p;
}
model{
vector[200] mu;
vector[200] nu;
p ~ beta( 2 , 2 );
k ~ exponential( 1 );
sigma ~ exponential( 1 );
tau ~ exponential( 1 );
a1 ~ normal( 0 , 0.5 );
a2 ~ normal( 0 , 0.5 );
b ~ normal( 0 , 0.5 );
m ~ normal( 0 , 0.5 );
U ~ normal( 0 , 1 );
B2 ~ bernoulli( p );
B1 ~ bernoulli( p );
for ( i in 1:200 ) {
nu[i] = a2 + b * B2[i] + m * M[i] + k * U[i];
}
D ~ normal( nu , tau );
for ( i in 1:200 ) {
mu[i] = a1 + b * B1[i] + k * U[i];
}
M ~ normal( mu , sigma );
}
```
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 3, 2022, 10:11am UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/8 "2022-02-03T10:11:14Z")
Thanks for the suggestion! That helped a lot!
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 3, 2022, 10:11am UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/9 "2022-02-03T10:11:37Z")
I hope just running time unless turing recompiles on every call?
---
### Author:  [@rikh](https://discourse.julialang.org/u/rikh)
#### Post date: [February 3, 2022, 10:45am UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/10 "2022-02-03T10:45:01Z")
As long as you keep the Julia process running in between calls via the REPL / Pluto, it‘s not recompiling on the second call no
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 3, 2022, 3:11pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/11 "2022-02-03T15:11:54Z")
In that case it should be without recompilation. On the second run I now get
```julia
Wall duration = 21.42 seconds
Compute duration = 65.88 seconds
```
What is the Wall duration?
---
### Author:  [@Christopher\_Fisher](https://discourse.julialang.org/u/Christopher_Fisher)
#### Post date: [February 3, 2022, 3:28pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/12 "2022-02-03T15:28:18Z")
Wall time is the time for all chains to finish and compute duration is the sum of each chain’s time to complete. Did your last benchmark use `MvNormal`?
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 3, 2022, 3:34pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/13 "2022-02-03T15:34:11Z")
Yes but I did more draws.
For this model
```julia
@model function mom(N, M, D, B1, B2)
p ~ Beta(2,2)
k ~ Exponential(1)
σ ~ Exponential(1)
τ ~ Exponential(1)
a1 ~ Normal(0, 0.5)
a2 ~ Normal(0, 0.5)
b ~ Normal(0, 0.5)
m ~ Normal(0, 0.5)
U ~ filldist(Normal(0,1), N)
B1 ~ Bernoulli(p)
B2 ~ Bernoulli(p)
ν = a2 .+ b * B2 + m * M + k * U
D ~ MvNormal(ν, τ * I)
μ = a1 .+ b * B1 + k * U
M ~ MvNormal(μ, σ * I)
end
Turing.setadbackend(:reversediff)
Turing.setrdcache(true)
flbi = sample(
mom(Int(dat_R[:N]), dat_R[:M], dat_R[:D], dat_R[:B1], dat_R[:B2]),
NUTS(1000, 0.65),
MCMCThreads(),
2_000, 4)
```
I get
```julia
Iterations = 1001:1:3000
Number of chains = 4
Samples per chain = 2000
Wall duration = 10.99 seconds
Compute duration = 37.53 seconds
```
---
### Author:  [@Christopher\_Fisher](https://discourse.julialang.org/u/Christopher_Fisher)
#### Post date: [February 3, 2022, 3:36pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/14 "2022-02-03T15:36:19Z")
Ok. Thanks. I was unable to run the Stan code. How does the new Turing model compare to Stan when the same number of samples are used?
Edit: My mistake. I think that is the comparison with the same number of samples. So it is within a factor of 4?
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 3, 2022, 3:39pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/15 "2022-02-03T15:39:35Z")
I actually think that 2000 iterations in stan means you run 2000 iterations but in Turing you get 2000 iterations with 1000 burn-in. So with 3000 iterations in stan I get
```julia
All 4 chains finished successfully.
Mean chain execution time: 3.2 seconds.
Total execution time: 3.8 seconds.
```
---
### Author:  [@Christopher\_Fisher](https://discourse.julialang.org/u/Christopher_Fisher)
#### Post date: [February 3, 2022, 3:43pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/16 "2022-02-03T15:43:22Z")
There is certainly room for improvement, but 11 seconds is reasonable. In my experience (unless something has changed recently), the problem with Julia’s autodiff is that it does not scale as well as Stan’s. I think the other problem is reversediff does not work well with loops, which can be easier to write in some cases.
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 3, 2022, 3:45pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/17 "2022-02-03T15:45:28Z")
I’ll try to test a couple of different AD backends. Is it ok to set the backend in a live session or would it be better to restart?
---
### Author:  [@Christopher\_Fisher](https://discourse.julialang.org/u/Christopher_Fisher)
#### Post date: [February 3, 2022, 3:48pm UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/18 "2022-02-03T15:48:28Z")
I think it is fine to start in the same session. Unfortunately, reversediff might be the best at the moment.
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 4, 2022, 10:29am UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/19 "2022-02-04T10:29:28Z")
Here is my benchmark for two runs of each:
```julia
Turing.setadbackend(:forwarddiff)
Iterations = 1001:1:2000
Number of chains = 4
Samples per chain = 1000
Wall duration = 102.14 seconds
Compute duration = 361.99 seconds
Iterations = 1001:1:2000
Number of chains = 4
Samples per chain = 1000
Wall duration = 95.9 seconds
Compute duration = 333.32 seconds
```
```julia
Turing.setadbackend(:tracker)
Iterations = 1001:1:2000
Number of chains = 4
Samples per chain = 1000
Wall duration = 99.84 seconds
Compute duration = 331.42 seconds
Iterations = 1001:1:2000
Number of chains = 4
Samples per chain = 1000
Wall duration = 86.06 seconds
Compute duration = 298.54 seconds
```
```julia
Turing.setadbackend(:zygote)
# gave up after a couple of minutes; will let it run over the weekend
```
```julia
Turing.setadbackend(:reversediff)
Turing.setrdcache(true)
Iterations = 1001:1:2000
Number of chains = 4
Samples per chain = 1000
Wall duration = 26.13 seconds
Compute duration = 101.67 seconds
Iterations = 1001:1:2000
Number of chains = 4
Samples per chain = 1000
Wall duration = 5.85 seconds
Compute duration = 20.05 seconds
```
---
### Author:  [@danielw2904](https://discourse.julialang.org/u/danielw2904)
#### Post date: [February 4, 2022, 10:30am UTC](https://discourse.julialang.org/t/turing-jl-for-causal-inference-model/75650/20 "2022-02-04T10:30:06Z")
Yep reversediff seems to be the clear winner in my benchmark