SearchLibrium 0.0.261__tar.gz → 0.0.262__tar.gz

{searchlibrium-0.0.261/src/SearchLibrium.egg-info → searchlibrium-0.0.262}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: SearchLibrium
-Version: 0.0.261
+Version: 0.0.262
 Summary: Automated specification search for discrete choice models (logit, mixed logit, nested logit, RRM and more) using metaheuristic optimisation
 Author: Zeke Ahern, Mario Fernando Taco Morales, Alexander Paz, Prithvi Beeramole, Robert Burdett
 Maintainer-email: Zeke Ahern <z.ahern@qut.edu.au>

{searchlibrium-0.0.261 → searchlibrium-0.0.262}/pyproject.toml

@@ -132,7 +132,7 @@ Documentation = "https://pypi.org/project/SearchLibrium/"
 realpython = "SearchLibrium.__main__:main"
 
 [tool.bumpver]
-current_version = "0.0.261"
+current_version = "0.0.262"
 version_pattern = "MAJOR.MINOR.PATCH"
 commit_message = "[skip ci] Bump version {old_version} -> {new_version}"
 commit = true

{searchlibrium-0.0.261 → searchlibrium-0.0.262}/src/SearchLibrium/MixedLogit.py

@@ -122,10 +122,19 @@ def _panel_weighted_shares(prob_np, panel_info):
     return ind, np.mean(ind, axis=0)
 
 class MixedLogit(DiscreteChoiceModel):
-    def __init__(self, halton_opts=None, distributions=['n', 'ln', 'nln', 't', 'tn', 'u'], _jax=None):
+    def __init__(self, halton_opts=None, distributions=['n', 'ln', 'nln', 't', 'tn', 'u'], _jax=None,
+                 save_prob_full=None):
         super().__init__(_jax)
         self.descr = "MXL"
         self.halton_opts = halton_opts
+        # MEMORY: prob_full is the full (N, P, J, R) probability tensor built in
+        # post-process for the summary/plots. For a 13k-case/20-alt/1000-draw
+        # model that is ~2 GB, and it is never read during estimation, so it is
+        # now opt-in. Pass save_prob_full=True (or set
+        # SEARCHLIBRIUM_SAVE_PROB_FULL=1) to retain the previous behaviour.
+        if save_prob_full is None:
+            save_prob_full = bool(os.environ.get("SEARCHLIBRIUM_SAVE_PROB_FULL", ""))
+        self.save_prob_full = bool(save_prob_full)
         self.draws_generator = Draws(k=len(distributions), halton_opts=halton_opts, rvdist=distributions)
         self.random_parameters = RandomParameters(distributions or [])
         self.softmax_r = True
@@ -1173,7 +1182,8 @@ class MixedLogit(DiscreteChoiceModel):
                 self.choice_pred_prob = np.mean(p, axis=3)
                 self.ind_pred_prob, self.pred_prob = _panel_weighted_shares(
                     self.choice_pred_prob, self.panel_info)
-                self.prob_full = p
+                if self.save_prob_full:
+                    self.prob_full = p
                 self._refresh_corr_from_varlist()
                 self.post_process(jax_result, self.Xnames, self.N)
                 return
@@ -1251,7 +1261,8 @@ class MixedLogit(DiscreteChoiceModel):
             self.ind_pred_prob, self.pred_prob = _panel_weighted_shares(
                 self.choice_pred_prob, self.panel_info)
 
-            self.prob_full = p
+            if self.save_prob_full:
+                self.prob_full = p
         # }
         self._refresh_corr_from_varlist()
         self.post_process(result, self.Xnames, self.N)
@@ -1499,7 +1510,11 @@ class MixedLogit(DiscreteChoiceModel):
     # process; during a heterogeneous GA search that is one retained model
     # per distinct parameter-count combination.
     # ------------------------------------------------------------------
-    _JIT_CACHE_MAXSIZE = 16
+    # MEMORY: cache size is in *live models*, each of which can pin its X/y
+    # arrays and a compiled XLA executable. The old default of 16 was tuned for
+    # speed; 4 keeps most of the reuse benefit while capping the worst case.
+    # Override with SEARCHLIBRIUM_JIT_CACHE_MAXSIZE (0 disables the cache).
+    _JIT_CACHE_MAXSIZE = max(0, int(os.environ.get("SEARCHLIBRIUM_JIT_CACHE_MAXSIZE", "4")))
     _mxl_jit_cache: "OrderedDict[tuple, tuple]" = OrderedDict()
 
