Hmm - does anyone know of a long-term design doc for Hatch? I'm curious as to whether there's a plan to add a
hatch add command (or something similar).
I want to fill a histogram many times with the same observations, the only thing that is different is the weights (systematic uncertainties). Instead of simply filling the histogram repeatedly, I assume one could save computation time here because finding the bin indices needs to be one only once. Is this possible in boost_histogram?
@jrueb: does the
weight argument cover your needs? https://boost-histogram.readthedocs.io/en/latest/usage/storage.html?highlight=weight#double
The idea behind boostorg/histogram#211 is just to make this a bit more efficient
A few small fixes in Hist 2.6.1, just released. It won’t try to produce a fancy repr if it can’t make a fast one now, and just falls back on standard reprs. Also doesn’t import SciPy unless you need it, important for something else we are working as well as using Hist in WebAssembly - SciPy is huge and takes several seconds to load in a browser (more with a bad connection) (it does cache, like pip).
Hello experts, I am trying to use the Mean storage using Hist in order to save the average of the weights in each bin. I currently am using the following Hist.
# fill hist with all variables
h = (
Hist.new.Reg(
30,
np.min(predictions["TwoBody_PT"]),
np.max(predictions["TwoBody_PT"]),
name="bpt"
)
.Reg(
30,
np.min(predictions["TwoBody_FDCHI2_OWNPV"]),
np.max(predictions["TwoBody_FDCHI2_OWNPV"]),
name="fd"
)
.Double() # would like to replace with .Mean()
.fill(
bpt=np.array(predictions.TwoBody_PT),
fd=np.array(predictions.TwoBody_FDCHI2_OWNPV),
weight=np.array(predictions.preds_per_cand) # the weight is really the response of a NN
)
)However, when I try using .Mean(), I get the following error: ValueError: Sample array must be 1D. Could you help me understand what am I doing wrong? Many thanks in advance.
It is sad to see you and many others using Hist instead of boost-histogram, the underlying library. The hist syntax goes against common design principles, namely that one should not make half-constructed objects.
Regarding the error, you need to check the shape of what you pass to
weight=
This should be a 1D array, not 2D
I'm not sure that that rule-of-thumb applies here: the
Hist.new method returns a construction proxy, which doesn't have invalid (aka half-constructed) states (it's just the builder pattern). One might prefer to construct the axes and construct the hist with them, à la boost-histogram, but I think that's a matter of API preference.
Angus is right - the construction proxy is not a partially contructed histogram. There are no methods other than adding a new axis and finalizing the proxy with the selection of a storage type - that produces the histogram. It’s not “bad API”, it’s just an alternative choice. Hist is intetend for users, and boost-histogram is intended to be built on and is not really intended to be used directly - you’ve said so yourself. Some users really enjoy the QuickConstruct system - it’s quite nice on the command line since you don’t need extra imports and can just type as you think. But Hist fully suppports classic contruction as well (and I tend toward that one if I’m not typing live).
@henryiii: I know this is fairly bad as bug reports go, but I'm wondering if I've overlooked something w.r.t to the
density argument of mplhep.histplot? I expect it to normalise the area (np.sum(np.diff(e) * b)) but it seems to be doing something else.
5 replies
(not urgent)
Could I get a sanity check that I'm seeing a bug rather than misusing Hist?
Second one will slice and sum on the slice (no underflow bin). First one will sum over the flow bins too, which includes everything you removed with the first cut
0:len:sum will avoid flow bins
Thanks!
Try again - should
project be ignoring my slice?
I thought I'd messed that up, but I think the example is correct.
How would I best go about scaling a histogram in a way that depends on a categorical axis?
import hist
h = hist.Hist.new.Reg(3, 0, 3).StrCat(["a", "b"], name="cat").Double()
h.fill([1,1,2], cat="a")
h.fill([0], cat="b")
h[:, "a"] *= 2This is what I naively tried out, which is hitting a TypeError: Not supported yet via boost-histogram. My use case in practice here is that I gather the same kind of distributions for a lot of different processes, and want to scale the distributions in a process-dependent manner.
2 replies
it is producing a histogram with 0 counts... but when i change
overflow=True the histogram gets filled
specifically it's this code here https://github.com/faroukmokhtar/boostedhiggs/blob/main/python/make_1dhists.py#L50
these are the results
@alexander-held: you can directly modify the contents, e.g
h[...] = h.values() * [[2, 1]]
Yes, that’s it. Use .values(). Or use Hist (built in plotting) or use mplhep (native support for histograms)
The view is a structured array
