IndexError using ordered transform

[sammosummo]

import pymc3 as pm

with pm.Model():
    x = pm.Uniform('x', 0., 1., shape=2, transform=pm.distributions.transforms.ordered)
    pm.sample()

Traceback (most recent call last):
  File "C:/Users/sm2286/PycharmProjects/ForgettingPitch/_testordered.py", line 4, in <module>
    x = pm.Uniform('x', 0., 1., shape=2, transform=pm.distributions.transforms.ordered)
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\pymc3\distributions\distribution.py", line 46, in __new__
    return model.Var(name, dist, data, total_size)
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\pymc3\model.py", line 833, in Var
    model=self)
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\pymc3\model.py", line 1557, in __init__
    transformed_name, transform.apply(distribution), total_size=total_size)
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\pymc3\distributions\transforms.py", line 95, in apply
    return TransformedDistribution.dist(dist, self)
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\pymc3\distributions\distribution.py", line 56, in dist
    dist.__init__(*args, **kwargs)
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\pymc3\distributions\transforms.py", line 120, in __init__
    testval = forward(dist.default())
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\pymc3\distributions\transforms.py", line 334, in forward
    y = tt.inc_subtensor(y[..., 0], x[..., 0])
  File "C:\Users\sm2286\Anaconda3\envs\PerceptualAnchors\lib\site-packages\theano\tensor\var.py", line 510, in __getitem__
    raise IndexError('too many indices for array')
IndexError: too many indices for array

[junpenglao]

Did you try specifying the test_val? I think this is still required currently.

[junpenglao]

Yep, you need to do:

import pymc3 as pm

with pm.Model():
    x = pm.Uniform('x', 0., 1., shape=2, transform=pm.distributions.transforms.ordered, testval=np.asarray([.5, .9]))
    pm.sample(init='adapt_diag')

You probably should use a Chain transformation to make it sample better, see eg: cell [22] in https://docs.pymc.io/notebooks/api_quickstart.html

[sammosummo]

Can you explain what you mean by Chain transformation and why this improves sampling?

[rpgoldman]

The transformations are now documented at https://docs.pymc.io/api/distributions/transforms.html

Please let us know if this material is not clear -- I did the extraction of docstrings from code, but I'm not responsible for any of the code itself. So it would be good to know if these leave unanswered questions, so we can fix them.

[eigenfoo]

@sammosummo you can check out this "release blog post" by Colin: https://colindcarroll.com/2018/07/20/why-im-excited-about-pymc3-v3.5.0/

Specifically, the section 5 on the ordered transform. Hope this helps!