tpt-sci-ppl 0.1.0

Probabilistic programming model/DSL and NUTS sampler for the tpt-science pillar (built from scratch)
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
//! A small probabilistic-programming model/DSL built on top of the from-scratch
//! [`crate::nuts()`] sampler.
//!
//! A [`ModelBuilder`] collects parameter priors (expressed as reverse-mode
//! differentiable closures, so the NUTS gradient is exact) and a likelihood,
//! then [`ModelBuilder::build`] freezes them into a model and [`Model::fit`] runs NUTS
//! to draw posterior samples.
//!
//! ```rust
//! use tpt_sci_ppl::{ModelBuilder, Rng};
//! use tpt_math_prob_core::SplitMix64;
//!
//! // Posterior of mu given data ~ N(mu, 1) with a weak N(0, 5) prior.
//! let data = [2.0, 3.0, 2.5, 3.5, 3.0];
//! let mut m = ModelBuilder::new();
//! m.gaussian_parameter(0.0, 5.0);
//! m.set_data(data.to_vec());
//! m.likelihood(|_tape, v, d| {
//!     let mut s = _tape.constant(0.0);
//!     for &x in d.iter() {
//!         let z = (v[0] - x) / 1.0;
//!         s = s + (-0.5 * z * z);
//!     }
//!     s
//! });
//! let model = m.build().unwrap();
//! let mut rng = SplitMix64::seed_from_u64(1);
//! let trace = model.fit(&mut rng, 500).unwrap();
//! let samples = trace.combined_samples();
//! let mean: f64 = samples.iter().map(|s| s[0]).sum::<f64>() / samples.len() as f64;
//! assert!((mean - 2.8).abs() < 0.6);
//! ```

use tpt_math_autodiff_rev::{GradientTape, Variable};
use tpt_math_prob_core::Rng;

use crate::error::PplError;
use crate::nuts::{nuts_with_divergences, randn};
use crate::trace::Trace;
use crate::{Beta, Gamma, Gaussian};

/// A prior over one scalar parameter, written in terms of its `autodiff`
/// variable. Stored boxed so a `Model` can hold a heterogeneous list.
type PriorFn = Box<dyn for<'t> Fn(&'t Variable<'t>) -> Variable<'t>>;

/// A likelihood over all parameters and the observed data, written in terms of
/// `autodiff` variables.
type LikelihoodFn = Box<dyn for<'t> Fn(&'t GradientTape, &[Variable<'t>], &[f64]) -> Variable<'t>>;

/// Builder for a [`Model`].
pub struct ModelBuilder {
    dim: usize,
    priors: Vec<PriorFn>,
    likelihood: Option<LikelihoodFn>,
    data: Vec<f64>,
}

impl ModelBuilder {
    /// Create an empty model (zero parameters, no data).
    pub fn new() -> Self {
        Self {
            dim: 0,
            priors: Vec::new(),
            likelihood: None,
            data: Vec::new(),
        }
    }

    /// Add a parameter with an arbitrary differentiable prior closure.
    pub fn parameter<F>(&mut self, prior: F) -> &mut Self
    where
        F: for<'t> Fn(&'t Variable<'t>) -> Variable<'t> + 'static,
    {
        self.priors.push(Box::new(prior));
        self.dim += 1;
        self
    }

    /// Add a Gaussian prior `N(mean, std²)` (constant terms dropped — they do
    /// not affect sampling).
    pub fn gaussian_parameter(&mut self, mean: f64, std: f64) -> &mut Self {
        self.parameter(move |x: &Variable<'_>| {
            let x = *x;
            let z = (x - mean) / std;
            -0.5 * z * z
        })
    }

    /// Add a Beta prior `Beta(a, b)` (for a parameter in `(0, 1)`).
    pub fn beta_parameter(&mut self, a: f64, b: f64) -> &mut Self {
        self.parameter(move |x: &Variable<'_>| {
            let x = *x;
            (a - 1.0) * x.ln() + (b - 1.0) * (1.0_f64 - x).ln()
        })
    }

    /// Add a Gamma prior `Gamma(a, b)` (rate parameterisation, for a parameter
    /// in `(0, ∞)`).
    pub fn gamma_parameter(&mut self, a: f64, b: f64) -> &mut Self {
        self.parameter(move |x: &Variable<'_>| {
            let x = *x;
            (a - 1.0) * x.ln() - b * x
        })
    }

