[Config] Add optional arguments (#1395)

* Optional Components

* uP

* finish

* finish

* finish

* Apply suggestions from code review

Co-authored-by: Pedro Cuenca <[email protected]>

* up

* Update src/diffusers/pipeline_utils.py

* improve

Co-authored-by: Pedro Cuenca <[email protected]>

Author: patrickvonplaten

src/diffusers/pipeline_utils.py

@@ -129,10 +129,13 @@ class DiffusionPipeline(ConfigMixin):
 
     Class attributes:
 
-        - **config_name** ([`str`]) -- name of the config file that will store the class and module names of all
+        - **config_name** (`str`) -- name of the config file that will store the class and module names of all
           components of the diffusion pipeline.
+        - **_optional_components** (List[`str`]) -- list of all components that are optional so they don't have to be
+          passed for the pipeline to function (should be overridden by subclasses).
     """
     config_name = "model_index.json"
+    _optional_components = []
 
     def register_modules(self, **kwargs):
         # import it here to avoid circular import
@@ -184,12 +187,19 @@ def save_pretrained(self, save_directory: Union[str, os.PathLike]):
         model_index_dict.pop("_diffusers_version")
         model_index_dict.pop("_module", None)
 
+        expected_modules, optional_kwargs = self._get_signature_keys(self)
+
+        def is_saveable_module(name, value):
+            if name not in expected_modules:
+                return False
+            if name in self._optional_components and value[0] is None:
+                return False
+            return True
+
+        model_index_dict = {k: v for k, v in model_index_dict.items() if is_saveable_module(k, v)}
+
         for pipeline_component_name in model_index_dict.keys():
             sub_model = getattr(self, pipeline_component_name)
-            if sub_model is None:
-                # edge case for saving a pipeline with safety_checker=None
-                continue
-
             model_cls = sub_model.__class__
 
             save_method_name = None
@@ -523,26 +533,27 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
         # some modules can be passed directly to the init
         # in this case they are already instantiated in `kwargs`
         # extract them here
-        expected_modules = set(inspect.signature(pipeline_class.__init__).parameters.keys()) - set(["self"])
+        expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class)
         passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+        passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
 
         init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs)
 
+        # define init kwargs
+        init_kwargs = {k: init_dict.pop(k) for k in optional_kwargs if k in init_dict}
+        init_kwargs = {**init_kwargs, **passed_pipe_kwargs}
+
+        # remove `null` components
+        init_dict = {k: v for k, v in init_dict.items() if v[0] is not None}
+
         if len(unused_kwargs) > 0:
             logger.warning(f"Keyword arguments {unused_kwargs} not recognized.")
 
-        init_kwargs = {}
-
         # import it here to avoid circular import
         from diffusers import pipelines
 
         # 3. Load each module in the pipeline
         for name, (library_name, class_name) in init_dict.items():
-            if class_name is None:
-                # edge case for when the pipeline was saved with safety_checker=None
-                init_kwargs[name] = None
-                continue
-
             # 3.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
             if class_name.startswith("Flax"):
                 class_name = class_name[4:]
@@ -570,7 +581,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                             f"{passed_class_obj[name]} is of type: {type(passed_class_obj[name])}, but should be"
                             f" {expected_class_obj}"
                         )
-                elif passed_class_obj[name] is None:
+                elif passed_class_obj[name] is None and name not in pipeline_class._optional_components:
                     logger.warning(
                         f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note"
                         f" that this might lead to problems when using {pipeline_class} and is not recommended."
@@ -651,11 +662,13 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
 
         # 4. Potentially add passed objects if expected
         missing_modules = set(expected_modules) - set(init_kwargs.keys())
-        if len(missing_modules) > 0 and missing_modules <= set(passed_class_obj.keys()):
+        passed_modules = list(passed_class_obj.keys())
+        optional_modules = pipeline_class._optional_components
+        if len(missing_modules) > 0 and missing_modules <= set(passed_modules + optional_modules):
             for module in missing_modules:
-                init_kwargs[module] = passed_class_obj[module]
+                init_kwargs[module] = passed_class_obj.get(module, None)
         elif len(missing_modules) > 0:
-            passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys()))
+            passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - optional_kwargs
             raise ValueError(
                 f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
             )
@@ -664,6 +677,14 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
         model = pipeline_class(**init_kwargs)
         return model
 
+    @staticmethod
+    def _get_signature_keys(obj):
+        parameters = inspect.signature(obj.__init__).parameters
+        required_parameters = {k: v for k, v in parameters.items() if v.default is not True}
+        optional_parameters = set({k for k, v in parameters.items() if v.default is True})
+        expected_modules = set(required_parameters.keys()) - set(["self"])
+        return expected_modules, optional_parameters
+
     @property
     def components(self) -> Dict[str, Any]:
         r"""
@@ -688,8 +709,10 @@ def components(self) -> Dict[str, Any]:
         Returns:
             A dictionaly containing all the modules needed to initialize the pipeline.
         """
-        components = {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
-        expected_modules = set(inspect.signature(self.__init__).parameters.keys()) - set(["self"])
+        expected_modules, optional_parameters = self._get_signature_keys(self)
+        components = {
+            k: getattr(self, k) for k in self.config.keys() if not k.startswith("_") and k not in optional_parameters
+        }
 
         if set(components.keys()) != expected_modules:
             raise ValueError(

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -67,6 +67,7 @@ class AltDiffusionPipeline(DiffusionPipeline):
         feature_extractor ([`CLIPFeatureExtractor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
+    _optional_components = ["safety_checker", "feature_extractor"]
 
     def __init__(
         self,
@@ -84,6 +85,7 @@ def __init__(
         ],
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPFeatureExtractor,
+        requires_safety_checker: bool = True,
     ):
         super().__init__()
 
