planGPT

KO | EN

gitlite — search

#javascript #python #hacktoberfest #typescript #llm #golang #android #machine-learning #linux #docker #macos #nodejs #awesome #awesome-list

(★ 11)

This repository includes the provided code for the paper Learning General Policies for Planning through GPT Models.

planGPT 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# planGPT -     Learning General Policies for Planning through GPT Models
This repository includes  the provided code for the paper ```Learning General Policies for Planning through GPT Models```[(link)](https://ojs.aaai.org/index.php/ICAPS/article/download/31510/33670).The scripts provided are:
- [scripts/plan_generator.py](scripts/plans_generator.py): script to generate plans given a dataset of planning problems.
- [scripts/validator.ipynb](scripts/validator.ipynb): script to validate the generated plans.
- [scripts/solver_pddl.py](scripts/solver_pddl.py): script to generate a planGPT plan using as input a PDDL problem.
- [scripts/metric.py](scripts/metric.py): script inspired by VAL that given an initial state, goal state, plan and domain verifies if the plan is valid.
- [scripts/model.py](scripts/model.py): script defining the planGPT model.
### Installation
The code is written in Python 3.8.10 and to install the required packages you can run the following command:
```bash 
cd scripts/
pip install -r requirements.txt
```
To note we provide more packages than needed, so it's possible to remove some of them if you have issues feel free to contact us by [email](mailto:massimiliano.tummolo@unibs.it).
In addition, for the dataset and models you will need to download them from the following links:
- *Models*: [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.10908361.svg)](https://doi.org/10.5281/zenodo.10908361)
- *Datasets*: [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.10925404.svg)](https://doi.org/10.5281/zenodo.10925404)

All domains are randomized, except for Floortile and Visitall, where the grid structure is kept fixed. For these two domains, we later conducted additional experiments by translating the grid along horizontal and vertical axes, achieving coverage results above 90%. During this process, we discovered around 2,000 duplicate problems between the training and testing sets due to generator issues. Even before removing these duplicates, the test set without overlap contained approximately 3,500–4,000 problems. After removing the duplicates, we repeated the experiments and obtained identical results, confirming the robustness of our findings. After further checks, no other domains showed any overlap between the training and test problems.

Furthermore, if you want, we provide all the *generations of the plans* in the following link: [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.10926333.svg)](https://doi.org/10.5281/zenodo.10926333) usable to be validated with the script [scripts/validator.ipynb](https://raw.githubusercontent.com/aiunibs/planGPT/main/scripts/validator.ipynb).
## Usage
### Plan Generation
The generation of plans by planGPT can be done using the script [scripts/plan_generator.py](scripts/plans_generator.py). The script requires a configuration file in JSON format. An example of the configuration file is the following:
```json
{
    "dataset_dir": "DATASET_FOLDER",
    "tokenizer_dir": "MODEL_FOLDER",
    "model_dir": "MODEL_FOLDER",
    "output_dir": "WHERE_TO_SAVE_PLANS",
    "num_beams": 1,
    "num_return_sequences": 1,
    "percentage_actions_seen": 0,
    "batch_size": 1,
    "save_after": 100,
    "domain": "DOMAIN_NAME"
}
```
The parameters are:
- *dataset_dir*: the folder containing the dataset.
- *tokenizer_dir*: the folder containing the tokenizer.
- *model_dir*: the folder containing the model.
- *output_dir*: the folder where to save the generated plans.
- *num_beams*: the number of beams to use during the generation. For example, 1 means greedy search, 5 means beam search with 5 beams.
- *num_return_sequences*: the number of sequences to return. For example, 1 means return only the best sequence, 5 means return the 5 best sequences.
- *percentage_actions_seen*: the percentage of actions seen during the generation (0 means no actions seen, 100 means all actions of a correct plan seen).
- *batch_size*: the batch size to use during the generation.
- *save_after*: the number of generations after which save the results.
- *domain*: the domain name of the dataset. For example, blocksworld.

Then after creating the configuration file you can run the script with the following command:
```bash
cd scripts/
python plan_generator.py configuration.json
```
### Plan Validation
The validation of the generated plans can be done using the script [scripts/validator.ipynb](scripts/validator.ipynb). The only modification needed is to set the path of the generated plans in the variable ```folder_data_path```. So you can provide the path of the folder containing the generated plans using a custom script or the generation provided in the link above.

planGPT

planGPT 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

planGPT

planGPT 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README