    /// Add a Gaussian prior derived from a typed [`Gaussian`] distribution
    /// `N(mean, std²)` (from `tpt-math-prob-bayes`). Equivalent to
    /// [`ModelBuilder::gaussian_parameter`] but takes the distribution struct
    /// directly, so a prior can be constructed and validated once and reused
    /// across models.
    pub fn gaussian_prior(&mut self, d: Gaussian) -> &mut Self {
        let (mean, std) = (d.mean(), d.std());
        self.parameter(move |x: &Variable<'_>| {
            let x = *x;
            let z = (x - mean) / std;
            -0.5 * z * z
        })
    }

    /// Add a Beta prior derived from a typed [`Beta`] distribution on `(0, 1)`
    /// (from `tpt-math-prob-bayes`). Equivalent to
    /// [`ModelBuilder::beta_parameter`] but takes the distribution struct.
    pub fn beta_prior(&mut self, d: Beta) -> &mut Self {
        let (a, b) = (d.alpha(), d.beta());
        self.parameter(move |x: &Variable<'_>| {
            let x = *x;
            (a - 1.0) * x.ln() + (b - 1.0) * (1.0_f64 - x).ln()
        })
    }

    /// Add a Gamma prior derived from a typed [`Gamma`] distribution in
    /// shape/rate parameterisation on `(0, ∞)` (from `tpt-math-prob-bayes`).
    /// Equivalent to [`ModelBuilder::gamma_parameter`] but takes the
    /// distribution struct.
    pub fn gamma_prior(&mut self, d: Gamma) -> &mut Self {
        let (a, b) = (d.shape(), d.rate());
        self.parameter(move |x: &Variable<'_>| {
            let x = *x;
            (a - 1.0) * x.ln() - b * x
        })
    }

    /// Set the observed data passed to the likelihood closure.
    pub fn set_data(&mut self, data: Vec<f64>) -> &mut Self {
        self.data = data;
        self
    }

    /// Set the likelihood `log p(data | params)`, written in terms of the
    /// parameter `autodiff` variables and the observed `data`.
    pub fn likelihood<F>(&mut self, f: F) -> &mut Self
    where
        F: for<'t> Fn(&'t GradientTape, &[Variable<'t>], &[f64]) -> Variable<'t> + 'static,
    {
        self.likelihood = Some(Box::new(f));
        self
    }

    /// Freeze the model into a [`Model`].
    ///
    /// # Errors
    ///
    /// Returns [`PplError::InvalidOption`] if no parameters were added.
    pub fn build(self) -> Result<Model, PplError> {
        if self.dim == 0 {
            return Err(PplError::InvalidOption("model has no parameters".into()));
        }
        Ok(Model {
            dim: self.dim,
            priors: self.priors,
            likelihood: self.likelihood,
            data: self.data,
        })
    }
}

impl Default for ModelBuilder {
    fn default() -> Self {
        Self::new()
    }
}

/// A frozen probabilistic model: parameter priors, likelihood, and data.
pub struct Model {
    dim: usize,
    priors: Vec<PriorFn>,
    likelihood: Option<LikelihoodFn>,
    data: Vec<f64>,
}

impl Model {
    /// Number of parameters.
    pub fn dim(&self) -> usize {
        self.dim
    }

    /// Build the differentiable target and draw `n_samples` posterior samples
    /// with NUTS, starting from `0.5` in every coordinate (a safe interior
    /// point for both bounded and unbounded priors).
    ///
    /// This is a convenience wrapper that runs [`Model::fit_chains`] with two
    /// dispersed chains, so the returned [`Trace`] always carries a meaningful
    /// split-R-hat ([`Trace::rhat`] is `NaN` for a single chain). Call
    /// `fit_chains` directly for finer control over the number of chains or
    /// `fit_from` for a single, explicitly-seeded chain.
    ///
    /// # Errors
    ///
    /// Propagates any error from the underlying sampler.
    pub fn fit(&self, rng: &mut impl Rng, n_samples: usize) -> Result<Trace, PplError> {
        self.fit_chains(2, n_samples, rng)
    }

