Closed yuedajiong closed 6 months ago
maybe you are right. In GS, the feature/color, the valid range is from 0 to 1, you can double check it. and the author split it into 2 parts, the only different is learning-rate. I asked author last year, he confirmed. if no any other activation-function control, it is easy be out of range.
great work, I am trying stage1.
We are working to solve this problem. It seems to be caused by a learning rate that is too large for the Gaussian features (starting at 0.05 and decreasing to 0.01). We tried using a constant rate of 0.01, and the phenomenon of colorful patterns weakened.
We need to test on more cases. The demo results and configuration files will be updated once we make progress.