PyStan refactor branch

Developers
1

@ariddell, I’ve got most of it working! I’m on this branch:
feature/refactor-services

Here’s how I’m building it:

rm pystan/*.so
python3 setup.py build_ext --inplace

And here’s what I’m using to test:

import pystan

schools_code = """
data {
    int<lower=0> J; // number of schools
    real y[J]; // estimated treatment effects
    real<lower=0> sigma[J]; // s.e. of effect estimates
}
parameters {
    real mu;
    real<lower=0> tau;
    real eta[J];
}
transformed parameters {
    real theta[J];
    for (j in 1:J)
    theta[j] = mu + tau * eta[j];
}
model {
    eta ~ normal(0, 1);
    y ~ normal(theta, sigma);
}
"""

schools_dat = {'J': 8,
               'y': [28,  8, -3,  7, -1,  1, 18, 12],
               'sigma': [15, 10, 16, 11,  9, 11, 10, 18]}



fit = pystan.stan(model_code=schools_code, data=schools_dat, verbose=True, seed=1, chains=1, algorithm = "NUTS", control = dict(adapt_engaged = True, metric = "dense_e"))

## 1. diagnose: FIXME
pystan.stan(fit = fit, data=schools_dat, test_grad = True, chains = 1)

## 2: optimize newton
sm = pystan.StanModel(model_code=schools_code)
op = sm.optimizing(data=schools_dat, algorithm = "Newton")

## 3: optimize bfgs
op = sm.optimizing(data=schools_dat, algorithm = "BFGS")

## 4: optimize lbfgs
op = sm.optimizing(data=schools_dat, algorithm = "LBFGS")
                               
## 5: sample fixed_param: FIXME
fit = pystan.stan(fit = fit, data=schools_dat, algorithm = "Fixed_param")

## 6: sample nuts dense_e adapt_false
fit = pystan.stan(fit = fit, data=schools_dat, algorithm = "NUTS", control = dict(metric = "dense_e", adapt_engaged = False))

## 7: sample nuts dense_e adapt_true
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "NUTS", control = dict(metric = "dense_e", adapt_engaged = True))

## 8: sample nuts diag_e adapt_false
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "NUTS", control = dict(metric = "diag_e", adapt_engaged = False))

## 9: sample nuts diag_e adapt_true
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "NUTS", control = dict(metric = "diag_e", adapt_engaged = True))

## 10: sample nuts unit_e adapt_false
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "NUTS", control = dict(metric = "unit_e", adapt_engaged = False))

## 11: sample nuts unit_e adapt_true
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "NUTS", control = dict(metric = "unit_e", adapt_engaged = True))

## 12: sample static dense_e adapt_false
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "HMC", control = dict(metric = "dense_e", adapt_engaged = False))

## 13: sample static dense_e adapt_true
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "HMC", control = dict(metric = "dense_e", adapt_engaged = True))

## 14: sample static diag_e adapt_false
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "HMC", control = dict(metric = "diag_e", adapt_engaged = False))

## 15: sample static diag_e adapt_true
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "HMC", control = dict(metric = "diag_e", adapt_engaged = True))

## 16: sample static unit_e adapt_false
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "HMC", control = dict(metric = "unit_e", adapt_engaged = False))

## 17: sample static unit_e adapt_true
fit = pystan.stan(fit = fit, data=schools_dat,  algorithm = "HMC", control = dict(metric = "unit_e", adapt_engaged = True))

## 18: variational fullrank
sm.vb(data=schools_dat, algorithm = "fullrank")

## 19: variational meanfield
sm.vb(data=schools_dat, algorithm = "meanfield")

I’m having trouble with:

  • test_grad. The default arguments aren’t propagating into the c++. I think we saw this before. It runs the code, but the epsilon and error variables are both 0 instead of 1e-6.
  • use of sample_file. I tried setting it, but it doesn’t get set. It’s always writing to the temp directory. I don’t think this has anything to do with what I’ve done, but I can’t be sure.

I haven’t tested as thoroughly as RStan, but that’s cause I’m really fumbling around with Python. @ariddell, mind taking a look? Most of the functionality looks ok as far as I can tell.

2

Glad to hear things are working.

I took a quick look and try to find time to review in more depth.

re: test_grad, I’m not sure I ever tested test_grad with arguments in
PyStan in the first instance – so it might not be a problem with your code.

Is RStan getting close to being done? Do you have a deadline in mind for
merging the refactor?

3

I think as soon as we have CmdStan, RStan, and PyStan working, we should merge Stan.

RStan is done and being tested now. See branch: feature/refactor-stan-services

The last thing I need to work through in the interfaces are the defaults propagating from Stan. Everything else is good.