Differentiable (binned) likelihoods in JAX.
python -m pip install evermore
From source:
git clone https://github.com/pfackeldey/evermore
cd evermore
python -m pip install .
See more in examples/
evermore in a nutshell:
from typing import NamedTuple
import equinox as eqx
import jax
import jax.numpy as jnp
from jaxtyping import Array, PyTree
import evermore as evm
jax.config.update("jax_enable_x64", True)
# define a simple model with two processes and two parameters
def model(params: PyTree, hists: dict[str, Array]) -> Array:
mu_modifier = params.mu.scale()
syst_modifier = params.syst.scale_log(up=1.1, down=0.9)
return mu_modifier(hists["signal"]) + syst_modifier(hists["bkg"])
def loss(
diffable: PyTree,
static: PyTree,
hists: dict[str, Array],
observation: Array,
) -> Array:
params = eqx.combine(diffable, static)
expectation = model(params, hists)
# Poisson NLL of the expectation and observation
log_likelihood = evm.pdf.Poisson(lamb=expectation).log_prob(observation)
# Add parameter constraints from logpdfs
constraints = evm.loss.get_log_probs(params)
log_likelihood += evm.util.sum_over_leaves(constraints)
return -jnp.sum(log_likelihood)
# setup data
hists = {"signal": jnp.array([3]), "bkg": jnp.array([10])}
observation = jnp.array([15])
# define parameters, can be any PyTree of evm.Parameters
class Params(NamedTuple):
mu: evm.Parameter
syst: evm.NormalParameter
params = Params(mu=evm.Parameter(1.0), syst=evm.NormalParameter(0.0))
diffable, static = evm.parameter.partition(params)
# Calculate negative log-likelihood/loss
loss_val = loss(diffable, static, hists, observation)
# gradients of negative log-likelihood w.r.t. diffable parameters
grads = eqx.filter_grad(loss)(diffable, static, hists, observation)
print(f"{grads.mu.value=}, {grads.syst.value=}")
# -> grads.mu.value=Array([-0.46153846]), grads.syst.value=Array([-0.15436207])
See CONTRIBUTING.md for instructions on how to contribute.
Distributed under the terms of the BSD license.