    /// Draw samples starting from an explicit initial point.
    ///
    /// # Errors
    ///
    /// Returns [`PplError::DimensionMismatch`] if `initial.len()` does not
    /// match the model dimension.
    pub fn fit_from(
        &self,
        initial: &[f64],
        rng: &mut impl Rng,
        n_samples: usize,
    ) -> Result<Trace, PplError> {
        if initial.len() != self.dim {
            return Err(PplError::DimensionMismatch {
                expected: self.dim,
                got: initial.len(),
            });
        }
        // A callable that is higher-ranked over the tape lifetime: it can be
        // invoked with `&'t GradientTape` and `&'t [Variable<'t>]` for *any*
        // `'t`, which is exactly what the NUTS gradient machinery requires.
        let target = ModelTarget {
            priors: &self.priors,
            likelihood: &self.likelihood,
            data: &self.data,
        };
        let (samples, n_div) = nuts_with_divergences(&target, initial, rng, n_samples)?;
        Ok(Trace::new(vec![samples], n_div))
    }

    /// Draw `n_samples` posterior samples from `n_chains` independently
    /// initialised NUTS chains. The chains start from dispersed points around
    /// `0.5` so that [`Trace::rhat`] (split-R-hat across chains) is a meaningful
    /// convergence diagnostic.
    ///
    /// # Errors
    ///
    /// Returns [`PplError::InvalidOption`] if `n_chains == 0`, or any error from
    /// the underlying sampler.
    pub fn fit_chains(
        &self,
        n_chains: usize,
        n_samples: usize,
        rng: &mut impl Rng,
    ) -> Result<Trace, PplError> {
        if n_chains == 0 {
            return Err(PplError::InvalidOption("n_chains must be > 0".into()));
        }
        let base = vec![0.5; self.dim];
        let target = ModelTarget {
            priors: &self.priors,
            likelihood: &self.likelihood,
            data: &self.data,
        };
        let mut chains = Vec::with_capacity(n_chains);
        let mut n_div = 0usize;
        for _ in 0..n_chains {
            // Disperse the start so the chains explore from different regions;
            // a single shared start would make R-hat trivially ~1.
            let start: Vec<f64> = base.iter().map(|&v| v + 0.1 * randn(rng)).collect();
            let (samples, d) = nuts_with_divergences(&target, &start, rng, n_samples)?;
            n_div += d;
            chains.push(samples);
        }
        Ok(Trace::new(chains, n_div))
    }
}

/// Bridges a frozen model's priors/likelihood/data into a differentiable
/// target usable by [`nuts`]. The [`Target`] trait is implemented with a
/// single higher-ranked method so the autodiff tape lifetime is free, which
/// keeps this on stable Rust (no manual `Fn`/`extern "rust-call"` impls).
struct ModelTarget<'a> {
    priors: &'a [PriorFn],
    likelihood: &'a Option<LikelihoodFn>,
    data: &'a [f64],
}

impl<'a> crate::nuts::Target for ModelTarget<'a> {
    fn eval<'t>(&self, tape: &'t GradientTape, vars: &'t [Variable<'t>]) -> Variable<'t> {
        let mut s = tape.constant(0.0);
        for (i, p) in self.priors.iter().enumerate() {
            s += p(&vars[i]);
        }
        if let Some(l) = self.likelihood {
            s += l(tape, vars, self.data);
        }
        s
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use approx::assert_abs_diff_eq;
    use tpt_math_prob_core::SplitMix64;

    #[test]
    fn gaussian_model_recovers_posterior_mean() {
        // mu ~ N(0, 5),  data | mu ~ N(mu, 1)
        let data = [2.0, 3.0, 2.5, 3.5, 3.0];
        let mut m = ModelBuilder::new();
        m.gaussian_parameter(0.0, 5.0);
        m.set_data(data.to_vec());
        m.likelihood(|t, v, d| {
            let mut s = t.constant(0.0);
            for &x in d.iter() {
                let z = (v[0] - x) / 1.0;
                s += -0.5 * z * z;
            }
            s
        });
        let model = m.build().unwrap();
        let mut rng = SplitMix64::seed_from_u64(1);
        let trace = model.fit(&mut rng, 800).unwrap();
        let samples = trace.combined_samples();
        let mean: f64 = samples.iter().map(|s| s[0]).sum::<f64>() / samples.len() as f64;
        // Data mean is 2.8; weak prior lets the posterior sit near it.
        assert_abs_diff_eq!(mean, 2.8, epsilon = 0.6);
        assert!(mean > 1.0 && mean < 4.0);
    }