     @staticmethod
@@ -1562,6 +1577,17 @@ class MixedLogit(DiscreteChoiceModel):
             pass
         gc.collect()
 
+    def release_memory(self):
+        """Free per-fit caches and large post-process tensors this model holds.
+
+        Safe to call between metaheuristic candidates: it clears the process-
+        wide JIT/XLA cache and drops ``prob_full`` unless it was explicitly
+        saved. Fitted coefficients, standard errors and covariance are kept.
+        """
+        if not getattr(self, "save_prob_full", False):
+            self.prob_full = None
+        type(self)._clear_jit_cache()
+
     def _run_jaxopt(self, betas, X_jax, y_jax, pi_jax, draws_jax,
                     _fn, _cache_key):
         """JAX likelihood + ``jaxopt.LBFGS`` solver (experiment).

{searchlibrium-0.0.261 → searchlibrium-0.0.262}/src/SearchLibrium/latent_class.py

@@ -1780,6 +1780,7 @@ class LatentClassMixedLogit(DiscreteChoiceModel):
         member_params_spec=None,
         class_params_spec=None,
         panels=None,
+        retain_all=False,
         ident_penalty_kwargs=None,
         **kwargs,
     ):
@@ -1797,8 +1798,17 @@ class LatentClassMixedLogit(DiscreteChoiceModel):
             Variables used in the class membership equation.
         member_params_spec : list or array, optional
             Per-class membership variable specification.
+        retain_all : bool
+            MEMORY: when False (default) only the best model is kept alive and
+            the returned list contains just that model; every intermediate
+            class-count model is released as the search advances. Each model
+            otherwise pins its X/y/draws design, JAX device copies and cached
+            JIT objectives, so retaining all of them across a GA search leaks
+            hundreds of MB per candidate. Set True to recover the previous
+            behaviour (all models returned and stored on
+            ``best_model.search_results``).
         """
-        fitted_models = []
+        fitted_models = [] if retain_all else None
         best_model = None
         _best_nsig = float('inf')
         prev_model = None
@@ -1833,7 +1843,8 @@ class LatentClassMixedLogit(DiscreteChoiceModel):
                 de_seed=de_seed,
             )
             model.get_loglik_null()
-            fitted_models.append(model)
+            if fitted_models is not None:
+                fitted_models.append(model)
 
             # Pareto-style significance prioritisation:
             #  1) prefer fewer insignificant variable groups,
@@ -1858,8 +1869,14 @@ class LatentClassMixedLogit(DiscreteChoiceModel):
                 _best_nsig = nsig
             prev_model = model
 
-        best_model.search_results = fitted_models
-        return best_model, fitted_models
+        if fitted_models is not None:
+            best_model.search_results = fitted_models
+            return best_model, fitted_models
+        # Default path: keep only the best model alive. Drop the warm-start
+        # chain's trailing reference so intermediate models can be collected.
+        best_model.search_results = None
+        prev_model = None
+        return best_model, [best_model]
 
     # ------------------------------------------------------------------
     # Internal helpers for numerical standard errors

{searchlibrium-0.0.261 → searchlibrium-0.0.262}/src/SearchLibrium/mixed_logit.py

@@ -54,6 +54,7 @@ in a panel or longitudinal dataset.
 ''' LIBRARIES                                                  '''
 ''' ---------------------------------------------------------- '''
 import itertools
+import os
 import numpy as np
 import scipy.stats as ss
 from scipy.optimize import minimize
@@ -273,6 +274,8 @@ class MixedLogit(DiscreteChoiceModel):
         self.rvdist = None          # Distributions for randvar
         self.correlated_vars, self.halton_opts = None, None
         self.batch_size, self.minimise_func = None, None
+        if not hasattr(self, "save_prob_full"):
+            self.save_prob_full = bool(os.environ.get("SEARCHLIBRIUM_SAVE_PROB_FULL", ""))
     # }
 
