Robust generalized bart inference (#4709)

* robust generalized bart inference

* update docstring

* update test

* clarify role of link function

* revert few changes

* raise a ValueError if inv_link string is not valid

* update release notes

RELEASE-NOTES.md

@@ -7,7 +7,7 @@
 + Fix bug in the computation of the log pseudolikelihood values (SMC-ABC). (see [#4672](https://github.com/pymc-devs/pymc3/pull/4672)).
 
 ### New Features
-+ BART with non-gaussian likelihoods (see [#4675](https://github.com/pymc-devs/pymc3/pull/4675)).
++ BART with non-gaussian likelihoods (see [#4675](https://github.com/pymc-devs/pymc3/pull/4675) and [#4709](https://github.com/pymc-devs/pymc3/pull/4709)).
 
 ## PyMC3 3.11.2 (14 March 2021)
 

pymc3/distributions/bart.py

@@ -15,6 +15,7 @@
 import numpy as np
 
 from pandas import DataFrame, Series
+from scipy.special import expit
 
 from pymc3.distributions.distribution import NoDistribution
 from pymc3.distributions.tree import LeafNode, SplitNode, Tree
@@ -30,7 +31,6 @@ def __init__(
         m=200,
         alpha=0.25,
         split_prior=None,
-        scale=None,
         inv_link=None,
         jitter=False,
         *args,
@@ -63,22 +63,32 @@ def __init__(
             )
         self.m = m
         self.alpha = alpha
-        self.y_std = Y.std()
-
-        if scale is None:
-            self.leaf_scale = NormalSampler(sigma=None)
-        elif isinstance(scale, (int, float)):
-            self.leaf_scale = NormalSampler(sigma=Y.std() / self.m ** scale)
 
         if inv_link is None:
-            self.inv_link = lambda x: x
+            self.inv_link = self.link = lambda x: x
+        elif isinstance(inv_link, str):
+            # The link function is just a rough approximation in order to allow the PGBART sampler
+            # to propose reasonable values for the leaf nodes.
+            if inv_link == "logistic":
+                self.inv_link = expit
+                self.link = lambda x: (x - 0.5) * 10
+            elif inv_link == "exp":
+                self.inv_link = np.exp
+                self.link = np.log
+                self.Y[self.Y == 0] += 0.0001
+            else:
+                raise ValueError("Accepted strings are 'logistic' or 'exp'")
         else:
-            self.inv_link = inv_link
+            self.inv_link, self.link = inv_link
+
+        self.init_mean = self.link(self.Y.mean())
+        self.Y_un = self.link(self.Y)
 
         self.num_observations = X.shape[0]
         self.num_variates = X.shape[1]
         self.available_predictors = list(range(self.num_variates))
         self.ssv = SampleSplittingVariable(split_prior, self.num_variates)
+        self.initial_value_leaf_nodes = self.init_mean / self.m
         self.trees = self.init_list_of_trees()
         self.all_trees = []
         self.mean = fast_mean()
@@ -96,7 +106,7 @@ def preprocess_XY(self, X, Y):
         return X, Y, missing_data
 
     def init_list_of_trees(self):
-        initial_value_leaf_nodes = self.Y.mean() / self.m
+        initial_value_leaf_nodes = self.initial_value_leaf_nodes
         initial_idx_data_points_leaf_nodes = np.array(range(self.num_observations), dtype="int32")
         list_of_trees = []
         for i in range(self.m):
@@ -110,7 +120,7 @@ def init_list_of_trees(self):
         # bartMachine: A Powerful Tool for Machine Learning in R. ArXiv e-prints, 2013
         # The sum_trees_output will contain the sum of the predicted output for all trees.
         # When R_j is needed we subtract the current predicted output for tree T_j.
-        self.sum_trees_output = np.full_like(self.Y, self.Y.mean())
+        self.sum_trees_output = np.full_like(self.Y, self.init_mean)
 
         return list_of_trees
 
@@ -181,14 +191,13 @@ def get_new_idx_data_points(self, current_split_node, idx_data_points):
 
     def get_residuals(self):
         """Compute the residuals."""
-        R_j = self.Y - self.inv_link(self.sum_trees_output)
-
+        R_j = self.Y_un - self.sum_trees_output
         return R_j
 
     def draw_leaf_value(self, idx_data_points):
         """Draw the residual mean."""
         R_j = self.get_residuals()[idx_data_points]
-        draw = self.mean(R_j) + self.leaf_scale.random()
+        draw = self.mean(R_j)
         return draw
 
     def predict(self, X_new):
@@ -278,24 +287,6 @@ def rvs(self):
                     return i
 
