Add support for discrete variables in plot_bpv #1379
Conversation
Codecov Report
@@ Coverage Diff @@
## master #1379 +/- ##
==========================================
- Coverage 91.72% 91.63% -0.10%
==========================================
Files 105 105
Lines 10941 10965 +24
==========================================
+ Hits 10036 10048 +12
- Misses 905 917 +12
Continue to review full report at Codecov.
|
|
I understood that the current loo-pit approximation is not appliable to discrete data. I had this in mind to implement some day https://doi.org/10.1111/j.1541-0420.2009.01191.x Disclaimer: I have yet to review the PR, will do it soon |
|
Right, loo-pit is not applicable to discrete data for the same reasons plot_bpv isn't . I will check that paper, thanks! |
aloctavodia
changed the title
|
It seems I need to think a little bit more about this. |
aloctavodia
changed the title
aloctavodia
force-pushed
the
bpv_discrete
branch
from
9c8431c
to
0bba9a5
Compare
| tstat_pit_dens.size, | ||
| n_ref, |
There was a problem hiding this comment.
I think this is also needed in bokeh backend
OriolAbril
changed the title
| @@ -83,6 +88,16 @@ def plot_bpv( | |||
| obs_vals = obs_vals.flatten() | |||
| pp_vals = pp_vals.reshape(total_pp_samples, -1) | |||
|
|
|||
| print(obs_vals.shape, pp_vals.shape) | |||
There was a problem hiding this comment.
I just want to keep the users informed
|
There are some duplication, but I think we can move things to the main function maybe in another PR when we have time? |
| fake = {"a": np.random.poisson(2.5, 1000)} | ||
| fake_model = from_dict(posterior_predictive=fake, observed_data=fake) | ||
| fake_obs = {"a": np.random.poisson(2.5, 100)} | ||
| fake_pp = {"a": np.random.poisson(2.5, (10, 100))} |
There was a problem hiding this comment.
As a 2 dim array this will be understood as a (chain, draw) array, you should add a dimension at the beggining.
| if pp_vals.ndim > 2: | ||
| pp_vals = pp_vals.reshape(total_pp_samples, -1) |
There was a problem hiding this comment.
I think this will break the ArviZ shape convention and interpret the draw dimension as the observations.
|
Will merge once/if tests pass and release |
Fix
plot_bpvto work with discrete variables. A similar fix could be applied toplot_loopitIt also add an option (default to false) to plot the mean square error between the uniform distribution and marginal p_value
Old
New
data =np.random.binomial(n=1, p=0.5, size=1000)
with pm.Model() as momo:
p = pm.Beta("p", 1., 1.)
data = np.random.binomial(n=1, p=0.5, size=50)
with pm.Model() as momo:
p = pm.Beta("p", 1., 1.)
data = np.random.binomial(n=1, p=0.5, size=100)
with pm.Model() as momo:
p = pm.Beta("p", 1., 1.)
data =np.random.binomial(n=1, p=0.5, size=100)
with pm.Model() as momo:
p = pm.Beta("p", 100., 1.)
data =np.random.binomial(n=1, p=0.8, size=100)
with pm.Model() as momo:
p = pm.Beta("p", 50, 50)
data = np.random.poisson(2.5, size=1000)
with pm.Model() as momo:
p = pm.Uniform('p', 0, 10)
data = np.random.poisson(2.5, size=1000)
with pm.Model() as momo:
p = pm.Uniform('p', 3, 10)
data = np.random.poisson(2.5, size=1000)
data[:100] = 4
with pm.Model() as momo:
p = pm.Uniform('p', 0, 10)
The two following
examples are not affected by this PR, but I add them here as reference.
normal model sample size= 1000
normal model sample size= 100