/DIG_MILP

The official implementation of DIG-MILP

Primary LanguagePython

DIG_MILP

The official implementation of DIG-MILP: A Deep Instance Generator for Mixed-Integer Linear Programming with Feasibility Guarantee

Environment Requirement

To implement the code, the environment below is required:

SCIP version 7.0.3
ecole       0.7.3
matplotlib          3.7.2
networkx            3.0
numpy               1.24.4
PySCIPOpt           3.5.0
PyYAML              6.0.1
scikit-learn        1.3.0
scipy               1.10.1
torch               2.0.0+cu117
torch-geometric     2.3.1
torchaudio          2.0.1+cu117
torchvision         0.15.1+cu117
tqdm                4.65.0
wandb               0.15.5

The corresponding CUDA version is 11.7 and CUDNN version is 8500. To install PYSCIPOpt, please first install SCIP.

Code implementation

We include the code implementation for all the four tasks (set covering, combinatorial auctions, CVS and IIS).

Combinatorial auctions (CA) (synthetic)

Data pre-process

  1. run _1_generate_lp_data_mixed.py to generate graphs for combinatorial auctions. Here the generation follows that in Exact Combinatorial Optimization with Graph Convolutional Neural Networks.

  2. run _2_infer_scip_primal.py to get a set of primal solution/slack to the original problems.

  3. run _3_generate_dual.py to get the dual format of linear relaxation to the original problems.

  4. run _4_1_infer_scip_dual_solution.py and _4_2_infer_scip_dual_slack.py to infer a set of solutions / slack variables for the dual format of linear relaxation.

  5. run _5_normalize_nodes.py to normalize the node inputs of the dataset.

  6. run _6_normalize_weight.py to normalize the edge weight in the dataset.

  7. run _7_generate_vcgraph.py to generate the VC-bipartite graph in torch-geometric in-memory dataset.

  8. run_8_normalize_degree.py to normalize the node degrees.

DIG-MILP training and inference

  1. run sh 9_train.sh to train the model. The pre-trained DIG-MILP model is in the train_files.
  2. run sh 10_generate.sh to generate new instances.

downstream task - parameter tuning without sharing data

enter /downstream_parameter/,

  1. run sh 17_infer_log.sh and sh 17_infer_log2.sh to infer the 45 different hyper-parameter seeds for SCIP. run the script for three times and change the target output path to 'time_log0', 'time_log1', and 'time_log2'.

  2. run ``paint_similarity.py``` to paint the Pearson correlation figure and get the score.

  • We include the SCIP solution time for all the four problems in the time_log folder and the corresponding figure in the paint folder.

downstream task - train machine learning models

enter / downstream_train_ml/, 13. run _1_infer_generate_primal.py to get the optimal solution for the generated MILP instances. 14. run _2_generate_train_data.py to generate the training data for downstream task. 15. run _7_generate_val_lp.py and _8_generate_val_data.py to generate the validation data, run _4_generate_test_lp.py amd _5_generate_test_graph.py to generate the testing data. 16. run sh 3_train.sh to train the downstream machine learning model. 17. run _6_test.py to test the model.

  • The implementation of Set covering (SC), CVS and IIS is similar to CA. But there could be some difference in data pre-processing:

CVS and IIS

Data pre-process

To pre-process the data downloaded from MIPLIB2017, we provide the script to pre-process and we also directly include the pre-processed version (in primal_format folder).