Xuefeng Xu

'-sr' means poisson sampling rate, but the code use '-b' batch size, so just change -sr=0.004 to -b=240, for example.

I'm ok with that. May be we can finish the premature overflow issue in `special.boxcox` then go to YJ. Here is what I thought for the premature overflow issue. 1....

I'm ok to reduce the buffer on ymax if we think it's necessacy.

Some useful lemma from the original YJ paper [1]   [1] [A new family of power transformations to improve normality or symmetry](https://doi.org/10.1093/biomet/87.4.954)

Thanks, I was able to run the FEMNIST demo, but fail in the HPO demo. Could you provide more details about how to resolve the logging module issue? ```bash python...

> ```python > if learning_rate == OPTIMAL: > typw = np.sqrt(1.0 / np.sqrt(alpha)) > # computing eta0, the initial learning rate > initial_eta0 = typw / max(1.0, loss.dloss(-typw, 1.0)) >...

I think it is because you set ``auto_rescale = False``

The modified formulation loses precision when: 1. $\lambda\approx0$ ```python import mpmath import numpy as np from scipy.special._mptestutils import ( Arg, assert_mpmath_equal, exception_to_nan) np.seterr(over='ignore') def boxcox(x, lmbda): if abs(lmbda) < 1e-14:...

Thanks @mdhaber @dschmitz89 ! I use the new formulation for large values, test passed.

From my experience, you can try build from source. Or run ``python3 setup.py bdist_wheel`` to create a wheel file, then ``pip3 install xx.whl``