 
-class NormalSampler:
-    def __init__(self, sigma):
-        self.size = 5000
-        self.cache = []
-        self.sigma = sigma
-
-    def random(self):
-        if self.sigma is None:
-            return 0
-        else:
-            if not self.cache:
-                self.update()
-        return self.cache.pop()
-
-    def update(self):
-        self.cache = np.random.normal(loc=0.0, scale=self.sigma, size=self.size).tolist()
-
-
 class BART(BaseBART):
     """
     BART distribution.
@@ -317,23 +308,17 @@ class BART(BaseBART):
         Each element of split_prior should be in the [0, 1] interval and the elements should sum
         to 1. Otherwise they will be normalized.
         Defaults to None, all variable have the same a prior probability
-    scale : float
-        Controls the variance of the proposed leaf value. The leaf values are computed as a
-        Gaussian with mean equal to the conditional residual mean and variance proportional to
-        the variance of the response variable, and inversely proportional to the number of trees
-        and the scale parameter. Defaults to None, i.e the variance is 0.
-    inv_link : numpy function
-        Inverse link function defaults to None, i.e. the identity function.
+    inv_link : str or tuple of functions
+        Inverse link function defaults to None, i.e. the identity function. Accepted strings are
+        ``logistic`` or ``exp``.
     jitter : bool
         Whether to jitter the X values or not. Defaults to False. When values of X are repeated,
         jittering X has the effect of increasing the number of effective spliting variables,
         otherwise it does not have any effect.
     """
 
-    def __init__(
-        self, X, Y, m=200, alpha=0.25, split_prior=None, scale=None, inv_link=None, jitter=False
-    ):
-        super().__init__(X, Y, m, alpha, split_prior, scale, inv_link)
+    def __init__(self, X, Y, m=200, alpha=0.25, split_prior=None, inv_link=None, jitter=False):
+        super().__init__(X, Y, m, alpha, split_prior, inv_link)
 
     def _str_repr(self, name=None, dist=None, formatting="plain"):
         if dist is None:

pymc3/step_methods/pgbart.py

@@ -75,29 +75,29 @@ def __init__(self, vars=None, num_particles=10, max_stages=5000, chunk="auto", m
         self.log_num_particles = np.log(num_particles)
         self.indices = list(range(1, num_particles))
         self.max_stages = max_stages
-        self.old_trees_particles_list = []
-        for i in range(self.bart.m):
-            p = ParticleTree(self.bart.trees[i], self.bart.prior_prob_leaf_node)
-            self.old_trees_particles_list.append(p)
 
         shared = make_shared_replacements(vars, model)
         self.likelihood_logp = logp([model.datalogpt], vars, shared)
+        self.init_leaf_nodes = self.bart.initial_value_leaf_nodes
+        self.init_likelihood = self.likelihood_logp(self.bart.inv_link(self.bart.sum_trees_output))
+        self.init_log_weight = self.init_likelihood - self.log_num_particles
+        self.old_trees_particles_list = []
+        for i in range(self.bart.m):
+            p = ParticleTree(
+                self.bart.trees[i],
+                self.bart.prior_prob_leaf_node,
+                self.init_log_weight,
+                self.init_likelihood,
+            )
+            self.old_trees_particles_list.append(p)
         super().__init__(vars, shared)
 
     def astep(self, _):
         bart = self.bart
+
         inv_link = bart.inv_link
-        num_observations = bart.num_observations
         variable_inclusion = np.zeros(bart.num_variates, dtype="int")
 
-        # For the tunning phase we restrict max_stages to a low number, otherwise it is almost sure
-        # we will reach max_stages given that our first set of m trees is not good at all.
-        # Can set max_stages as a function of the number of variables/dimensions? XXX
-        if self.tune:
-            max_stages = 5
-        else:
-            max_stages = self.max_stages
-
         if self.idx == bart.m:
             self.idx = 0
 
