LeiBAI/AGCRN

The data scale in trainer isn't consistent

Factos001 opened this issue · 3 comments

Factos001 commented

Hi,

I found that in BasicTrainer.py, the scale of the model's outputs and labels isn't consistent when setting args.real_value=True. You may need to inverse transform the model's outputs to make it right.

jiangxinke commented

Hi,

I found that in BasicTrainer.py, the scale of the model's outputs and labels isn't consistent when setting args.real_value=True. You may need to inverse transform the model's outputs to make it right.

Yeap, I also found this, you will find AGCRN model's label's scaler is really high while the output's is smaller. He forgets to inverse transform.

jiangxinke commented

I still wonder why so many papers in Spatio-Temp's mae and rmse are all the same scaler !!! They use the real data to compare with the normalize data and got this answer, wtf??

LeiBAI commented

Hi,

I found that in BasicTrainer.py, the scale of the model's outputs and labels isn't consistent when setting args.real_value=True. You may need to inverse transform the model's outputs to make it right.

Thanks for the comment. When setting args.real_value=True, the output is the real value, not the normalized value. The corresponding codes are Line 47, 74, and 182 in BasicTrainer.py.