Diagnostic a pathological behavior: solution? #28
Replies: 4 comments 5 replies
-
|
Dear @Joshuaalbert, |
Beta Was this translation helpful? Give feedback.
-
|
I'd be happy to help. Firstly, can you downgrade jax to 0.0.6 for these tests, until 1.0 is released? |
Beta Was this translation helpful? Give feedback.
-
|
Heuu it is jaxns that I should downgrade, isn't it? |
Beta Was this translation helpful? Give feedback.
-
|
Hello @Joshuaalbert, Here is my Nympyro code def model():
# Cosmological params (test13 JEC)
Omega_c = numpyro.sample('Omega_c', dist.Uniform(0.1, 0.4))
sigma8 = numpyro.sample('sigma8', dist.Uniform(0.5, 1.2))
Omega_b = numpyro.sample('Omega_b', dist.Uniform(0.01, 0.09))
h = numpyro.sample('h', dist.Uniform(0.4, 1.0))
n_s = numpyro.sample('n_s', dist.Uniform(0.5, 1.25))
w0 = numpyro.sample('w0', dist.Uniform(-2.0, -0.0001))
# Astrophysical params
A = numpyro.sample('A', dist.Uniform(0., 2.5))
eta = numpyro.sample('eta', dist.Uniform(0., 6.))
# parameters for systematics
m = [numpyro.sample('m%d'%i, dist.Normal(0.012, 0.023))
for i in range(1,5)]
dz1 = numpyro.sample('dz1', dist.Normal(0.001, 0.016))
dz2 = numpyro.sample('dz2', dist.Normal(-0.019, 0.013))
dz3 = numpyro.sample('dz3', dist.Normal(0.009, 0.011))
dz4 = numpyro.sample('dz4', dist.Normal(-0.018, 0.022))
# Now that params are defined, here is the forward model
cosmo = FiducialCosmo(Omega_c=Omega_c, sigma8=sigma8, Omega_b=Omega_b,
h=h, n_s=n_s, w0=w0)
signal = model_fn(get_params_vec(cosmo, m, [dz1, dz2, dz3, dz4], [A, eta]))
# And here we define the likelihood
numpyro.sample('cl_wl', dist.MultivariateNormal(signal, C), obs=cl_obs)Here is my tentative to setup a JaxNS code: #@jax.jit
#code from https://github.com/google/jax/issues/2314
def multi_gauss_logpdf(x, mean, cov):
""" Calculate the probability density of a
sample from the multivariate normal. """
D = mean.shape[0]
(sign, logdet) = np.linalg.slogdet(cov)
p1 = D*np.log(2*np.pi) + logdet
p2 = (x-mean).T @ np.linalg.inv(cov) @ (x-mean)
return -1./2 * (p1 + p2)
def solve(cov,cl_obs):
def log_lik(Omega_c, sigma8, Omega_b,h, n_s, w0, A, eta,
m1, m2,m3, m4,
dz1, dz2, dz3, dz4,
**kwargs):
cosmo = FiducialCosmo(Omega_c=Omega_c, sigma8=sigma8, Omega_b=Omega_b,
h=h, n_s=n_s, w0=w0)
signal = model_fn(get_params_vec(cosmo, [m1, m2,m3, m4], [dz1, dz2, dz3, dz4], [A, eta]))
### Comment faire numpyro.sample('cl_wl', dist.MultivariateNormal(signal, cov), obs='cl_obs')
return multi_gauss_logpdf(cl_obs,signal,cov)
prior_chain = PriorChain(UniformPrior('Omega_c',0.1, 0.4),
UniformPrior('sigma8',0.5, 1.2),
UniformPrior('Omega_b',0.01, 0.09),
UniformPrior('h',0.4, 1.0),
UniformPrior('n_s',0.5, 1.25),
UniformPrior('w0',-2.0, -0.0001),
UniformPrior('A',0.,2.5),
UniformPrior('eta',0.,6.),
NormalPrior('m1',0.012, 0.023),
NormalPrior('m2',0.012, 0.023),
NormalPrior('m3',0.012, 0.023),
NormalPrior('m4',0.012, 0.023),
NormalPrior('dz1',0.001, 0.016),
NormalPrior('dz2',-0.019, 0.013),
NormalPrior('dz3',0.009, 0.011),
NormalPrior('dz4',-0.018, 0.022)
)
print('num_live_points:',prior_chain.U_ndims*500)
ns = jaxns.nested_sampling.NestedSampler(log_lik, prior_chain,
num_live_points=prior_chain.U_ndims*500)
print('Go...')
results = ns(jax.random.PRNGKey(32564))
return results
## Go.
results = solve(C,data)Now I try to run this code on GPU... 2021-10-20 15:11:15.189842: W external/org_tensorflow/tensorflow/core/common_runtime/bfc_allocator.cc:461] Allocator (GPU_0_bfc) ran out of memory trying to allocate 933.17GiB (rounded to 1001989427968)requested by op The stack trace below excludes JAX-internal frames. The above exception was the direct cause of the following exception: Traceback (most recent call last): |
Beta Was this translation helpful? Give feedback.
-
|
The memory load clearly depend on the # of live points, as setting to 1,000 get a 100GiB order (that crashes too) so, I down grade to num live pts 100. But what is strange, that is the Numpyro implementation runs with 10,000 live points ns = NestedSampler(cond_model, num_slices=10,
num_live_points=10000)
ns.run(jax.random.PRNGKey(42))on the same kind of GPU hardware. So I do not understand what is going wrong in my native JaxNS calls... With 100 num live points, the job runs up to the end, but the results are very poor, worse than the Numpyro implementation results. By the way a NUTS run with (Normal priors instead of Uniform priors) gives this kind of results to give you an idea |
Beta Was this translation helpful? Give feedback.
-
|
I just discover that for some parameter sets, it results NaN likelihood value . So, I guess these NaN are making the mess in your algorithm. If it is the case, it would be good to rise an exception, no? |
Beta Was this translation helpful? Give feedback.
-
|
@jecampagne can you pull the latest and try again. The API has changed slightly, see the tutorial here: https://github.com/Joshuaalbert/jaxns/blob/master/examples/simple/mvn_data_mvn_prior.ipynb |
Beta Was this translation helpful? Give feedback.
-
|
@jecampagne does this example require any data? If not, I could add it to the examples section. |
Beta Was this translation helpful? Give feedback.
-
|
Ho, I will let you know but currently I have an other hot topic to finish. Sorry. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Dear experts,
I have posted this message to
Numpyroforum, and they invite me to ask the question here as theirNestedSamplerclass is a wrapper toJaxNS(see here) . Also I must confess that I 'm not an experienced user of Nested SAmpler so my question may be not well formulated.So, I have a
modelwhich I sample with theNUTSsampler. Schematically, I do the following (sorry this is in the Numpyro language)numpyro.samplestatements to define priors (which are for this exemple all Gaussian distributions) and likelihoodfix_cond_model = numpyro.handlers.condition(model, <parmeters defult values>to generate somedatathanks toThen, I proceed to the MCMC run et finally get the samples:
So far so good, now I wander if I can use the
NestedSampler? I have triedFrom the function calls point of view, it seems ok from
Numpyrodevelopers, but the sampling of the variables is clearly pathologic,And here are the arviz kde plots
I certainly miss something. Any idea are welcome. Thanks
Beta Was this translation helpful? Give feedback.
All reactions