I am trying to code up Chapter 1 in the book “Modeling-based Machine Learning”.
Here is my code, which models a crime scene and two possible murderers, modeled by a Bernoulli function. See the function model5(). There are two possible weapons, also modeled by a Bernoulli. Note that the probability p of the Bernoulli depends on who is the murderer. Finally, a bullet is found on the scene, which is an observation of the murder weapon. When I run this program I get a warning error on line 71 that
"/Users/erlebach/anaconda3/lib/python3.7/site-packages/pyro/primitives.py:71: RuntimeWarning: trying to observe a value outside of inference at weapon
RuntimeWarning) (LIne 71 is past the last line of code)
"
Why does this warning occur?
I also get a runtime error: RuntimeError: output with type torch.LongTensor doesn’t match the desired type torch.FloatTensor
which I can trace back to the line
emp = EmpiricalMarginal(posterior.run(), sites=“murderer”)
I include the error below the source code.
import sys, os
import numpy as np
import torch
import pyro
import pyro.infer
import torch.tensor as tensor
import pyro.optim
import pyro.distributions as dist
from pyro.optim import Adam
from pyro.infer import SVI, Trace_ELBO, TraceEnum_ELBO, config_enumerate, infer_discrete
from pyro.infer import EmpiricalMarginal
import torch.distributions.constraints as constraints
from torch.distributions.beta import Beta
from torch.distributions import Normal, Bernoulli
from pyro.infer.mcmc import HMC, NUTS, MCMC
from functools import reduce, partial
import time
assert pyro.__version__.startswith('0.3.0')
# Check function arguments
pyro.enable_validation()
# Generate new random numbers on each run
pyro.set_rng_seed(0)
def model5():
print("enter model5")
# murderer = 1 (Mr. Auburn)
# murderer = 0 (Ms. Gray)
murderer = dist.Bernoulli(0.7)
# prior
murderer = pyro.sample("murderer", murderer)
# weapon == 1 (Gun)
# weapon == 0 (dagger)
if murderer == 0: # Ms. Gray
p = 0.9
else: # Mr. Auburn
p = 0.2
# conditional probability: weapon given the murderer
weapon = dist.Bernoulli(p)
# A bullet is found on the murder scene. That means that the weapon is observed.
obs = pyro.sample("weapon", weapon, obs=torch.tensor(1))
# Latent variable: murderer
# Observed variable: weapon
model5()
posterior = pyro.infer.Importance(model5, None, num_samples=10)
print("posterior= ", posterior)
emp = EmpiricalMarginal(posterior.run(), sites="murderer")
print(emp)
===================================
RuntimeError Traceback (most recent call last)
~/Documents/src/2018/model-based_machine_learning/chapter1/infer.py in <module>()
59 posterior = pyro.infer.Importance(model5, None, num_samples=10)
60 print("posterior= ", posterior)
---> 61 emp = EmpiricalMarginal(posterior.run(), sites="murderer")
62 print(emp)
63
~/anaconda3/lib/python3.7/site-packages/pyro/infer/abstract_infer.py in run(self, *args, **kwargs)
198 self._reset()
199 with poutine.block():
--> 200 for i, vals in enumerate(self._traces(*args, **kwargs)):
201 if len(vals) == 2:
202 chain_id = 0
~/anaconda3/lib/python3.7/site-packages/pyro/infer/importance.py in _traces(self, *args, **kwargs)
43 model_trace = poutine.trace(
44 poutine.replay(self.model, trace=guide_trace)).get_trace(*args, **kwargs)
---> 45 log_weight = model_trace.log_prob_sum() - guide_trace.log_prob_sum()
46 yield (model_trace, log_weight)
47
~/anaconda3/lib/python3.7/site-packages/pyro/poutine/trace_struct.py in log_prob_sum(self, site_filter)
134 else:
135 try:
--> 136 log_p = site["fn"].log_prob(site["value"], *site["args"], **site["kwargs"])
137 except ValueError:
138 _, exc_value, traceback = sys.exc_info()
~/anaconda3/lib/python3.7/site-packages/torch/distributions/bernoulli.py in log_prob(self, value)
92 self._validate_sample(value)
93 logits, value = broadcast_all(self.logits, value)
---> 94 return -binary_cross_entropy_with_logits(logits, value, reduction='none')
95
96 def entropy(self):
~/anaconda3/lib/python3.7/site-packages/torch/nn/functional.py in binary_cross_entropy_with_logits(input, target, weight, size_average, reduce, reduction, pos_weight)
2075 raise ValueError("Target size ({}) must be the same as input size ({})".format(target.size(), input.size()))
2076
-> 2077 return torch.binary_cross_entropy_with_logits(input, target, weight, pos_weight, reduction_enum)
2078
2079
RuntimeError: output with type torch.LongTensor doesn't match the desired type torch.FloatTensor
