Skip to content
master
Go to file
Code
This branch is 38 commits behind gradhep:master.

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
nbs
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

README.md

neos

nice end-to-end optimized statistics ;)

DOI CI Binder

neos logo

Install

Just run

python -m pip install neos

Contributing

Please read CONTRIBUTING.md before making a PR, as this project is maintained using nbdev, which operates completely using Jupyter notebooks. One should make their changes in the corresponding notebooks in the nbs folder (including README changes -- see nbs/index.ipynb), and not in the library code, which is automatically generated.

Example using a sigmoid-based neural network:

import jax
import neos.makers as makers
import neos.cls as cls
import numpy as np
import jax.experimental.stax as stax
import jax.experimental.optimizers as optimizers
import jax.random
import time

Initialise network using jax.experimental.stax

init_random_params, predict = stax.serial(
    stax.Dense(1024),
    stax.Relu,
    stax.Dense(1024),
    stax.Relu,
    stax.Dense(2),
    stax.Softmax,
)

Initialse tools from neos:

The way we initialise in neos is to define functions that make a statistical model from histograms, which in turn are themselves made from a predictive model, such as a neural network. Here's some detail on the unctions used below:

  • hists_from_nn_three_blobs(predict) uses the nn decision function predict defined in the cell above to form histograms from signal and background data, all drawn from multivariate normal distributions with different means. Two background distributions are sampled from, which is meant to mimic the situation in particle physics where one has a 'nominal' prediction for a nuisance parameter and then an alternate value (e.g. from varying up/down by one standard deviation), which then modifies the background pdf. Here, we take that effect to be a shift of the mean of the distribution. The value for the background histogram is then the mean of the resulting counts of the two modes, and the uncertainty can be quantified through the count standard deviation.
  • nn_hepdata_like(hmaker) uses hmaker to construct histograms, then feeds them into the neos.models.hepdata_like function that constructs a pyhf-like model. This can then be used to call things like logpdf and expected_data downstream.
  • cls_maker takes a model-making function as it's primary argument, which is fed into functions from neos.fit that minimise the logpdf of the model in both a constrained (fixed parameter of interest) and a global way. Moreover, these fits are wrapped in a function that allows us to calculate gradients through the fits using fixed-point differentiation. This allows for the calculation of both the profile likelihood and its gradient, and then the same for cls :)

All three of these methods return functions. in particular, cls_maker returns a function that differentiably calculates CLs values, which is our desired objective to minimise.

hmaker = makers.hists_from_nn_three_blobs(predict)
nnm = makers.nn_hepdata_like(hmaker)
loss = cls.cls_maker(nnm, solver_kwargs=dict(pdf_transform=True))
_, network = init_random_params(jax.random.PRNGKey(2), (-1, 2))
/home/phinate/envs/neos/lib/python3.7/site-packages/jax-0.1.59-py3.7.egg/jax/lib/xla_bridge.py:122: UserWarning: No GPU/TPU found, falling back to CPU.

Define training loop!

opt_init, opt_update, opt_params = optimizers.adam(1e-3)

def update_and_value(i, opt_state, mu):
    net = opt_params(opt_state)
    value, grad = jax.value_and_grad(loss)(net, mu)
    return opt_update(i, grad, opt_state), value, net

def train_network(N):
    cls_vals = []
    _, network = init_random_params(jax.random.PRNGKey(1), (-1, 2))
    state = opt_init(network)
    losses = []
    
    for i in range(N):
        start_time = time.time()
        state, value, network = update_and_value(i,state,1.0)
        epoch_time = time.time() - start_time
        losses.append(value)
        metrics = {"loss": losses}
        yield network, metrics, epoch_time

Let's run it!!

maxN = 20 # make me bigger for better results!

# Training
for i, (network, metrics, epoch_time) in enumerate(train_network(maxN)):
    print(f"epoch {i}:", f'CLs = {metrics["loss"][-1]}, took {epoch_time}s') 
epoch 0: CLs = 0.06680655092981347, took 5.355436325073242s
epoch 1: CLs = 0.4853891149072429, took 1.5733795166015625s
epoch 2: CLs = 0.3379355596004474, took 1.5171947479248047s
epoch 3: CLs = 0.1821927415636535, took 1.5081253051757812s
epoch 4: CLs = 0.09119136931683047, took 1.5193650722503662s
epoch 5: CLs = 0.04530559823843272, took 1.5008423328399658s
epoch 6: CLs = 0.022572851867672883, took 1.499192476272583s
epoch 7: CLs = 0.013835564056077887, took 1.5843737125396729s
epoch 8: CLs = 0.01322058601444187, took 1.520324468612671s
epoch 9: CLs = 0.013407422454837725, took 1.5050244331359863s
epoch 10: CLs = 0.011836452218993765, took 1.509469985961914s
epoch 11: CLs = 0.00948507486266359, took 1.5089364051818848s
epoch 12: CLs = 0.007350505632595539, took 1.5106918811798096s
epoch 13: CLs = 0.005755974539907838, took 1.5267891883850098s
epoch 14: CLs = 0.0046464301411786035, took 1.5851080417633057s
epoch 15: CLs = 0.0038756402968267434, took 1.8452086448669434s
epoch 16: CLs = 0.003323640670405803, took 1.9116990566253662s
epoch 17: CLs = 0.0029133909840759475, took 1.7648999691009521s
epoch 18: CLs = 0.002596946123608612, took 1.6314191818237305s
epoch 19: CLs = 0.0023454051342963744, took 1.5911424160003662s

And there we go!!

If you want to reproduce the full animation, a version of this code with plotting helpers can be found in demo_training.ipynb! :D

Thanks

A big thanks to the teams behind jax, fax, and pyhf for their software and support.

About

Upstream optimisation for downstream inference

Resources

License

Packages

No packages published
You can’t perform that action at this time.