SJTU CS3308 Machine Learning for Bin Packing Problem Muti-person Collaboration
Group members:
- YINGMING ZHENG
- YIHENG WANG
- KUNYAN LI
This repository retains all the historical codes of our group during the research on the 3D bin packing problem, but not all the codes work well. We have abandoned some ineffective methods. The main code files used are mainly the two folders A_Dynamic_Multi-Modal_Deep_Reinforcement_Learning_Framework_for_3D_Bin_Packing_Problem and DMRL-BPP. The latter contains the implementation codes for Task1 and Task2, while the former contains the implementation code for Task3. Among them, the sub-repository of Task3 has two branches, a main branch and an ablation experiment branch. We've also added folders of Transformer and RL based model for Task1 and Task2, you could use it for analysis or baseline. ML_BPP.pdf contains detailed introduction of our model and performance.
To reproduce this experiment, you may need:
- Partially clone this repository. Not all files are necessary.
- Configure the experimental environment according to the environment.yml file corresponding to the experiment.
- Specifically, for Task 3, you need to prepare the api_key by yourself.
- Run python main.py. By default, for tasks 1 and 2, the model will be retrained from scratch, and it takes about 3 days to train for 40,000 epochs on 4 A10; for Task 3, by default, the best model we trained will be loaded, and it takes 8 to 10 hours to perform a single round of inference on all data using a proxy api_key. Using a direct connection api or loading only a small amount of data can significantly speed up this process.
If you have any other questions, please feel free to contact us.