search

tsinghua-fib-lab / DRL-urban-planning

A deep reinforcement learning (DRL) based approach for spatial layout of land use and roads in urban communities. (Nature Computational Science)
https://rdcu.be/dlRPZ
MIT License
203 stars 40 forks source link

Question about the model transferability #5

Closed Sabrina2040 closed 1 year ago

Sabrina2040 commented 1 year ago

The overall idea of the code is also very clear. But I have one question about the embodiment of model transferability within the code. I don't seem to find any code that embodies the training process related to model transferability? Or is it that different types of community environments can be switched directly without making other types of modifications, such as dimensions of action outputs, etc. I envision that the number of edges and points in the graphs they compose are different for different sized communities, so there needs to be some code additions(either fine-tuning or re-training) in the current version of the repo.

DavyMorgan commented 1 year ago

Thanks, we use finetuning to achieve transfer across different communities.

Specifically, we pretrain a model on a 3x3 grid community, and finetune the model on other communities. To achieve the finetuning, you just need to change the cfg file and set the iteration cmd param to a specified iteration. Please refer to the flags in train,py: flags.DEFINE_string('iteration', '0', 'The start iteration. Can be number or best. If not 0, the agent will load from ' 'a saved checkpoint.')