Reproduce the imitation result.

[Jarvis-K]

1. In what format should I prepare the collecting data? Can u provide the imitation data collection script?
2. Can u provide the imitation data in one road to help a quick reproduction?

[Gungnir2099]

You can run run_chatgpt.py to collect the fine-tuning data. It will be stored in ./GPT_logs.

[Gungnir2099]

The file is named XXX_state_action_prompt_commonsense_no_calculation.json.

[DA21S321D]

Since the simulation time is 3600 for anon_4_4_hangzhou_real.json, how many rounds are needed to collect data from gpt4?

[Gungnir2099]

The duration of the green signal phase is 30s. So, the collection needs about 120 rounds.

[Gungnir2099]

In response to your concern, we provide the budget (in USD) for training LightGPT (with A800 GPUs) on the Jinan 1 dataset.

	Data Collection	Imitation Fine-tuning (1 GPU)	Critic-Guided Policy Refinement (4 GPUs)	Sum
Cost	$135.53	$58.00	$5.52	$199.05

[DA21S321D]

Could you share just one map data collected with gpt? So that we don't have to run it again for the same thing. Thanks

[Jarvis-K]

You can run run_chatgpt.py to collect the fine-tuning data. It will be stored in ./GPT_logs.

The format of GPT_logs/xxx.json is not compatible with imitation learning. How can I transform the {state, state_incoming, prompt, action} to {instruction, output} which is required for imitation finetuning.

[Gungnir2099]

Could you share just one map data collected with gpt? So that we don't have to run it again for the same thing. Thanks

I'm sorry, but our collected fine-tuning data for LLMLight isn't publicly available at this time because we're constantly refining it. However, the LightGPT trained on the Jinan 1 dataset is accessible. You're welcome to use its responses and gather data for additional fine-tuning purposes.

[Gungnir2099]

You can run run_chatgpt.py to collect the fine-tuning data. It will be stored in ./GPT_logs.

The format of GPT_logs/xxx.json is not compatible with imitation learning. How can I transform the {state, state_incoming, prompt, action} to {instruction, output} which is required for imitation finetuning.

I have included the code for compiling {state, state_incoming, prompt, action} into {instruction, output} format. You can run python finetune/state_action_2_instructions.py --input_file STATE_ACTION_PROMPT_FILE --output_file INSTRUCTION_FILE to perform the transformation. Thank you for your feedback.

[DA21S321D]

Thanks for your kindness. I also checked this script. Since there are different files for different intersections : Hangzhou-4_4-anon_4_4_hangzhou_real-intersection_1_1-gpt-4-4_state_action_prompt_commonsense_no_calculation.json and Hangzhou-4_4-anon_4_4_hangzhou_real-intersection_1_2-gpt-4-4_state_action_prompt_commonsense_no_calculation.json ..., should i merge their output json (generated by running finetune/state_action_2_instructions.py) together as the instruction file to finetune this? :

python ./finetune/run_imitation_finetune.py --base_model MODEL_PATH 
                                            --data_path instruction file merged 
                                            --output_dir OUTPUT_DIR

You can run run_chatgpt.py to collect the fine-tuning data. It will be stored in ./GPT_logs.

The format of GPT_logs/xxx.json is not compatible with imitation learning. How can I transform the {state, state_incoming, prompt, action} to {instruction, output} which is required for imitation finetuning.

I have included the code for compiling {state, state_incoming, prompt, action} into {instruction, output} format. You can run python finetune/state_action_2_instructions.py --input_file STATE_ACTION_PROMPT_FILE --output_file INSTRUCTION_FILE to perform the transformation. Thank you for your feedback.

[Gungnir2099]

Thanks for your kindness. I also checked this script. Since there are different files for different intersections : Hangzhou-4_4-anon_4_4_hangzhou_real-intersection_1_1-gpt-4-4_state_action_prompt_commonsense_no_calculation.json and Hangzhou-4_4-anon_4_4_hangzhou_real-intersection_1_2-gpt-4-4_state_action_prompt_commonsense_no_calculation.json ..., should i merge their output json (generated by running finetune/state_action_2_instructions.py) together as the instruction file to finetune this? :

python ./finetune/run_imitation_finetune.py --base_model MODEL_PATH 
                                            --data_path instruction file merged 
                                            --output_dir OUTPUT_DIR

You can run run_chatgpt.py to collect the fine-tuning data. It will be stored in ./GPT_logs.

The format of GPT_logs/xxx.json is not compatible with imitation learning. How can I transform the {state, state_incoming, prompt, action} to {instruction, output} which is required for imitation finetuning.

I have included the code for compiling {state, state_incoming, prompt, action} into {instruction, output} format. You can run python finetune/state_action_2_instructions.py --input_file STATE_ACTION_PROMPT_FILE --output_file INSTRUCTION_FILE to perform the transformation. Thank you for your feedback.

Yes. Each file includes the interaction records at an intersection. You can merge them all to build your instruction fine-tuning dataset.

[Jarvis-K]

Thanks for your response! It does address my issues. This issue gonna be closed.