@@ -110,25 +110,28 @@ def astep(self, _):
             bart.sum_trees_output -= old_prediction
             # Generate an initial set of SMC particles
             # at the end of the algorithm we return one of these particles as the new tree
-            particles = self.init_particles(tree.tree_id, num_observations, inv_link)
+            particles = self.init_particles(tree.tree_id)
 
-            for t in range(1, max_stages):
+            for t in range(1, self.max_stages):
                 # Get old particle at stage t
                 particles[0] = self.get_old_tree_particle(tree.tree_id, t)
                 # sample each particle (try to grow each tree)
                 for c in range(1, self.num_particles):
                     particles[c].sample_tree_sequential(bart)
                 # Update weights. Since the prior is used as the proposal,the weights
                 # are updated additively as the ratio of the new and old log_likelihoods
-                for p_idx, p in enumerate(particles):
-                    new_likelihood = self.likelihood_logp(inv_link(p.tree.predict_output()))
+                for p in particles:
+                    new_likelihood = self.likelihood_logp(
+                        inv_link(bart.sum_trees_output + p.tree.predict_output())
+                    )
                     p.log_weight += new_likelihood - p.old_likelihood_logp
                     p.old_likelihood_logp = new_likelihood
 
                 # Normalize weights
                 W, normalized_weights = self.normalize(particles)
                 # Resample all but first particle
                 re_n_w = normalized_weights[1:] / normalized_weights[1:].sum()
+
                 new_indices = np.random.choice(self.indices, size=len(self.indices), p=re_n_w)
                 particles[1:] = particles[new_indices]
 
@@ -149,8 +152,7 @@ def astep(self, _):
             new_tree = np.random.choice(particles, p=normalized_weights)
             self.old_trees_particles_list[tree.tree_id] = new_tree
             bart.trees[idx] = new_tree.tree
-            new_prediction = new_tree.tree.predict_output()
-            bart.sum_trees_output += new_prediction
+            bart.sum_trees_output += new_tree.tree.predict_output()
 
             if not self.tune:
                 self.iter += 1
@@ -194,26 +196,28 @@ def get_old_tree_particle(self, tree_id, t):
         old_tree_particle.set_particle_to_step(t)
         return old_tree_particle
 
-    def init_particles(self, tree_id, num_observations, inv_link):
+    def init_particles(self, tree_id):
         """
         Initialize particles
         """
-        # The first particle is from the tree we are trying to replace
         prev_tree = self.get_old_tree_particle(tree_id, 0)
-        likelihood = self.likelihood_logp(inv_link(prev_tree.tree.predict_output()))
+        likelihood = self.likelihood_logp(self.bart.inv_link(prev_tree.tree.predict_output()))
         prev_tree.old_likelihood_logp = likelihood
         prev_tree.log_weight = likelihood - self.log_num_particles
         particles = [prev_tree]
 
-        # The rest of the particles are identically initialized
-        initial_idx_data_points_leaf_nodes = np.arange(num_observations, dtype="int32")
+        initial_idx_data_points_leaf_nodes = np.arange(self.bart.num_observations, dtype="int32")
         new_tree = Tree.init_tree(
             tree_id=tree_id,
-            leaf_node_value=0,
+            leaf_node_value=self.init_leaf_nodes,
             idx_data_points=initial_idx_data_points_leaf_nodes,
         )
-        for i in range(1, self.num_particles):
-            particles.append(ParticleTree(new_tree, self.bart.prior_prob_leaf_node, 0, 0))
+
+        prior_prob = self.bart.prior_prob_leaf_node
+        for _ in range(1, self.num_particles):
+            particles.append(
+                ParticleTree(new_tree, prior_prob, self.init_log_weight, self.init_likelihood)
+            )
 
         return np.array(particles)
 

pymc3/tests/test_bart.py

@@ -1,7 +1,5 @@
 import numpy as np
 
-from scipy.special import expit
-
 import pymc3 as pm
 
 
@@ -106,7 +104,7 @@ def test_model():
 
     Y = np.repeat([0, 1], 50)
     with pm.Model() as model:
-        mu = pm.BART("mu", X, Y, m=50, inv_link=expit, scale=0.25)
+        mu = pm.BART("mu", X, Y, m=50, inv_link="logistic")
         y = pm.Bernoulli("y", mu, observed=Y)
         trace = pm.sample(1000, random_seed=212480)