Hi,
I cloned your repository with python 3.7, as recommended by the documentation in a conda virtual environment. I installed the dependencies and tried to run the quickstart model example on the METR_LA dataset. I ran into some issues that I believe should get your attention.
This is the command:
python run_model.py --task traffic_state_pred --model GRU --dataset METR_LA
The issues:
- Ray tuning is specified in the requirements, but the version specified only works with protobuf<4.x. Therefore, one must downgrade protobuf version in order for ray tuning to work. I recommend adding protobuf to the requirements and fixating its version to 3.20.3, for instance.
- Conda supports openssl 1.1.0 (maybe associated with a python 3.7 environment). Urllib only works if its version is < 2.x. I also recommend fixating a version in the requirements, like 1.26.15.
- It seems to me that the device in the config parser is hardcoded to always be True. I had to change the iniit_device() method in order to run the command on a laptop without GPU. I recommend adding the device to the arguments or reviewing the conditional statement where the GPU use is always considered as True.
- There are libraries needed to run the GRU model that are not supported by default, such as: einops, infomap, torch-geometric, pywavelets.
- Torch.profiler is not working in the recommended version of torch 1.7.1. The problem is fixed though when upgrading to version 1.9.0.
Libcity is the only unified framework I found in the field of traffic prediction with a code-based abstraction ready for implementation, with standardization of data formats, implemented models, and customization capabilities. Your continuous support to the library would facilitate breakthroughs in the field of traffic prediction. Clearly, there are a lot of issues in 2026 for the usage of this library. I would appreciate if these issues can be considered and fixed, such as counting on a support for the library by the authors.
Hi,
I cloned your repository with python 3.7, as recommended by the documentation in a conda virtual environment. I installed the dependencies and tried to run the quickstart model example on the METR_LA dataset. I ran into some issues that I believe should get your attention.
This is the command:
python run_model.py --task traffic_state_pred --model GRU --dataset METR_LA
The issues:
Libcity is the only unified framework I found in the field of traffic prediction with a code-based abstraction ready for implementation, with standardization of data formats, implemented models, and customization capabilities. Your continuous support to the library would facilitate breakthroughs in the field of traffic prediction. Clearly, there are a lot of issues in 2026 for the usage of this library. I would appreciate if these issues can be considered and fixed, such as counting on a support for the library by the authors.