ENH/BUG: Allow multinomial to check pvals with other float types #16732
Conversation
b6243eb to
66ac658
Compare
|
The condition is kind of a mouthful, so may not be the best approach. The alternative would be to use a fused type for kahan_sum which would then be checked also using a fused type for the comparison value. |
66ac658 to
a8f6a0a
Compare
|
The two commits use different approaches.
Are either of these reasonable ways to handle it? The third option, which I haven't implemented would be
|
|
My first impression is that the last version looks slightly nicer than the previous. Does the |
No. It is just 1.0. Could use the value of np.nextafter(np.float32(1.0), np.float32(2.0)).
Is last here 3 or 2? |
|
Oh, yeah, I was thinking about option 3, which I wouldn't mind if there was some simple solution to give to the user running into the problem... |
2244453 to
bce9a67
Compare
|
I've not included the 3rd version in the 3rd commit. It is easily the simplest. |
bce9a67 to
95b0a34
Compare
Add additional check when original input is an array that does not have dtype double closes numpy#8317
Use a fused type and a kahan sum to check condition
Improve error message when the sum of pvals is larger than 1 when the input data is an ndarray
95b0a34 to
359d04f
Compare
Improve error message when the sum of pvals is larger than 1 when the input data is an ndarray closes numpy#8317 xref numpy#16732
Add additional check when original input is an array that does not have dtype double
closes #8317