KIDA: Knowledge Inheritance in Data Aggregation

This project releases our 1st place solution on NeurIPS2021 ML4CO Dual Task.

Slide and model weights are available.

Environment Setup

Follow the tutorials of ml4co-competition to download ML4CO instance dataset and set up your Python dependencies. Suppose the directory of instance dataset is /YOUR_PATH/ml4co-competition/instances. The dependencies of KIDA can be found in ./Nuri/conda.yaml . Anaconda environment of ml4co can be created by source init.sh

Testing

Download model weights and put them in /YOUR_PATH/NeurIPS2021-ML4CO-KIDA/Nuri/ .

Copy the submission files to ml4co-competition directory.

cp -r /YOUR_PATH/NeurIPS2021-ML4CO-KIDA/Nuri/ /YOUR_PATH/ml4co-competition/submissions/

Evaluate models in dual task by running the following commands.

cd /YOUR_PATH/ml4co-competition/submissions/Nuri
conda activate ml4co
python ../../common/evaluate.py dual item_placement
python ../../common/evaluate.py dual load_balancing
python ../../common/evaluate.py dual anonymous

Training

Modify the configuration file in ./train_files/configs to meet the needs of the local environment.

./train_files/configs/dataset.ini is the configuration file for data generation process. DATASET_DIR decides the directory where the instance dataset is stored. STORE_DIR decides the directory where generated data is stored. NODE_RECORD_PROB decides the probability of using Strong Branching when collecting data. TIME_LIMIT decides the time limit for SCIP solver. POLICY_TYPE decides the model we use (0 for Item Placement Benchmark, 1 for Workload Apportionment Benchmark and 2 for Anonymous Benchmark).

./train_files/configs/train.ini is the configuration file for training process. STORE_DIR decides the directory where training files are stored (It must keep the same as STORE_DIR in dataset.ini). TRAIN_NUM and VALID_NUM decide the number of data used in training process for each epoch.

Then, run the following commands to train the model.

cd /YOUR_PATH/NeurIPS2021-ML4CO-KIDA/train_files
# Train Item Placement Benchmark
source item.sh
# Train Workload Apportionment Benchmark 
source load.sh
# Train Anonymous Benchmark
source ano.sh

Resouces

参赛方案介绍 - 旷视研究院