Start value doesn't work for Uniform distribution

[davidbrochart]

It looks like when a variable has a Uniform distribution, its start value is ignored. This example shows that alpha never gets initialized to 5 when it has a Uniform distribution, but it does when it has a Normal distribution.

import numpy as np
import pymc3 as pm
import theano

# Initialize random number generator
np.random.seed(123)

# True parameter values
alpha, sigma = 1, 1
beta = [1, 2.5]

# Size of dataset
size = 100

# Predictor variable
X1 = np.random.randn(size)
X2 = np.random.randn(size) * 0.2

# Simulate outcome variable
Y = alpha + beta[0]*X1 + beta[1]*X2 + np.random.randn(size)*sigma

# Specify model
@theano.as_op([theano.tensor.dscalar, theano.tensor.dscalar, theano.tensor.dscalar], [theano.tensor.dvector])
def _mu(alpha, beta_0, beta_1):
    print(alpha, beta_0, beta_1)
    return alpha + beta_0*X1 + beta_1*X2

basic_model = pm.Model()

with basic_model:
    alpha = pm.Uniform('alpha', -10, 10)
    #alpha = pm.Normal('alpha', mu=0, sd=10)
    beta_0 = pm.Normal('beta_0', mu=0, sd=10)
    beta_1 = pm.Normal('beta_1', mu=0, sd=10)
    sigma = pm.HalfNormal('sigma', sd=1)
    mu = pm.Deterministic('mu', _mu(alpha, beta_0, beta_1))
    Y_obs = pm.Normal('Y_obs', mu=mu, sd=sigma, observed=Y)

with basic_model:
    step = pm.Metropolis([alpha, beta_0, beta_1, sigma])
    start = {'alpha': 5., 'beta_0': 4., 'beta_1': 3.}
    trace = pm.sample(10, step=step, start=start, init=None)

[fonnesbeck]

I will have a peek at it.

BTW, is there a reason you are creating an op function for mu? You should just be able to specify:

mu = Deterministic('mu', alpha + beta_0*X1 + beta_1*X2)

directly in the model.

[fonnesbeck]

Yes, it appears to happen for any variables that are auto-transformed prior to sampling (i.e. those not supported over the real line).

[davidbrochart]

I created an op function in order to insert a print statement.

[fonnesbeck]

Just pushed a fix; please test.

[davidbrochart]

With your fix, I get this error in the previous example:

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-5-da07ba7e1eef> in <module>()
      2     step = pm.Metropolis([alpha, beta_0, beta_1, sigma])
      3     start = {'alpha': 5., 'beta_0': 4., 'beta_1': 3.}
----> 4     trace = pm.sample(10, step=step, start=start, init=None)

/home/david/soft/anaconda3/lib/python3.5/site-packages/pymc3/sampling.py in sample(draws, step, init, n_init, start, trace, chain, njobs, tune, nuts_kwargs, step_kwargs, progressbar, model, random_seed, live_plot, **kwargs)
    257         sample_func = _sample
    258 
--> 259     return sample_func(**sample_args)
    260 
    261 

/home/david/soft/anaconda3/lib/python3.5/site-packages/pymc3/sampling.py in _sample(draws, step, start, trace, chain, tune, progressbar, model, random_seed, live_plot, **kwargs)
    272     try:
    273         strace = None
--> 274         for it, strace in enumerate(sampling):
    275             if live_plot:
    276                 if it >= skip_first:

/home/david/soft/anaconda3/lib/python3.5/site-packages/tqdm/_tqdm.py in __iter__(self)
    828 """, fp_write=getattr(self.fp, 'write', sys.stderr.write))
    829 
--> 830             for obj in iterable:
    831                 yield obj
    832                 # Update and print the progressbar.

/home/david/soft/anaconda3/lib/python3.5/site-packages/pymc3/sampling.py in _iter_sample(draws, step, start, trace, chain, tune, model, random_seed)
    352         _soft_update(start, strace.point(-1))
    353     else:
--> 354         _soft_update(start, model.test_point)
    355 
    356     try:

/home/david/soft/anaconda3/lib/python3.5/site-packages/pymc3/sampling.py in _soft_update(a, b)
    485     """As opposed to dict.update, don't overwrite keys if present.
    486     """
--> 487     b = _transformed_init(a, b)
    488     a.update({k: v for k, v in b.items() if k not in a})
    489 

/home/david/soft/anaconda3/lib/python3.5/site-packages/pymc3/sampling.py in _transformed_init(a, b)
    477             if tname.startswith(name) and tname!=name:
    478                 transform = tname.split(name)[-1][1:-1]
--> 479                 transform_func = distributions.__dict__[transform]
    480                 b[tname] = transform_func.forward(a[name]).eval()
    481 

KeyError: 'interval'

[davidbrochart]

Sorry, closed the issue by mistake.

[fonnesbeck]

Looks like that transform was missing from __all__. Fixed now.

[fonnesbeck]

Actually, there is a problem with the interval transform in that the bounds of the transformation is not included in the name. Need to address this.

[fonnesbeck]

OK, please test the updated PR.

[davidbrochart]

It works fine, thanks a lot!

[fonnesbeck]

Ok, good. I want the PR to be reviewed before it is merged, as my solution is a bit hacky.

[denadai2]

same error here

[fonnesbeck]

Fixed in master now.