Fix target model initialisation

CHANGELOG.rst

@@ -1,6 +1,7 @@
 Changelog
 =========
 
+- Update surrogate model initialisation to use all initial evidence
 - Use kernel copy to avoid pickle issue and allow BOLFI parallelisation with non-default kernel
 - Restrict matplotlib version < 3.9 for compatibility with GPy
 - Add option to use additive or multiplicative adjustment in any acquisition method

elfi/methods/inference/bolfi.py

@@ -95,7 +95,7 @@ def __init__(self,
         if precomputed is not None:
             params = batch_to_arr2d(precomputed, self.target_model.parameter_names)
             n_precomputed = len(params)
-            self.target_model.update(params, precomputed[target_name])
+            self.target_model.update(params, precomputed[target_name], optimize=True)
 
         self.batches_per_acquisition = batches_per_acquisition or self.max_parallel_batches
 
@@ -115,6 +115,10 @@ def __init__(self,
         self.state['last_GP_update'] = self.n_initial_evidence
         self.state['acquisition'] = []
 
+        if self.target_model.n_evidence < 1 and self.n_initial_evidence > 0:
+            self.init_x = np.zeros((self.n_initial_evidence, self.target_model.input_dim))
+            self.init_y = np.zeros((self.n_initial_evidence, 1))
+
     def _resolve_initial_evidence(self, initial_evidence):
         # Some sensibility limit for starting GP regression
         precomputed = None
@@ -215,10 +219,20 @@ def update(self, batch, batch_index):
         params = batch_to_arr2d(batch, self.target_model.parameter_names)
         self._report_batch(batch_index, params, batch[self.target_name])
 
-        optimize = self._should_optimize()
-        self.target_model.update(params, batch[self.target_name], optimize)
-        if optimize:
-            self.state['last_GP_update'] = self.target_model.n_evidence
+        if self.target_model.n_evidence < 1 and self.n_initial_evidence > 0:
+            # accumulate initialisation data
+            n = self.state['n_evidence']
+            self.init_x[n - self.batch_size:n] = params
+            self.init_y[n - self.batch_size:n] = batch[self.target_name].reshape(-1, 1)
+            if self.state['n_evidence'] >= self.n_initial_evidence:
+                # initialise model
+                self.target_model.update(self.init_x, self.init_y, optimize=True)
+        else:
+            # update model
+            optimize = self._should_optimize()
+            self.target_model.update(params, batch[self.target_name], optimize)
+            if optimize:
+                self.state['last_GP_update'] = self.state['n_evidence']
 
     def prepare_new_batch(self, batch_index):
         """Prepare values for a new batch.

elfi/methods/inference/bolfire.py

@@ -95,6 +95,8 @@ def __init__(self,
         # Initialize BO
         self.n_initial_evidence = self._resolve_n_initial_evidence(n_initial_evidence)
         self.acquisition_method = self._resolve_acquisition_method(acquisition_method)
+        self.init_x = np.zeros((self.n_initial_evidence, self.target_model.input_dim))
+        self.init_y = np.zeros((self.n_initial_evidence, 1))
 
         # Initialize state dictionary
         self.state['n_evidence'] = 0
@@ -390,10 +392,19 @@ def _process_simulated(self):
         # BO part
         self.state['n_evidence'] += 1
         parameter_values = self.current_params
-        optimize = self._should_optimize()
-        self.target_model.update(parameter_values, negative_log_ratio_value, optimize)
-        if optimize:
-            self.state['last_GP_update'] = self.target_model.n_evidence
+        if self.target_model.n_evidence < 1 and self.n_initial_evidence > 0:
+            # accumulate initialisation data
+            self.init_x[self.state['n_evidence'] - 1] = parameter_values
+            self.init_y[self.state['n_evidence'] - 1] = negative_log_ratio_value
+            if self.state['n_evidence'] >= self.n_initial_evidence:
+                # initialise model
+                self.target_model.update(self.init_x, self.init_y, optimize=True)
+        else:
+            # update model
+            optimize = self._should_optimize()
+            self.target_model.update(parameter_values, negative_log_ratio_value, optimize)
+            if optimize:
+                self.state['last_GP_update'] = self.target_model.n_evidence
 
     def _generate_training_data(self, likelihood, marginal):
         """Generate training data."""