Implementation of An Efficient Combinatorial Optimization Model Using Learning-to-Rank Distillation

This is an offiical implementation of our [work](Paper Link should be here) and baselines and reproducible experiments for the paper.

Environment Settings

We assume that you have installed Python using Anaconda. It should be possible to use other Python distributions, but we have not tested. Your environment should be equipped with CUDA.

We use Python, PyTorch with CUDA as the backend for our ML implementations.

• Python >= 3.6
• PyTorch >= 1.8.0
• CUDA >= 10.1

We use the packages as listed below(alphabetical order):

• Cython == 0.29.22
• Google OR-tools 9.0.9048

External Libraries:

• PyTorch
• fast-soft-sort
  • Follow instruction guide to install this package.
  • CUDA Extension is required.

Installation and how-to-use

---

If you are using Conda,

conda create --name rlrd python=3.6.9
conda activate rlrd

or refer to virtualenv. We divided our implementation into two discrete directories, mdkp/ for MDKP, and gfps/ for GFPS.

1. We assume the working directory to be either mdkp/. or gfps/

2. Install PyTorch with CUDA. For Conda: conda install pytorch==1.8.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch For usual Python: pip install torch==1.8.0+cu101 torchvision==0.6.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html These commands will install appropriate PyTorch with Cuda 10.1 version as the same environment as we test. Please refer https://pytorch.org/get-started/locally/ and https://pytorch.org/get-started/previous-versions/ to see appropriate PyTorch version for your environment.

3. Install dependencies

   pip install -r requirements.txt

4. Dataset Preparation a. MDKP

   • Dataset made on the fly.

   b. GFPS

   • Go to mdkp/create_dataset.py, and set the NUM_PROC_LIST and NUM_TASKS_LIST variables.
   • Run create_dataset.py, then the dataset will be stored in ../gfpsdata/tr, te, val.

5. Global model train.

   For MDKP, 
      num_items: A number of items we should package.
      item_dim: Dimension of weights for each item. Constraints are on each dimension.
      w: Weight of each item is sampled in Uniform([0, w])
      c: A correlation parameter, corresponding to "alpha" in the main paper.
   

   a. MDKP

   • Run train.py --num_items 200 --item_dim 20 --w 300 --c 0. Add --use_cuda if possible.
   • Model will be saved in ./mdkpmodels/globalrl

   b. GFPS

   • Run train.py --num_tasks 32 --num_procs 4, or according to your dataset created.
   • Model will be saved in ./gfpsmodels/globalrl

6. Model train. Each model will be trained by tuning the global model. a. MDKP

   • Run localtrain.py --num_items 50 --item_dim 3 --w 200 --c 0. Add --use_cuda if possible.
   • Model will be saved in ../mdkpmodels/localrl

   b. GFPS

   • Run localtrain.py --num_tasks 32 --num_procs 4 --range_r 3.00 --range_l 3.00. Please set range_r = range_l.
   • Model will be saved in ../gfpsmodels/localrl

7. Distillation a. MDKP

   • Run distillation.py --num_items 50 --item_dim 3 --w 200 --c 0.
   • Each parameters num_items, item_dim, w, and c must be trained on localRL by the previous step 6.
   • Model will be saved in ../mdkpmodels/distillation

   b. GFPS

   • Run distillation.py --num_tasks 32 --num_procs 4 --range_r 3.00 --range_l 3.00. Please set range_r = range_l.
   • Each parameters num_tasks, num_procs and range must be trained on localRL by the previous step 6.

Evaluation:

a. MDKP

• Run test.py --num_items 50 --item_dim 3 --w 200 --c 0.
• You can see ratio of performance and inference time.

b. GFPS

• Run test.py --num_tasks 32 --num_procs 4 --range_r 3.00 --range_l 3.00. Please set range_r = range_l.
• You can see ratio of performance and inference time.