Fix bootstrap fixed random state (#232)

* Fix bootstrap fixed random state (closes #231)

Use old-school optional type

Enable colour output in CI for pytest

Fix typo in arg name

Rethink strategy; provide way to fix random seed but still call resample multiple times

Be explicit about number of samples

* Attempt to stratify bootstrap if classification props are present

* Pass `targets_groups` up to Ensemble from member

* use first prop for stratification

---------

Co-authored-by: ppdebreuck <pierre-paul.debreuck@student.uclouvain.be>

.github/workflows/ci.yml

@@ -5,6 +5,11 @@ on:
   push:
     branches:
       - master
+
+env:
+  # Enable color output for pytest
+  FORCE_COLOR: true
+
 jobs:
 
   lint:

modnet/models/ensemble.py

@@ -47,6 +47,7 @@ def __init__(
         bootstrap=True,
         models=None,
         modnet_models=None,
+        random_state: Optional[int] = None,
         **kwargs,
     ):
         """
@@ -55,11 +56,13 @@ def __init__(
             n_models: number of inner MODNetModels, each model has the same architecture defined by the args nd kwargs.
             bootstrap: whether to bootstrap the samples for each inner MODNet fit.
             models: List of user provided MODNetModels. Enables to have different architectures. n_models is discarded in this case.
+            random_state: fix a random state for use with this model.
             modnet_model: Deprecated. Same argument as models. For backward compatibility only.
             **kwargs: See MODNetModel
         """
         self.__modnet_version__ = __version__
         self.bootstrap = bootstrap
+        self.random_state = random_state
         if modnet_models is not None and models is None:
             models = modnet_models
         if models is None:
@@ -74,6 +77,7 @@ def __init__(
         self.targets = self.models[0].targets
         self.weights = self.models[0].weights
         self.num_classes = self.models[0].num_classes
+        self.targets_groups = self.models[0].targets_groups
         self.out_act = self.models[0].out_act
 
     def fit(
@@ -89,18 +93,31 @@ def fit(
 
         if self.bootstrap:
             LOG.info("Generating bootstrap data...")
+            if self.random_state is None:
+                random_state = self.n_models * [None]
+            else:
+                random_state = np.arange(self.n_models) + self.random_state
+
+            # Loop over all targets and check if any involve classification, if so, stratify
+            stratify = None
+            for prop in self.targets_groups:
+                if self.num_classes[prop[0]] >= 2:  # Classification
+                    stratify = training_data.df_targets[prop[0]].array
+                    break
+            train_indices = [
+                resample(
+                    np.arange(len(training_data.df_targets)),
+                    replace=True,
+                    n_samples=len(training_data.df_targets),
+                    random_state=random_state[i],
+                    stratify=stratify,
+                )
+                for i in range(self.n_models)
+            ]
+
             train_datas = [
-                training_data.split(
-                    (
-                        resample(
-                            np.arange(len(training_data.df_targets)),
-                            replace=True,
-                            random_state=2943,
-                        ),
-                        [],
-                    )
-                )[0]
-                for _ in range(self.n_models)
+                training_data.split((train_indices[i], []))[0]
+                for i in range(self.n_models)
             ]
         else:
             train_datas = [training_data for _ in range(self.n_models)]