     ''' -------------------------------------------------------------------- '''
@@ -781,11 +784,17 @@ class MixedLogit(DiscreteChoiceModel):
             self.ind_pred_prob, self.pred_prob = _panel_weighted_shares(
                 self.choice_pred_prob, self.panel_info)
 
-            self.prob_full = p
+            if self.save_prob_full:
+                self.prob_full = p
         # }
         self.post_process(result, self.Xnames, self.N)
     # }
 
+    def release_memory(self):
+        """Drop prob_full if not explicitly saved. Safe between GA candidates."""
+        if not getattr(self, "save_prob_full", False):
+            self.prob_full = None
+
     ''' ---------------------------------------------------------- '''
     ''' Function. This code segments the betas array into different 
         parts and stores these parts in var_list under keys from 

{searchlibrium-0.0.261 → searchlibrium-0.0.262}/src/SearchLibrium/ordered_logit.py

@@ -69,6 +69,8 @@ from scipy import stats
 from scipy.optimize import minimize
 
 import inspect
+import os
+from collections import OrderedDict
 
 minval = 1E-30
 
@@ -2481,7 +2483,15 @@ class SA_ORDLOG():
         self.mod = OrderedLogit(X=X, y=y, J=J, distr='logit', start=None, fit_intercept=True)
 
         # SA specific parameters
-        self.archive = {}  # Define a dictionary
+        # MEMORY: the archive memoises spec -> loglik so repeated specs skip
+        # the expensive refit. It used to be an unbounded dict, so a long run
+        # retained one entry per distinct predictor mask (2**K in the limit).
+        # It is now an LRU-bounded OrderedDict; evicting an entry only costs a
+        # future refit, never correctness. Set
+        # SEARCHLIBRIUM_ORDLOG_ARCHIVE_MAX=0 to restore unbounded behaviour.
+        self.archive = OrderedDict()
+        self.archive_maxsize = int(
+            os.environ.get("SEARCHLIBRIUM_ORDLOG_ARCHIVE_MAX", "50000"))
         self.tI = 100
         self.tF = 0.001
         self.t = self.tI
@@ -2544,6 +2554,12 @@ class SA_ORDLOG():
             print("Solution already seen")
             candidate_score = self.archive[key]
 
+        # LRU touch + bounded eviction (see archive_maxsize above).
+        self.archive.move_to_end(key)
+        _amax = getattr(self, "archive_maxsize", 0)
+        while _amax and len(self.archive) > _amax:
+            self.archive.popitem(last=False)
+
         # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
         if accept_change(self.current_score, candidate_score, self.t):

{searchlibrium-0.0.261 → searchlibrium-0.0.262}/src/SearchLibrium/test_lc_de.py

@@ -231,6 +231,7 @@ def run_class_search(
         max_classes=max_classes,
         criterion="bic",
         warm_start=True,
+        retain_all=True,
         de_init=de_init,
         de_popsize=de_popsize,
         de_maxiter=de_maxiter,

searchlibrium-0.0.262/src/SearchLibrium/version.txt

@@ -0,0 +1 @@
+0.0.262

{searchlibrium-0.0.261 → searchlibrium-0.0.262/src/SearchLibrium.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: SearchLibrium
-Version: 0.0.261
+Version: 0.0.262
 Summary: Automated specification search for discrete choice models (logit, mixed logit, nested logit, RRM and more) using metaheuristic optimisation
 Author: Zeke Ahern, Mario Fernando Taco Morales, Alexander Paz, Prithvi Beeramole, Robert Burdett
 Maintainer-email: Zeke Ahern <z.ahern@qut.edu.au>

searchlibrium-0.0.261/src/SearchLibrium/version.txt

@@ -1 +0,0 @@
-0.0.261