Hello, and thank you for releasing this interesting and well-documented work!
I am currently reproducing the results of L2O-pMINLP and have a question regarding the model’s ability to generalize across different problem sizes.
When running the provided scripts (e.g., run_qp.py, run_nlp.py), the --size argument defines the problem dimension (i.e., the number of decision variables).
During training, I use a fixed problem size (e.g., --size 10), and I would like to know whether it is possible to perform inference on a different problem size (e.g., --size 20) using the same trained model.
Specifically:
Can the current implementation (or model architecture) support inference on problem instances whose dimension (--size) differs from the one used during training?
If not, would it require architectural modifications — such as input-layer resizing, shared parameterization, or GNN-style variable handling — to enable this type of cross-size generalization?
Example scenario:
Training: --size 10
Inference: --size 20
Thank you very much for your time and for sharing this excellent project! Any clarification on this point would be greatly appreciated.
Hello, and thank you for releasing this interesting and well-documented work!
I am currently reproducing the results of L2O-pMINLP and have a question regarding the model’s ability to generalize across different problem sizes.
When running the provided scripts (e.g., run_qp.py, run_nlp.py), the --size argument defines the problem dimension (i.e., the number of decision variables).
During training, I use a fixed problem size (e.g., --size 10), and I would like to know whether it is possible to perform inference on a different problem size (e.g., --size 20) using the same trained model.
Specifically:
Can the current implementation (or model architecture) support inference on problem instances whose dimension (--size) differs from the one used during training?
If not, would it require architectural modifications — such as input-layer resizing, shared parameterization, or GNN-style variable handling — to enable this type of cross-size generalization?
Example scenario:
Thank you very much for your time and for sharing this excellent project! Any clarification on this point would be greatly appreciated.