Generalize samplers to arbitrary function of machine
Created by: gcarleo
This PR allows to sample from other functions of the machine. In general it is now possible to sample from F(machine(x)), for arbitrary real-valued F. By default, F(X)=|X|^2. e.g. Sampling from |machine(x)| is useful when dealing with density matrices for example.
Merge request reports
Merged by
(Apr 3, 2025 2:59am UTC)
Activity
requested review from @filippo.vicentini
requested review from @filippo.vicentini
Created by: twesterhout
How about generalising the approach a bit further? Any function
Ffor whichstd::is_invocable_r<double, F, std::complex<double>>::value == truecan be used asMachineNorm. So why not just makestd::function<double(std::complex<double>)>a parameter rather than anintdenoting order?-
Created by: gcarleo
How about generalising the approach a bit further? Any function
Ffor whichstd::is_invocable_r<double, F, std::complex<double>>::value == truecan be used asMachineNorm. So why not just makestd::function<double(std::complex<double>)>a parameter rather than anintdenoting order?I originally wanted to pass just a
std::functionactually, but the problem is that at some point I needed to compare two functions, to make sure that the initialization of the sampler was correctly done. (see what happens inSampler/ExactSampler.hpp) and this is not too safe, as far as I understand.... In practical cases however one never needs to sample from crazy functions of the machine, so I settled to just the norm int. -
Created by: twesterhout
Well, if a sampler already has
orderas a member variable, why pass it toSweepas function argument as well?As for weird functions... you can never predict user needs in advance. Nice thing about
std::functionis that pybind11 can wrap Python callables instd::functions which means you can experiment with different norms without ever touching C++! -
Created by: gcarleo
Well, if a sampler already has order as a member variable, why pass it to Sweep as function argument as well?
That's the thing, I don't want to pass order in the constructor, but only in
Sweep, because that's more flexible. However there might be some samplers, likeExactSampler, that can speed up the sampling if they know in advanceorder, so that's why it is stored as a private variable in there. Notice that this variable is just used to check if the current order requested bySweepis equal to previous one, if it is not, it recomputes some quantities from scratch, otherwise it can do it faster.If I do this with a generic function, I'd need to check if the currently stored function is the same I am asking for when I call
Sweep... -
Created by: twesterhout
That's the thing, I don't want to pass order in the constructor, but only in Sweep, because that's more flexible.
Could you, please, give an example when using a member variable is not flexible enough?
The design I have in mind is having a function as a member variable + a member function allowing you to change it. This functionality can even be implemented in some base class, no need to redo it in every sampler. Then, 1) if you want to change the norm in the middle of a Monte Carlo simulation, just call the setter; and 2)
ExactSamplercan cache values since it knows the norm beforehand. It also knows when to clear the cache -- whenever someone calls the setter. requested review from @filippo.vicentini
- Last reply by Vicentini Filippo
- Last reply by Vicentini Filippo
- Last reply by Vicentini Filippo
- Last reply by Vicentini Filippo
- Last reply by Vicentini Filippo
- Last reply by Vicentini Filippo
86 88 &AbstractSampler::SetVisible, 87 89 R"EOF( 88 90 numpy.array: The quantum numbers being sampled, 89 and distributed according to $$|\Psi(v)|^2$$ )EOF") 91 and distributed according to $$F(\Psi(v))$$ )EOF") 90 92 .def_property_readonly("acceptance", &AbstractSampler::Acceptance, R"EOF( 91 93 numpy.array: The measured acceptance rate for the sampling. 92 94 In the case of rejection-free sampling this is always equal to 1. )EOF") 93 95 .def_property_readonly("hilbert", &AbstractSampler::GetHilbert, R"EOF( 94 96 netket.hilbert: The Hilbert space used for the sampling. )EOF") 95 97 .def_property_readonly("machine", &AbstractSampler::GetMachine, R"EOF( 96 netket.machine: The machine used for the sampling. )EOF"); 98 netket.machine: The machine used for the sampling. )EOF") 99 .def_property("machine_func", &AbstractSampler::GetMachineFunc,