rlouf
commented
3 years ago @brandonwillard it looks like I still have a problem with a RandomStream node that I do not understand. The stack trace I guet when executing test_dynamic_expansion in test_trajectory.py seems to indicates that the problems lies comes from the dynamic integration. However the test test_dynamic_integration_divergence does pass. The problem would appear when I share the RandomStream between two functions. Is there a workaround?
Stack trace
``` WARNING (aesara.tensor.blas): Using NumPy C-API based implementation for BLAS functions. Traceback (most recent call last): File "/home/remi/projects/ampersand/aehmc/example.py", line 64, inThe following simplified example does work as expected:
from typing import Callable
import aesara
import aesara.tensor as aet
from aesara.tensor.random.utils import RandomStream
rng = RandomStream(59)
def bernoulli_adder(srng: RandomStream):
def add_bernoulli(b):
def one_step(x):
a = aet.where(srng.bernoulli(0.5), 1, -1)
c = a + x
return c
res, _ = aesara.scan(one_step, outputs_info=(b,), n_steps=10)
return res
return add_bernoulli
def normal_adder(srng: RandomStream, add_bernoulli: Callable):
def add_normal(init):
def one_step(a):
res = add_bernoulli(a)
final = res[-1]
norm = srng.normal(0, 1)
return final + norm
res, _ = aesara.scan(one_step, outputs_info=(init,), n_steps=10)
return res
return add_normal
init = aet.scalar("init")
add_bernoulli = bernoulli_adder(rng)
add_normal = normal_adder(rng, add_bernoulli)
result = add_normal(init)
fn = aesara.function((init,), result)
print(fn(10))Which makes me think that the issue may be more trivial. But at this point I need another pair of eyes to look at the code.
brandonwillard
codecov[bot]
In this PR we add the NUTS kernel.