@@ -114,7 +116,7 @@ def __init__(
             new_config["clip_sample"] = False
             scheduler._internal_dict = FrozenDict(new_config)
 
-        if safety_checker is None:
+        if safety_checker is None and requires_safety_checker:
             logger.warning(
                 f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                 " that you abide to the conditions of the Alt Diffusion license and do not expose unfiltered"
@@ -124,6 +126,12 @@ def __init__(
                 " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
             )
 
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
         self.register_modules(
             vae=vae,
             text_encoder=text_encoder,
@@ -133,6 +141,7 @@ def __init__(
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     def enable_xformers_memory_efficient_attention(self):
         r"""

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -80,6 +80,7 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline):
         feature_extractor ([`CLIPFeatureExtractor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
+    _optional_components = ["safety_checker", "feature_extractor"]
 
     def __init__(
         self,
@@ -97,6 +98,7 @@ def __init__(
         ],
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPFeatureExtractor,
+        requires_safety_checker: bool = True,
     ):
         super().__init__()
 
@@ -127,7 +129,7 @@ def __init__(
             new_config["clip_sample"] = False
             scheduler._internal_dict = FrozenDict(new_config)
 
-        if safety_checker is None:
+        if safety_checker is None and requires_safety_checker:
             logger.warning(
                 f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                 " that you abide to the conditions of the Alt Diffusion license and do not expose unfiltered"
@@ -137,6 +139,12 @@ def __init__(
                 " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
             )
 
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
         self.register_modules(
             vae=vae,
             text_encoder=text_encoder,
@@ -146,6 +154,7 @@ def __init__(
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
         r"""

src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

@@ -132,6 +132,7 @@ class CycleDiffusionPipeline(DiffusionPipeline):
         feature_extractor ([`CLIPFeatureExtractor`]):
             Model that extracts features from generated images to be used as inputs for the `safety_checker`.
     """
+    _optional_components = ["safety_checker", "feature_extractor"]
 
     def __init__(
         self,
@@ -142,6 +143,7 @@ def __init__(
         scheduler: DDIMScheduler,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPFeatureExtractor,
+        requires_safety_checker: bool = True,
     ):
         super().__init__()
 
@@ -159,7 +161,7 @@ def __init__(
             new_config["steps_offset"] = 1
             scheduler._internal_dict = FrozenDict(new_config)
 
-        if safety_checker is None:
+        if safety_checker is None and requires_safety_checker:
             logger.warning(
                 f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                 " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
@@ -169,6 +171,12 @@ def __init__(
                 " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
             )
 
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
         self.register_modules(
             vae=vae,
             text_encoder=text_encoder,
@@ -178,6 +186,7 @@ def __init__(
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing
     def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py

@@ -51,6 +51,7 @@ def __init__(
         scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
         safety_checker: OnnxRuntimeModel,
         feature_extractor: CLIPFeatureExtractor,
+        requires_safety_checker: bool = True,
     ):
         super().__init__()
 
@@ -81,6 +82,22 @@ def __init__(
             new_config["clip_sample"] = False
             scheduler._internal_dict = FrozenDict(new_config)
 
+        if safety_checker is None and requires_safety_checker:
+            logger.warning(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
         self.register_modules(
             vae_encoder=vae_encoder,
             vae_decoder=vae_decoder,
@@ -91,6 +108,7 @@ def __init__(
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     def _encode_prompt(self, prompt, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
         r"""

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py

@@ -87,6 +87,7 @@ def __init__(
         scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
         safety_checker: OnnxRuntimeModel,
         feature_extractor: CLIPFeatureExtractor,
+        requires_safety_checker: bool = True,
     ):
         super().__init__()
 
@@ -117,7 +118,7 @@ def __init__(
             new_config["clip_sample"] = False
             scheduler._internal_dict = FrozenDict(new_config)
 
-        if safety_checker is None:
+        if safety_checker is None and requires_safety_checker:
             logger.warning(
                 f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                 " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
@@ -127,6 +128,12 @@ def __init__(
                 " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
             )
 
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
         self.register_modules(
             vae_encoder=vae_encoder,
             vae_decoder=vae_decoder,
@@ -137,6 +144,7 @@ def __init__(
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
     def _encode_prompt(self, prompt, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):

src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py

@@ -100,6 +100,7 @@ def __init__(
         scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
         safety_checker: OnnxRuntimeModel,
         feature_extractor: CLIPFeatureExtractor,
+        requires_safety_checker: bool = True,
     ):
         super().__init__()
         logger.info("`OnnxStableDiffusionInpaintPipeline` is experimental and will very likely change in the future.")
@@ -131,7 +132,7 @@ def __init__(
             new_config["clip_sample"] = False
             scheduler._internal_dict = FrozenDict(new_config)
 
-        if safety_checker is None:
+        if safety_checker is None and requires_safety_checker:
             logger.warning(
                 f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                 " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
@@ -141,6 +142,12 @@ def __init__(
                 " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
             )
 
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
         self.register_modules(
             vae_encoder=vae_encoder,
             vae_decoder=vae_decoder,
@@ -151,6 +158,7 @@ def __init__(
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
         )
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
     def _encode_prompt(self, prompt, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):