eas342
commented
1 year ago The error goes away if I change the sampling from pymc3 to pymc3_ext. I wish I'd noticed that sooner!
This gives an EOFError:
import pymc3_ext as pmx
import pymc3 as pm
trace = pm.sample(
tune=250,
draws=500,
start=map_soln,
chains=4,
cores=4,
target_accept=0.9,
)This works with no problem. Notice the X in pmx!:
import pymc3_ext as pmx
import pymc3 as pm
trace = pmx.sample(
tune=250,
draws=500,
start=map_soln,
chains=4,
cores=4,
target_accept=0.9,
)
dfm
catrionamurray
Describe the bug I had an issue with sampling in starry on multiple cores when the orbit is variable. This had me running everything more slowly on one core for a while. The error traceback I got was:
Error traceback
--------------------------------------------------------------------------- EOFError Traceback (most recent call last) Input In [24], inTo Reproduce
Minimal-ish example adapted from the "Hot jupiter phase curve example"
import starry import matplotlib.pyplot as plt import numpy as np import pymc3 as pm import pymc3_ext as pmx import exoplanet starry.config.quiet = True np.random.seed(1) # In[7]: A = starry.Primary(starry.Map(ydeg=0, udeg=2, amp=1.0), m=1.0, r=1.0, prot=1.0) A.map[1] = 0.4 A.map[2] = 0.2 # In[8]: # These are the parameters we're going to try to infer log_amp_true = -3.0 offset_true = 30.0 b = starry.Secondary( starry.Map(ydeg=1, udeg=0, amp=10 ** log_amp_true, inc=90.0, obl=0.0), m=0.0, r=0.1, inc=90.0, prot=1.0, porb=1.0, ) b.map[1, 0] = 0.5 b.theta0 = 180.0 + offset_true # In[9]: sys = starry.System(A, b) # In[10]: t = np.linspace(-0.3, 1.3, 1000) flux_true = sys.flux(t).eval() ferr = 1e-4 flux = flux_true + ferr * np.random.randn(len(t)) plt.figure(figsize=(12, 5)) plt.plot(t, flux, "k.", alpha=0.3, ms=3) plt.plot(t, flux_true) plt.xlabel("Time [days]", fontsize=24) plt.ylabel("Flux [normalized]", fontsize=24); # In[21]: with pm.Model() as model: # These are the variables we're solving for; # here we're placing wide Gaussian priors on them. #offset = pm.Normal("offset", 0.0, 50.0, testval=0.11) offset=offset_true log_amp = pm.Normal("log_amp", -4.0, 2.0, testval=-3.91) porb = pm.Normal('porb',mu=1.0,sigma=0.02) #porb = 1.0 # Instantiate the star; all its parameters are assumed # to be known exactly A = starry.Primary( starry.Map(ydeg=0, udeg=2, amp=1.0, inc=90.0, obl=0.0), m=1.0, r=1.0, prot=1.0 ) A.map[1] = 0.4 A.map[2] = 0.2 # Instantiate the planet. Everything is fixed except for # its luminosity and the hot spot offset. b = starry.Secondary( starry.Map(ydeg=1, udeg=0, amp=10 ** log_amp, inc=90.0, obl=0.0), m=0.0, r=0.1, prot=1.0, porb=porb, ) b.map[1, 0] = 0.5 b.theta0 = 180.0 + offset # Instantiate the system as before sys = starry.System(A, b) # Our model for the flux flux_model = pm.Deterministic("flux_model", sys.flux(t)) # This is how we tell `pymc3` about our observations; # we are assuming they are ampally distributed about # the true model. This line effectively defines our # likelihood function. pm.Normal("obs", flux_model, sd=ferr, observed=flux) # In[22]: with model: map_soln = pmx.optimize() # In[23]: plt.figure(figsize=(12, 5)) plt.plot(t, flux, "k.", alpha=0.3, ms=3) plt.plot(t, map_soln["flux_model"]) plt.xlabel("Time [days]", fontsize=24) plt.ylabel("Flux [normalized]", fontsize=24); # In[24]: with model: trace = pm.sample( tune=250, draws=500, start=map_soln, chains=4, cores=4, target_accept=0.9, )Expected behavior Should sample the posterior and calculate a trace object. Instead, I get the error.
Your setup (please complete the following information):
Additional context This happened to me in a fairly specific set of circumstances:
starry.I have found a solution/workaround and wanted to share it in case anyone else gets EOFError.