PC fix #164
PC fix #164
Conversation
|
@galapaegos , if you get a chance, please compile with trace on and look at the remaining mismatches. I'm getting some normalize should be 0 errors that probably are an issue, and haven't inspected 3 body yet. |
|
The problem seems to be that the normalize (binned) call uses 3N-element non-incrementing "normalize" values (min, max, and bins for N observables), and this causes the PC increment to either expect 0 increment (no argument case) or 1 increment (argument case). |
import numpy as np
mean = 1.5
sigma = 2.5
x = np.array((1.2, 7.2))
f = np.exp(-(x - mean)**2 / sigma**2)
fn = f / np.sqrt(2 * np.pi * sigma**2)
print("g({}) = {} -> {} normalized".format(x[0], f[0], fn[0]))
print("g({}) = {} -> {} normalized".format(x[1], f[1], fn[1]))
# g(1.2) = 0.985703184122443 -> 0.15729547042910427 normalized
# g(7.2) = 0.005525397988803928 -> 0.0008817259495115716 normalized |
|
Right now, these are the same number: CHECK(gauss.getValue(EvalFunc::Prob) == 0.99282585_a);
CHECK(gauss.getValue(EvalFunc::Eval) == 0.99282585_a);And they shouldn't be. The Prob one should be multiplying by the normalization, which is not supposed to be one (and isn't, according to the check above this one). |
|
Ahh, I was looking at the wrong side. I believe I know the cause of this issue. The normalization value is being reset due to the second setIndices call under getValue. The idea is to pre-compute the values, then use them for other calculations. First, we need to add a 'normalize value' under PdfBase that will cache the computed value. This value will be written under populateArray. The default value needs to be 1.0. I think this will cover the cases. |
|
Thanks for fixing those irritating consts! I think it looks great. Will try to give one last look over and merge very soon. |
|
Nice; cached normalizations may be very useful; they could replace the host array at some point. |
…. Removed const from normalize routines
ParameterContainer is now checked when trace is on, and mismatched increments are errors. Several problems with 4 body amps mismatching fixed.
Also fixing extra white space and normalize spelling. Nicer makefile too. Bug in usage of setData fixed.