Some suggestions for requirement version up & multi-gpu experiment

[yoom618]

I'm not sure if my code modifications will work well with all models in all experimental settings,
but I'd like to suggest three code modifications just in case.

0) Intro

• Right now, I'm using the latest versions of the libraries that are mentioned in requirements.

  einops==0.8.0
  local-attention==1.11.1
  matplotlib==3.10.0
  numpy==2.1.3
  pandas==2.2.3
  patool==3.1.0
  reformer-pytorch==1.4.4
  scikit-learn==1.5.2
  scipy==1.15.1
  sktime==0.35.1
  sympy==1.13.1
  torch==2.6.0
  tqdm==4.67.1
  PyWavelets==1.8.0
  

• There were some problems when using numpy==2.1.3, so I want to share my two modifications.
• Also, I'm using multi GPUs, and I found that the CUDA_VISIBLE_DEVICES is not working properly in the multi-gpu mode.

1) Change np.Inf to np.inf (see commit for details)

• Since np.Inf is removed in numpy>=2.0.0, it should be fixed to np.inf for those who using upper version of numpy.

2) Change data_loader in short term prediction (see commit for details)

• As you guys already know, variable length arrays are not supported now in np.
• I found that adding dtype=np.ndarray can prevent the problem. (reference)
• Also, there should be correction in true variable in validation phase.

3) Setting CUDA_VISIBLE_DEVICES before importing torch (see commit for details)

• Right now, os.["CUDA_VISIBLE_DEVICES"] aren't working properly.
• Also there are some duplicates in device setting between run.py and exp_basic.py.
• So I
  • moved torch-related codes below os setting
  • merged all device-related arguments codes in run.py
  • change args.device_ids and args.device since the numbering changes after os setting.

[wuhaixu2016]

Hi, great suggestions. It will take some time for us to fix the issue that you proposed. We will make it soon.

[frank142857]

Hi, what's the python version do you choose to run the tasks?

[yoom618]

@frank142857 right now i'm using 3.12 version

[iganggang]

thank u very much.