
Hi @zytzrh ,I've reviewed the source code of TorchDrug and noticed an issue related to the computation of the explicit_valence property. This can be found in the TorchDrug repository here: https://github.com/DeepGraphLearning/torchdrug/blob/6066fbd82360abb5f270cba1eca560af01b8cc90/torchdrug/data/molecule.py#L511C5-L514C47. It's important to note that this doesn't affect the operation of DiffPack. Regarding the graph construction in DiffPack, observed here:https://github.com/DeepGraphLearning/DiffPack/blob/81ee51612abfab5af23a8d33fbb081a024f83a1c/diffpack/task.py#L116C1-L121C61 , it appears that additional edge types are incorporated into the edge_list . However, bond2valence lacks corresponding values for these new edge types.