    #[test]
    fn fit_from_rejects_wrong_dimension() {
        let mut m = ModelBuilder::new();
        m.gaussian_parameter(0.0, 1.0);
        let model = m.build().unwrap();
        let mut rng = SplitMix64::seed_from_u64(1);
        assert!(matches!(
            model.fit_from(&[0.0, 0.0], &mut rng, 10),
            Err(PplError::DimensionMismatch { .. })
        ));
    }

    #[test]
    fn typed_gaussian_prior_matches_raw_prior() {
        // A Gaussian prior built from the typed `tpt-math-prob-bayes`
        // distribution must recover the same posterior as the raw
        // `(mean, std)` prior.
        let data = [2.0, 3.0, 2.5, 3.5, 3.0];
        let mut m = ModelBuilder::new();
        m.gaussian_prior(Gaussian::new(0.0, 5.0));
        m.set_data(data.to_vec());
        m.likelihood(|t, v, d| {
            let mut s = t.constant(0.0);
            for &x in d.iter() {
                let z = (v[0] - x) / 1.0;
                s += -0.5 * z * z;
            }
            s
        });
        let model = m.build().unwrap();
        let mut rng = SplitMix64::seed_from_u64(1);
        let trace = model.fit(&mut rng, 800).unwrap();
        let samples = trace.combined_samples();
        let mean: f64 = samples.iter().map(|s| s[0]).sum::<f64>() / samples.len() as f64;
        assert_abs_diff_eq!(mean, 2.8, epsilon = 0.6);
    }

    #[test]
    fn typed_beta_prior_recovers_posterior() {
        // Posterior of p (a rate in (0, 1)) given 7 successes / 3 failures
        // under a uniform Beta(1, 1) prior. The posterior mean is 0.7.
        let mut m = ModelBuilder::new();
        m.beta_prior(Beta::uniform());
        m.set_data(vec![7.0, 3.0]);
        m.likelihood(|_t, v, d| {
            let s = d[0];
            let f = d[1];
            s * v[0].ln() + f * (1.0_f64 - v[0]).ln()
        });
        let model = m.build().unwrap();
        let mut rng = SplitMix64::seed_from_u64(42);
        let trace = model.fit(&mut rng, 2000).unwrap();
        let samples = trace.combined_samples();
        let mean: f64 = samples.iter().map(|s| s[0]).sum::<f64>() / samples.len() as f64;
        assert_abs_diff_eq!(mean, 0.7, epsilon = 0.06);
    }

    #[test]
    fn multi_chain_fit_reports_diagnostics() {
        // mu ~ N(0, 5), data | mu ~ N(mu, 1), data mean 2.8.
        let data = [2.0, 3.0, 2.5, 3.5, 3.0];
        let mut m = ModelBuilder::new();
        m.gaussian_parameter(0.0, 5.0);
        m.set_data(data.to_vec());
        m.likelihood(|t, v, d| {
            let mut s = t.constant(0.0);
            for &x in d.iter() {
                let z = (v[0] - x) / 1.0;
                s += -0.5 * z * z;
            }
            s
        });
        let model = m.build().unwrap();
        let mut rng = SplitMix64::seed_from_u64(123);
        let trace = model.fit_chains(4, 800, &mut rng).unwrap();
        assert_eq!(trace.n_chains(), 4);
        assert_eq!(trace.dim(), 1);
        // Converged chains give split-R-hat near 1.
        let rhat = trace.rhat(0);
        assert!(rhat.is_finite() && rhat < 1.1, "rhat = {rhat}");
        // ESS should be a positive fraction of the raw draw count.
        let ess = trace.ess(0);
        assert!(
            ess > 0.0 && ess < trace.n_draws() as f64 + 1.0,
            "ess = {ess}"
        );
        // Divergence rate is well defined (0.0 for this well-behaved target).
        assert!((0.0..=1.0).contains(&trace.divergence_rate()));
        // The posterior mean is recovered.
        assert_abs_diff_eq!(trace.mean(0), 2.8, epsilon = 0.6);
    }

    #[test]
    fn fit_chains_rejects_zero_chains() {
        let mut m = ModelBuilder::new();
        m.gaussian_parameter(0.0, 1.0);
        let model = m.build().unwrap();
        let mut rng = SplitMix64::seed_from_u64(1);
        assert!(matches!(
            model.fit_chains(0, 10, &mut rng),
            Err(PplError::InvalidOption { .. })
        ));
    }
}