statsmodels 的 cuda 替代？

otonoco · 2025 年3 月 17 日 03:01

rt
wondering if we have a pytorch version of statsmodels?

RandomTrash · 2025 年3 月 17 日 03:13

您是否在寻找 jax

otonoco · 2025 年3 月 17 日 03:16

I don’t think Jax includes the regression models & tsa methods??

问的原因主要是不想自己写一个。。。

RandomTrash · 2025 年3 月 17 日 03:18

但是可以套皮然后拿进去call。。。

Wi-Fi · 2025 年3 月 17 日 03:47

github.com/statsmodels/statsmodels

ENH: speeding up model training by moving it to GPU?

已打开 02:37PM - 19 Jun 18 UTC

vss888

I am searching for a way to speed up model training since we need to train many …of them (i.e. dozens of millions). To profile things up, I took an example model, which takes 190 seconds to train. Out of those 75% are spent in kalman_filter.py:733(_filter) and the next 17% are spent in numpy.core.multiarray.dot. Looking at statsmodels/tsa/kalmanf/kalman_loglike.pyx, I can see that the code is already heavily optimized for CPU by using Fortran BLAS libraries. So one idea is to move at least the two functions to run on GPUs instead of CPUs. Please, let me know if you have a feeling of much it could help and which Python CUDA library (CuPy, Numba, MinPy, PyTorch, TensorFlow, CULA, scikit-cuda, PyCUDA, PyOpenCL, pyculib, theano, etc) would be a good choice to proceed? Thank you for your help!

This was back in 2018, I guess the devs dones’t care that much about GPUs.

这年头AI写代码也很快，你要不要仿照dask-ml搞一个custatsmudels？把底层numpy无缝换乘cuPy就行