Transformed variable failure during Metropolis sampling

[fonnesbeck]

I've come across a curious error related to using transformed variables in Metropolis sampling. Specifically, a MissingInputError is raised when sample is called. Replacing the variable in question with a constant results in the next transformed variable failing for the same reason.

Here is an example of a model that fails because of this

BTW, setting exception_verbisity to "high" does nothing to improve the verbosity of the output.

[twiecki]

Did this stop working after #1215 was merged?

[fonnesbeck]

I don't know, as this is a new model, but it seems unlikely since transformations have always been renamed to something. I will try it though.

[fonnesbeck]

Verified that this is not the issue. After reverting to before the pull request was merged:

MissingInputError: An input of the graph, used to compute Elemwise{exp,no_inplace}(cold_mort_interval), was not provided and not given a value.Use the Theano flag exception_verbosity='high',for more information on this error.

NUTS also fails, but because it receives NaNs in the gradient. So, I believe something deeper is the issue.

[fonnesbeck]

Running the same model with ADVI yields:

NullTypeGradError: tensor.grad encountered a NaN. This variable is Null because the grad method 
for input 3 (Elemwise{true_div,no_inplace}.0) of the RandomFunction{multinomial_helper} op is 
mathematically undefined. No gradient defined through raw random numbers op

[twiecki]

It's quite a complex model it seems, perhaps try to work backwards?