pass transforms around instead of making duplicates (facebookresearch…

…#416)

Summary:

Instead of creating duplicate transforms whenever we need one, we create a single transform from the config and initialize the wrapped model and wrapped generators with that one transform. This passes the same transform object around and allows the transformations to learn parameters and still be synced up across wrapped objects.

Differential Revision: D65155103

aepsych/strategy.py

@@ -410,9 +410,9 @@ def from_config(cls, config: Config, name: str) -> Strategy:
         stimuli_per_trial = config.getint(name, "stimuli_per_trial", fallback=1)
         outcome_types = config.getlist(name, "outcome_types", element_type=str)
 
-        generator = GeneratorWrapper.from_config(name, config)
+        generator = GeneratorWrapper.from_config(name, config, transforms)
 
-        model = ModelWrapper.from_config(name, config)
+        model = ModelWrapper.from_config(name, config, transforms)
 
         acqf_cls = config.getobj(name, "acqf", fallback=None)
         if acqf_cls is not None and hasattr(generator, "acqf"):

aepsych/transforms/parameters.py

@@ -162,12 +162,14 @@ def from_config(
         cls,
         name: str,
         config: Config,
+        transforms: Optional[ChainedInputTransform] = None,
     ):
         gen_cls = config.getobj(name, "generator", fallback=SobolGenerator)
-        transforms = ParameterTransforms.from_config(config)
+        if transforms is None:
+            transforms = ParameterTransforms.from_config(config)
 
         # We need transformed values from config but we don't want to edit config
-        transformed_config = transform_options(config)
+        transformed_config = transform_options(config, transforms)
 
         gen = gen_cls.from_config(transformed_config)
 
@@ -281,27 +283,32 @@ def from_config(
         cls,
         name: str,
         config: Config,
+        transforms: Optional[ChainedInputTransform] = None,
     ):
         # We don't always have models
         model_cls = config.getobj(name, "model", fallback=None)
         if model_cls is None:
             return None
 
-        transforms = ParameterTransforms.from_config(config)
+        if transforms is None:
+            transforms = ParameterTransforms.from_config(config)
 
         # Need transformed values
-        transformed_config = transform_options(config)
+        transformed_config = transform_options(config, transforms)
 
         model = model_cls.from_config(transformed_config)
 
         return cls(model, transforms)
 
 
-def transform_options(config: Config) -> Config:
+def transform_options(
+    config: Config, transforms: Optional[ChainedInputTransform] = None
+) -> Config:
     """
     Return a copy of the config with the options transformed. The config
     """
-    transforms = ParameterTransforms.from_config(config)
+    if transforms is None:
+        transforms = ParameterTransforms.from_config(config)
 
     configClone = deepcopy(config)
 

tests/test_transforms.py

@@ -100,6 +100,14 @@ def test_model_init_equivalent(self):
 
     def test_transforms_in_strategy(self):
         for _strat in self.strat.strat_list:
+            # Check if the same transform is passed around everywhere
+            self.assertTrue(id(_strat.transforms) == id(_strat.generator.transforms))
+            if _strat.model is not None:
+                self.assertTrue(
+                    id(_strat.generator.transforms) == id(_strat.model.transforms)
+                )
+
+            # Check all the transform bits are the same
             for strat_transform, gen_transform in zip(
                 _strat.transforms.items(), _strat.generator.transforms.items()
             ):