Simple Turing model fails for InexactError

Specific Domains Probabilistic programming
1

Hi all,

The model is a simple binomial on binary data:

@model function coinflip_binomial_broken(y::Array{Int64})
    # prior on p
    p ~ Beta(1, 1)
    
    # updates on p
    heads = sum(y)
    heads ~ Binomial(length(y), p)
end

When run, I get InexactError: Int64(3.0555688441753306) (the value changes each time) which seems to be coming from the logpdf calculation for the Binomial.

I can make it work by pulling the sum and length functions out of the model block:

@model function coinflip_binomial(heads::Int64, flips::Int64)
    # prior on p
    p ~ Beta(1, 1)
    
    # update on p
    heads ~ Binomial(flips, p)
end

That said, why does the first one not work???

It seems to be due to sum(y) (rather than length(y)), which I deduced by redundantly passing the heads variable to the model instead of calculating it within the block while leaving the length call as is.

As a beginner to Turing and to Julia, I’d really like to understand what’s happening here. I suspect my understanding of how to write and interpret code in an AD setting is where my gap is. Any ideas on the above or suggestions for resources would be much appreciated.

Thanks!

2

The stack trace would help as we can’t see where your error happens. But I’d take a guess and say you have a Vector{Int} somewhere and calculate a Float64 which you then assign to that vector. That causes an InexactError because you can’t convert any float to an integer. Or any other container parameterized with Int.

3

I am unable to produce your result. Please be sure to post a working example that reproduces the error that you find. Also, please indicate which version of Turing you are using.

On my system, using Turing 0.15.8, I experience tons of gradient errors. I am guessing that the macro is not parsing the length(y) in the Binomial function.

using Turing, Random

@model function coinflip_binomial_broken(y::Array{Int64})
    # prior on p
    p ~ Beta(1, 1)
    
    # updates on p
    heads = sum(y)
    heads ~ Binomial(length(y), p)
end

Random.seed!(5484)
y = [1,0,0,1]
chain = sample(coinflip_binomial_broken(y), NUTS(1000,.65), 2000)

Output:

Summary Statistics
  parameters      mean       std   naive_se      mcse       ess      rhat 
      Symbol   Float64   Float64    Float64   Float64   Float64   Float64 

       heads    2.0000    0.0000     0.0000    0.0000       NaN       NaN
           p    0.2672    0.0000     0.0000    0.0000    7.1985    1.0187

Quantiles
  parameters      2.5%     25.0%     50.0%     75.0%     97.5% 
      Symbol   Float64   Float64   Float64   Float64   Float64 

       heads    2.0000    2.0000    2.0000    2.0000    2.0000
           p    0.2672    0.2672    0.2672    0.2672    0.2672

As you can see, the sampler does not move beyond its initial value.

4

You are correct. It is not length, but rather heads = sum(y) that is the culprit. The following works.

using Turing, Random

@model function coinflip_binomial_broken(y::Array{Int64})
    # prior on p
    p ~ Beta(1, 1)
    
    # updates on p
    heads = sum(y)
    2 ~ Binomial(length(y), p)
end

Random.seed!(5484)
y = [1,0,0,1]
chain = sample(coinflip_binomial_broken(y), NUTS(1000,.65), 2000)
5

I opened up an issue because it seems like a bug.

6

Thank you! I’ll follow along in the issue thread.