Blobs example a bit confusing #352
Comments
|
Thanks for reporting! You're absolutely right that there are several typos in the docs. The first example should read: def log_prob(params):
lp = log_prior(params)
if not np.isfinite(lp):
return -np.inf, None
ll = log_like(params)
if not np.isfinite(ll):
return -np.inf, None
return lp + ll, lpAnd the second should read: def log_prob(params):
lp = log_prior(params)
if not np.isfinite(lp):
return -np.inf, None, None
ll = log_like(params)
if not np.isfinite(ll):
return -np.inf, None, None
return lp + ll, lp, np.mean(params)In summary:
Let me know if this helps your use cases and we can update the docs. |
|
Ah, you're right. Returning I'm not sure why, exactly, but it's easily avoidable, just by returning |
|
Can you post exactly the code you're running to get this, with a fixed seed and all the relevant imports and data, etc? I wouldn't expect this to ever be hit: it shouldn't inspect the blobs if logprob is |
|
Sure! import emcee
import numpy as np
def log_prob(p):
mu, sig = p
if not 0 < mu < 1:
ll = -np.inf
return -np.inf, 0,0,0
ll = np.sum(-(data - mu)**2 / (2 * sig**2)) - sig
return ll, mu*sig, mu+sig, (data - mu)/sig
np.random.seed(1234)
data = np.random.normal(loc=0.2, scale=0.3, size=25)
sampler = emcee.EnsembleSampler(
nwalkers = 100,
ndim = 2,
log_prob_fn = log_prob,
)
init_pos = np.array([0.2, 0.3]) + 1e-4 * np.random.normal(size=(100, 2))
sampler.run_mcmc(init_pos, nsteps=1000, progress=True) |
|
Hi, I'm very glad I found this issue after also being confused by the blobs documentation. It has a lot of info that could be useful to add in the docs. Here's what I gathered so far (assuming I understand the intended behaviour correctly):
Now regarding 3. I think there might be a bug indeed. Playing with the example from the previous post, it seems that |
|
Some of this is the result of how numpy has changed their handling of dtypes (for the better!) since I originally wrote this forever ago. But either way, I'd say that the best approach in this particular case (whenever you have non-trival blob types) is to provide the dtype explicitly. And don't use blobs_dtype=np.dtype([("a", float), ("b", float), ("c", float, len(data))])Then your resulting blobs array with be a structured array that will be easy to slice, etc. |
The blobs docs were found to be a bit confusing by some users (including me). This merge request updates them following the discussion in dfm#352 to clarify how this interface can be used and the intent behind it.
General information:
Problem description:
Expected behavior:
There are actually a few things that are confusing me (based on the docs), and one thing that (given I understand correctly) could potentially be made better in the code. The first confusing thing in the docs is that (on the "Blobs" page) it gives the following example:
This looks like it's returning
lpas the log-likelihood whenllis infinite. That doesn't seem right... ifllis infinite, then shouldn't the returned probability be-np.inf? Are the returns backwards?Now the next example, for "named blobs and custom dtypes" has this
log_prob:Again, this has the same seemingly backwards return when
llis not finite. But it also has a different number of returns when eitherlporllare infinite (2 instead of 3). Is this really allowed? If so, great (though see below), but then I think it should be more clear that this is really allowed and encouraged.However, in my own application, I have a
log_probthat has dynamic blobs (I let the user specify a bunch of keys that they want to pull out of an object to store as blobs). This works perfectly fine, except that when I do a similarif not np.isfinite(ll): return -np.inf, -np.inf,emceecomplains:Minimal Example(s)
Let
data = np.random.normal(loc=0.2, scale=0.3, size=25)andinit_pos = np.array([0.2, 0.3]) + 1e-4 * np.random.normal(size=(100, 2))Example 1:
This is almost the same as the docs example -- returning two blobs, but only one blob if returning a
-np.infprior. Setup sampler withIt exits with
Example 2:
New
log_probwhere one of the blobs is an array:Exits with
Example 3:
Using the same log prob as example 2, but with a HDF5 backend, and sampler defined as:
Exits with
If I explicitly return
-np.inf, -np.inf, -np.inf, np.zeros(25)when the prior is infinite, then it works fine. However, returning an array of zeros of length 1 but with the correct dtype does not work.This makes it kind of hard to dynamically return blobs. What I could do is create the relevant blob dtype, then unpack it as a tuple, add the likelihood to the front of the tuple, and return that. In this case, I'd have to write the following:
This works, but is clearly a lot more brittle and complex. In cases like these, we know that the blobs will never actually be saved because the likelihood is
-np.infanyway. I wonder if you couldn't just ignore the blobs in these cases?The text was updated successfully, but these errors were encountered: