sinapsis-huggingface 0.2.10__tar.gz → 0.2.12__tar.gz

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sinapsis-huggingface
-Version: 0.2.10
+Version: 0.2.12
 Summary: Package for HuggingFace-based templates
 Author-email: SinapsisAI <dev@sinapsis.tech>
 Project-URL: Homepage, https://sinapsis.tech

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sinapsis-huggingface
-Version: 0.2.10
+Version: 0.2.12
 Summary: Package for HuggingFace-based templates
 Author-email: SinapsisAI <dev@sinapsis.tech>
 Project-URL: Homepage, https://sinapsis.tech

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_diffusers/src/sinapsis_huggingface_diffusers/templates/base_diffusers.py

@@ -1,12 +1,13 @@
 # -*- coding: utf-8 -*-
 
+import gc
 from abc import ABC, abstractmethod
-from typing import Any, Literal
+from typing import Literal
 
 import numpy as np
 import torch
 from diffusers import DiffusionPipeline
-from pydantic import BaseModel, ConfigDict
+from pydantic import BaseModel, ConfigDict, Field
 from sinapsis_core.data_containers.data_packet import DataContainer, ImagePacket
 from sinapsis_core.template_base import Template
 from sinapsis_core.template_base.base_models import (
@@ -20,6 +21,37 @@ from sinapsis_core.utils.env_var_keys import SINAPSIS_CACHE_DIR
 from sinapsis_huggingface_diffusers.helpers.tags import Tags
 
 
+class BaseGenerationParams(BaseModel):
+    """Defines the core parameters for a diffusers generation pipeline.
+
+    This model is flexible and allows any other parameters (e.g., `strength`)
+    to be passed, which will be forwarded to the underlying pipeline.
+
+    Attributes:
+        prompt (str | list[str] | None): The text prompt(s) to guide image generation.
+        height (int | None): The height in pixels of the generated image.
+        width (int | None): The width in pixels of the generated image.
+        negative_prompt (str | list[str] | None): Prompt(s) to guide the model away
+            from generating certain things.
+        num_inference_steps (int | None): The number of denoising steps. More steps
+            typically result in higher quality but are slower. Defaults to 50.
+        guidance_scale (float | None): Controls how much the prompt influences the
+            output. Higher values mean stronger adherence. Defaults to 7.5.
+        num_images_per_prompt (int | None): The number of images to generate per
+            prompt. Defaults to 1.
+    """
+
+    model_config = ConfigDict(extra="allow")
+
+    prompt: str | list[str] | None = None
+    height: int | None = None
+    width: int | None = None
+    negative_prompt: str | list[str] | None = None
+    num_inference_steps: int | None = 50
+    guidance_scale: float | None = 7.5
+    num_images_per_prompt: int | None = 1
+
+
 class BaseDiffusersAttributes(TemplateAttributes):
     """Configuration attributes for setting up a diffusion pipeline and generating images.
 
@@ -33,7 +65,7 @@ class BaseDiffusersAttributes(TemplateAttributes):
         device (Literal["cuda", "cpu"]): Device for computations, either "cpu" or "cuda".
         torch_dtype (Literal["float16", "float32"]): Data type for PyTorch tensors.
         enable_model_cpu_offload (bool): If True, enables CPU offloading to reduce GPU memory usage.
-        generation_params (dict): Parameters for image generation (e.g., prompt, guidance_scale).
+        generation_params (BaseGenerationParams): Parameters for image generation (e.g., prompt, guidance_scale).
         seed (int | list[int] | None): Random seed(s) for reproducibility.
         overwrite_images (bool): Whether to overwrite the existing images in the container.
             Defaults to False.
@@ -44,7 +76,7 @@ class BaseDiffusersAttributes(TemplateAttributes):
     device: Literal["cuda", "cpu"]
     torch_dtype: Literal["float16", "float32"] = "float16"
     enable_model_cpu_offload: bool = False
-    generation_params: dict[str, Any]
+    generation_params: BaseGenerationParams = Field(default_factory=BaseGenerationParams)
     seed: int | list[int] | None = None
     overwrite_images: bool = False
 
@@ -76,10 +108,17 @@ class BaseDiffusers(Template, ABC):
 
     def __init__(self, attributes: TemplateAttributeType) -> None:
         super().__init__(attributes)
+        self.initialize()
 
+    def initialize(self) -> None:
+        """Initializes the template's common state for creation or reset.
+
+        This method is called by both `__init__` and `reset_state` to ensure
+        a consistent state. Can be overriden by subclasses for specific behaviour.
+        """
         self.pipeline = self._make_pipeline()
         self.pipeline.set_progress_bar_config(disable=True)
-        self.num_images_per_prompt = self.attributes.generation_params.get("num_images_per_prompt", 1)
+        self.num_images_per_prompt = self.attributes.generation_params.num_images_per_prompt
         self.generator = self._make_generator()
 
         if self.attributes.enable_model_cpu_offload:
@@ -124,7 +163,7 @@ class BaseDiffusers(Template, ABC):
         pipeline_class = self._pipeline_class()
         return pipeline_class.from_pretrained(
             self.attributes.model_path,
-            torch_dtype=self.TORCH_DTYPE.get(self.attributes.torch_dtype),
+            dtype=self.TORCH_DTYPE.get(self.attributes.torch_dtype),
             cache_dir=self.attributes.model_cache_dir,
         ).to(self.attributes.device)
 
@@ -163,13 +202,15 @@ class BaseDiffusers(Template, ABC):
             inputs = {}
         output = self.pipeline(
             **inputs,
-            **self.attributes.generation_params,
+            **self.attributes.generation_params.model_dump(exclude_none=True),
             generator=self.generator,
             output_type="np",
         )
         generated_images = output.images if output_attribute == "images" else output.frames[0]
+        images_as_uint8 = [(image * 255).clip(0, 255).astype(np.uint8) for image in generated_images]
+        del output
 
-        return [(image * 255).clip(0, 255).astype(np.uint8) for image in generated_images]
+        return images_as_uint8
 
     def _update_images_in_container(self, container: DataContainer, new_packets: list[ImagePacket]) -> None:
         """Updates the container with new image packets based on the `overwrite_images` attribute.
@@ -187,17 +228,39 @@ class BaseDiffusers(Template, ABC):
         else:
             container.images.extend(new_packets)
 
-    def _clear_memory(self) -> None:
+    @staticmethod
+    def clear_memory() -> None:
         """Clears memory to free up resources.
 
         This method performs garbage collection and clears GPU memory (if applicable) to prevent memory leaks
         and ensure efficient resource usage.
         """
-        del self.generator
-        if self.attributes.device == "cuda":
+        gc.collect()
+        if torch.cuda.is_available():
             torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
 
     def reset_state(self, template_name: str | None = None) -> None:
-        self._clear_memory()
-        super().reset_state(template_name)
+        """Releases the pipeline and processor from memory and re-instantiates the template.
+
+        Args:
+            template_name (str | None, optional): The name of the template instance being reset. Defaults to None.
+        """
+        _ = template_name
+
+        if hasattr(self, "pipeline") and self.pipeline is not None:
+            components = self.pipeline.components
+            for component in components.values():
+                if isinstance(component, torch.nn.Module):
+                    component.to("cpu")
+
+            del self.pipeline
+
+        if hasattr(self, "processor"):
+            del self.processor
+
+        if hasattr(self, "generator"):
+            del self.generator
+
+        self.clear_memory()
+        self.initialize()
+        self.logger.info(f"Reset template instance `{self.instance_name}`")

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_diffusers/src/sinapsis_huggingface_diffusers/templates/image_to_image_diffusers.py

@@ -8,7 +8,6 @@ import numpy as np
 from diffusers import AutoPipelineForImage2Image
 from PIL import Image
 from sinapsis_core.data_containers.data_packet import DataContainer, ImagePacket
-from sinapsis_core.template_base.base_models import TemplateAttributeType
 
 from sinapsis_huggingface_diffusers.helpers.tags import Tags
 from sinapsis_huggingface_diffusers.templates.base_diffusers import BaseDiffusers
@@ -52,9 +51,13 @@ class ImageToImageDiffusers(BaseDiffusers):
 
     UIProperties = ImageToImageDiffusersUIProperties
 
-    def __init__(self, attributes: TemplateAttributeType) -> None:
-        super().__init__(attributes)
-        self.requires_pil = False
+    def initialize(self) -> None:
+        """Initializes the template's common state for creation or reset.
+
+        This method is called by both `__init__` and `reset_state` to ensure
+        a consistent state. Can be overriden by subclasses for specific behaviour.
+        """
+        super().initialize()
         self.output_attribute: Literal["images", "frames"] = "images"
         self.num_duplicates = self.num_images_per_prompt
 
@@ -67,7 +70,8 @@ class ImageToImageDiffusers(BaseDiffusers):
         """
         return AutoPipelineForImage2Image
 
-    def _convert_image_format(self, image_packet: ImagePacket) -> np.ndarray | Image.Image:
+    @staticmethod
+    def _convert_image_format(image_packet: ImagePacket) -> Image.Image:
         """Converts the input image into the appropriate format for the pipeline.
 
         The format depends on the `requires_pil` attribute:
@@ -78,12 +82,9 @@ class ImageToImageDiffusers(BaseDiffusers):
             image_packet (ImagePacket): The input image packet.
 
         Returns:
-            np.ndarray | Image.Image: The converted image, either as a normalized NumPy array or a
-                PIL Image.
+            Image.Image: The converted image as a PIL Image.
         """
-        if self.requires_pil:
-            return Image.fromarray(image_packet.content)
-        return image_packet.content / 255.0
+        return Image.fromarray(image_packet.content)
 
     def preprocess_inputs(self, image_packet: ImagePacket) -> dict[str, np.ndarray | list[np.ndarray]]:
         """Prepares the input image for the image-to-image pipeline.
@@ -157,6 +158,6 @@ class ImageToImageDiffusers(BaseDiffusers):
         new_packets = [ImagePacket(content=image) for image in all_generated_images]
         processed_packets, _ = self.post_processing_packets(new_packets, old_packets)
         self._update_images_in_container(container, processed_packets)
-        self._clear_memory()
+        self.clear_memory()
 
         return container

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_diffusers/src/sinapsis_huggingface_diffusers/templates/image_to_video_gen_xl_diffusers.py

@@ -1,9 +1,11 @@
 # -*- coding: utf-8 -*-
+from typing import Literal
 
 from diffusers import I2VGenXLPipeline
-from sinapsis_core.template_base.base_models import TemplateAttributeType
+from pydantic import Field
 
 from sinapsis_huggingface_diffusers.helpers.tags import Tags
+from sinapsis_huggingface_diffusers.templates.base_diffusers import BaseDiffusersAttributes, BaseGenerationParams
 from sinapsis_huggingface_diffusers.templates.image_to_image_diffusers import (
     ImageToImageDiffusers,
 )
@@ -12,6 +14,32 @@ ImageToVideoGenXLDiffusersUIProperties = ImageToImageDiffusers.UIProperties
 ImageToVideoGenXLDiffusersUIProperties.tags.extend([Tags.VIDEO, Tags.IMAGE_TO_VIDEO])
 
 
+class ImageToVideoGenerationParams(BaseGenerationParams):
+    """Defines the specific parameters for image-to-video generation pipelines.
+
+    Attributes:
+        target_fps (int | None): The target frames per second for the generated video.
+        num_frames (int | None): The total number of frames to generate in the video. Defaults to 16.
+        num_videos_per_prompt (int | None): The number of different videos to generate
+            from the same input image and prompt.
+    """
+
+    target_fps: int | None = None
+    num_frames: int | None = 16
+    num_videos_per_prompt: int | None = None
+
+
+class ImageToVideoGenXLDiffusersAttributes(BaseDiffusersAttributes):
+    """Defines the complete set of attributes for the ImageToVideoGenXLDiffusers template.
+
+    Attributes:
+        generation_params (ImageToVideoGenerationParams): Task-specific parameters for
+            video generation, such as `num_frames` and `target_fps`.
+    """
+
+    generation_params: ImageToVideoGenerationParams = Field(default_factory=ImageToVideoGenerationParams)
+
+
 class ImageToVideoGenXLDiffusers(ImageToImageDiffusers):
     """This class implements a specific template for image-to-video generation using Hugging Face's
     diffusers. The `ImageToVideoGenXLDiffusers` class inherits from the `ImageToImageDiffusers` template
@@ -42,14 +70,18 @@ class ImageToVideoGenXLDiffusers(ImageToImageDiffusers):
 
     """
 
+    AttributesBaseModel = ImageToVideoGenXLDiffusersAttributes
     UIProperties = ImageToVideoGenXLDiffusersUIProperties
-    DEFAULT_NUM_FRAMES = 16
 
-    def __init__(self, attributes: TemplateAttributeType) -> None:
-        super().__init__(attributes)
-        self.num_duplicates = self.attributes.generation_params.get("num_frames", self.DEFAULT_NUM_FRAMES)
-        self.requires_pil = True
-        self.output_attribute = "frames"
+    def initialize(self) -> None:
+        """Initializes the template's common state for creation or reset.
+
+        This method is called by both `__init__` and `reset_state` to ensure
+        a consistent state. Can be overriden by subclasses for specific behaviour.
+        """
+        super().initialize()
+        self.output_attribute: Literal["images", "frames"] = "frames"
+        self.num_duplicates = self.attributes.generation_params.num_frames
 
     @staticmethod
     def _pipeline_class() -> I2VGenXLPipeline:

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_diffusers/src/sinapsis_huggingface_diffusers/templates/inpainting_diffusers.py

@@ -7,7 +7,6 @@ import cv2
 import numpy as np
 from diffusers import AutoPipelineForInpainting
 from sinapsis_core.data_containers.data_packet import ImageAnnotations, ImagePacket
-from sinapsis_core.template_base.base_models import TemplateAttributeType
 
 from sinapsis_huggingface_diffusers.helpers.tags import Tags
 from sinapsis_huggingface_diffusers.templates.base_diffusers import (
@@ -75,8 +74,13 @@ class InpaintingDiffusers(ImageToImageDiffusers):
     UIProperties = InpaintingDiffusersUIProperties
     AttributesBaseModel = InpaintingDiffusersAttributes
 
-    def __init__(self, attributes: TemplateAttributeType) -> None:
-        super().__init__(attributes)
+    def initialize(self) -> None:
+        """Initializes the template's common state for creation or reset.
+
+        This method is called by both `__init__` and `reset_state` to ensure
+        a consistent state. Can be overriden by subclasses for specific behaviour.
+        """
+        super().initialize()
         if self.attributes.preserve_outside_content and self.attributes.dilation_radius is None:
             raise ValueError("Need to specify a dilation_radius if preserve_outside_content=True")
 
@@ -327,6 +331,6 @@ class InpaintingDiffusers(ImageToImageDiffusers):
             if old_packet.annotations:
                 new_packet.annotations = old_packet.annotations
                 for ann in new_packet.annotations:
-                    ann.label_str = str(self.attributes.generation_params.get("prompt"))
+                    ann.label_str = str(self.attributes.generation_params.prompt)
 
         return new_packets, old_packets

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_diffusers/src/sinapsis_huggingface_diffusers/templates/text_to_image_diffusers.py

@@ -80,6 +80,6 @@ class TextToImageDiffusers(BaseDiffusers):
         image_packets = [ImagePacket(content=image) for image in generated_images]
         self._set_packet_sources(image_packets)
         self._update_images_in_container(container, image_packets)
-        self._clear_memory()
+        self.clear_memory()
 
         return container

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_grounding_dino/src/sinapsis_huggingface_grounding_dino/templates/grounding_dino.py

@@ -111,11 +111,18 @@ class GroundingDINO(Template):
             attributes (dict[str, Any]): Dictionary containing configuration parameters.
         """
         super().__init__(attributes)
-        self.device = self.attributes.device
+        self.initialize()
+
+    def initialize(self) -> None:
+        """Initializes the template's common state for creation or reset.
+
+        This method is called by both `__init__` and `reset_state` to ensure
+        a consistent state. Can be overriden by subclasses for specific behaviour.
+        """
         self.processor = AutoProcessor.from_pretrained(
             self.attributes.model_path, cache_dir=self.attributes.model_cache_dir
         )
-        self.model = self._set_model().to(self.device)
+        self.model = self._set_model().to(self.attributes.device)
         self.max_tokens = self.processor.tokenizer.model_max_length
         self.text_input = self.validate_and_format_text_input(self.attributes.text_input)
 
@@ -171,7 +178,7 @@ class GroundingDINO(Template):
             images=image_packet.content,
             text=self.text_input,
             return_tensors="pt",
-        ).to(self.device)
+        ).to(self.attributes.device)
 
         with torch.no_grad():
             outputs = self.model(**inputs)
@@ -350,19 +357,32 @@ class GroundingDINO(Template):
 
         return container
 
-    def _clear_memory(self) -> None:
+    @staticmethod
+    def clear_memory() -> None:
         """Clears memory to free up resources.
 
         This method performs garbage collection and clears GPU memory (if applicable) to prevent memory leaks
         and ensure efficient resource usage.
         """
-        for child in self.model.children():
-            child.cpu()
-            gc.collect()
-        if self.attributes.device == "cuda":
+        gc.collect()
+        if torch.cuda.is_available():
             torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
 
     def reset_state(self, template_name: str | None = None) -> None:
-        self._clear_memory()
-        super().reset_state(template_name)
+        """Releases the pipeline and processor from memory and re-instantiates the template.
+
+        Args:
+            template_name (str | None, optional): The name of the template instance being reset. Defaults to None.
+        """
+        _ = template_name
+
+        if hasattr(self, "model") and self.model is not None:
+            self.model.to("cpu")
+            del self.model
+
+        if hasattr(self, "processor"):
+            del self.processor
+
+        self.clear_memory()
+        self.initialize()
+        self.logger.info(f"Reset template instance `{self.instance_name}`")

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/base_transformers.py

@@ -6,7 +6,7 @@ from abc import abstractmethod
 from typing import Any, Literal
 
 import torch
-from pydantic import Field
+from pydantic import BaseModel, ConfigDict, Field
 from sinapsis_core.data_containers.data_packet import DataContainer
 from sinapsis_core.template_base import Template
 from sinapsis_core.template_base.base_models import TemplateAttributes, TemplateAttributeType, UIPropertiesMetadata
@@ -17,6 +17,29 @@ from transformers.pipelines import Pipeline
 from sinapsis_huggingface_transformers.helpers.tags import Tags
 
 
+class BaseInferenceKwargs(BaseModel):
+    """A flexible container for keyword arguments passed during inference.
+
+    Attributes:
+        generate_kwargs (dict[str, Any] | None): A dictionary of advanced parameters passed directly to the
+            model's `generate` method for fine-tuning the pipeline generation.
+    """
+
+    generate_kwargs: dict[str, Any] | None = None
+    model_config = ConfigDict(extra="allow")
+
+
+class PipelineKwargs(BaseModel):
+    """A flexible container for keyword arguments used to create the pipeline.
+
+    This model allows any extra parameters to be passed during pipeline instantiation.
+    """
+
+    device: Literal["cuda", "cpu"]
+    torch_dtype: Literal["float16", "float32", "auto"] = "float16"
+    model_config = ConfigDict(extra="allow")
+
+
 class TransformersBaseAttributes(TemplateAttributes):
     """Attributes for configuring the TransformersPipelineTemplate.
 
@@ -31,20 +54,18 @@ class TransformersBaseAttributes(TemplateAttributes):
         seed (int | None): Random seed for reproducibility. If provided, this seed will ensure
             consistent results for pipelines that involve randomness. If not provided, a random seed
             will be generated internally.
-        pipeline_kwargs (dict[str, Any]): Keyword arguments passed during the instantiation of the
+        pipeline_kwargs (PipelineKwargs): Keyword arguments passed during the instantiation of the
             Hugging Face pipeline.
-        inference_kwargs (dict[str, Any]): Keyword arguments passed during the task execution or
+        inference_kwargs (BaseInferenceKwargs): Keyword arguments passed during the task execution or
             inference phase. These allow dynamic customization of the task, such as `max_length`
             and `min_length` for summarization, or `max_new_tokens` for image-to-text.
     """
 
     model_path: str
     model_cache_dir: str = str(SINAPSIS_CACHE_DIR)
-    device: Literal["cuda", "cpu"]
-    torch_dtype: Literal["float16", "float32"] = "float16"
     seed: int | None = None
-    pipeline_kwargs: dict[str, Any] = Field(default_factory=dict)
-    inference_kwargs: dict[str, Any] = Field(default_factory=dict)
+    pipeline_kwargs: PipelineKwargs = Field(default_factory=PipelineKwargs)
+    inference_kwargs: BaseInferenceKwargs = Field(default_factory=BaseInferenceKwargs)
 
 
 class TransformersBase(Template):
@@ -123,9 +144,7 @@ class TransformersBase(Template):
         return pipeline(
             task=self.task,
             model=self.attributes.model_path,
-            device=self.attributes.device,
-            torch_dtype=self._TORCH_DTYPE.get(self.attributes.torch_dtype),
-            **self.attributes.pipeline_kwargs,
+            **self.attributes.pipeline_kwargs.model_dump(),
             **kwargs,
         )
 

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/image_to_text_transformers.py

@@ -2,17 +2,47 @@
 
 import numpy as np
 from PIL import Image
+from pydantic import Field
 from sinapsis_core.data_containers.data_packet import DataContainer, TextPacket
 from sinapsis_core.template_base.base_models import OutputTypes
 
 from sinapsis_huggingface_transformers.helpers.tags import Tags
-from sinapsis_huggingface_transformers.templates.base_transformers import TransformersBase
+from sinapsis_huggingface_transformers.templates.base_transformers import (
+    BaseInferenceKwargs,
+    TransformersBase,
+    TransformersBaseAttributes,
+)
 
 ImageToTextTransformersUIProperties = TransformersBase.UIProperties
 ImageToTextTransformersUIProperties.output_type = OutputTypes.TEXT
 ImageToTextTransformersUIProperties.tags.extend([Tags.IMAGE, Tags.TEXT, Tags.IMAGE_TO_TEXT])
 
 
+class ImageToTextInferenceKwargs(BaseInferenceKwargs):
+    """Specific keyword arguments for the image-to-text pipeline.
+
+    Attributes:
+        max_new_tokens (int | None): The maximum number of tokens to generate in the description.
+        timeout (float | None): The maximum time in seconds to wait for fetching images from the web.
+    """
+
+    max_new_tokens: int | None = None
+    timeout: float | None = None
+
+
+class ImageToTextTransformersAttributes(TransformersBaseAttributes):
+    """Defines the complete set of attributes for the ImageToTextTransformers template.
+
+    Inherits general transformer settings from TransformersBaseAttributes.
+
+    Attributes:
+        inference_kwargs (ImageToTextInferenceKwargs): Task-specific parameters for the image-to-text pipeline,
+            such as `max_new_tokens`.
+    """
+
+    inference_kwargs: ImageToTextInferenceKwargs = Field(default_factory=ImageToTextInferenceKwargs)
+
+
 class ImageToTextTransformers(TransformersBase):
     """ImageToTextTransformers template to generate text from an image.
 
@@ -37,6 +67,7 @@ class ImageToTextTransformers(TransformersBase):
 
     """
 
+    AttributesBaseModel = ImageToTextTransformersAttributes
     GENERATED_TEXT_KEY = "generated_text"
     UIProperties = ImageToTextTransformersUIProperties
 
@@ -78,7 +109,9 @@ class ImageToTextTransformers(TransformersBase):
         """
         for image_packet in container.images:
             image = self._convert_to_pil(image_packet.content)
-            text_description = self.pipeline(image, **self.attributes.inference_kwargs)[0][self.GENERATED_TEXT_KEY]
-            text_packet = TextPacket(content=text_description)
-            container.texts.append(text_packet)
+            results = self.pipeline(image, **self.attributes.inference_kwargs.model_dump(exclude_none=True))
+            if results:
+                text_description = results[0].get(self.GENERATED_TEXT_KEY)
+                if text_description:
+                    container.texts.append(TextPacket(content=text_description))
         return container

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/speech_to_text_transformers.py

@@ -1,11 +1,17 @@
 # -*- coding: utf-8 -*-
+from typing import Literal
 
 import numpy as np
+from pydantic import Field
 from sinapsis_core.data_containers.data_packet import DataContainer, TextPacket
 from sinapsis_core.template_base.base_models import OutputTypes
 
 from sinapsis_huggingface_transformers.helpers.tags import Tags
-from sinapsis_huggingface_transformers.templates.base_transformers import TransformersBase
+from sinapsis_huggingface_transformers.templates.base_transformers import (
+    BaseInferenceKwargs,
+    TransformersBase,
+    TransformersBaseAttributes,
+)
 
 SpeechToTextTransformersUIProperties = TransformersBase.UIProperties
 SpeechToTextTransformersUIProperties.output_type = OutputTypes.TEXT
@@ -14,6 +20,30 @@ SpeechToTextTransformersUIProperties.tags.extend(
 )
 
 
+class SpeechToTextInferenceKwargs(BaseInferenceKwargs):
+    """Specific keyword arguments for the automatic-speech-recognition pipeline.
+
+    Attributes:
+        return_timestamps (Literal["char", "word"] | bool | None ): If set, controls the granularity of
+            timestamps returned with the transcribed text. Can be "char", "word", or True for segments.
+    """
+
+    return_timestamps: Literal["char", "word"] | bool | None = None
+
+
+class SpeechToTextTransformersAttributes(TransformersBaseAttributes):
+    """Defines the set of attributes for the SpeechToTextTransformers template.
+
+    Inherits general transformer settings from TransformersBaseAttributes.
+
+    Attributes:
+        inference_kwargs (SpeechToTextInferenceKwargs): Task-specific parameters for the speech-to-text pipeline,
+            such as `return_timestamps`.
+    """
+
+    inference_kwargs: SpeechToTextInferenceKwargs = Field(default_factory=SpeechToTextInferenceKwargs)
+
+
 class SpeechToTextTransformers(TransformersBase):
     """Template to perform speech-to-text actions
     using the HuggingFace module through the 'transformers' architecture.
@@ -40,6 +70,7 @@ class SpeechToTextTransformers(TransformersBase):
 
     """
 
+    AttributesBaseModel = SpeechToTextTransformersAttributes
     TEXT_KEY = "text"
     UIProperties = SpeechToTextTransformersUIProperties
 
@@ -65,11 +96,15 @@ class SpeechToTextTransformers(TransformersBase):
         for audio_packet in container.audios:
             audio = audio_packet.content
             audio = audio.astype(np.float32)
-            transcribed_text = self.pipeline(audio, **self.attributes.inference_kwargs)[self.TEXT_KEY]
-            transcribed_text_textpacket = TextPacket(
-                content=transcribed_text,
-                source=audio_packet.source,
-            )
-            self.logger.info(f"Speech-to-text transcription: {transcribed_text}")
-            container.texts.append(transcribed_text_textpacket)
+            results = self.pipeline(audio, **self.attributes.inference_kwargs.model_dump(exclude_none=True))
+            if results:
+                transcribed_text = results.get(self.TEXT_KEY)
+                if transcribed_text:
+                    self.logger.info(f"Speech-to-text transcription: {transcribed_text}")
+                    container.texts.append(
+                        TextPacket(
+                            content=transcribed_text,
+                            source=audio_packet.source,
+                        )
+                    )
         return container

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/summarization_transformers.py

@@ -1,16 +1,48 @@
 # -*- coding: utf-8 -*-
 
+from pydantic import Field
 from sinapsis_core.data_containers.data_packet import DataContainer
 from sinapsis_core.template_base.base_models import OutputTypes
 
 from sinapsis_huggingface_transformers.helpers.tags import Tags
-from sinapsis_huggingface_transformers.templates.base_transformers import TransformersBase
+from sinapsis_huggingface_transformers.templates.base_transformers import (
+    BaseInferenceKwargs,
+    TransformersBase,
+    TransformersBaseAttributes,
+)
 
 SummarizationTransformersUIProperties = TransformersBase.UIProperties
 SummarizationTransformersUIProperties.output_type = OutputTypes.TEXT
 SummarizationTransformersUIProperties.tags.extend([Tags.SUMMARIZATION, Tags.TEXT])
 
 
+class SummarizationInferenceKwargs(BaseInferenceKwargs):
+    """Specific keyword arguments for the summarization pipeline.
+
+    Attributes:
+        return_text (bool | None): Whether or not to include the decoded texts in the outputs.
+        return_tensors (bool | None): Whether or not to include the tensors of predictions.
+        clean_up_tokenization_spaces (bool | None): Whether or not to clean up the potential extra spaces.
+    """
+
+    return_text: bool | None = True
+    return_tensors: bool | None = False
+    clean_up_tokenization_spaces: bool | None = False
+
+
+class SummarizationTransformersAttributes(TransformersBaseAttributes):
+    """Defines the complete set of attributes for the SummarizationTransformers template.
+
+    Inherits general transformer settings from TransformersBaseAttributes.
+
+    Attributes:
+        inference_kwargs: Task-specific parameters for the summarization pipeline,
+            such as `clean_up_tokenization_spaces`.
+    """
+
+    inference_kwargs: SummarizationInferenceKwargs = Field(default_factory=SummarizationInferenceKwargs)
+
+
 class SummarizationTransformers(TransformersBase):
     """Template for text summarization using a Hugging Face Transformers pipeline.
 
@@ -39,6 +71,7 @@ class SummarizationTransformers(TransformersBase):
 
     """
 
+    AttributesBaseModel = SummarizationTransformersAttributes
     SUMMARY_TEXT_KEY = "summary_text"
     UIProperties = SummarizationTransformersUIProperties
 
@@ -63,9 +96,11 @@ class SummarizationTransformers(TransformersBase):
             DataContainer: DataContainer including the summarized text.
         """
         for text_packet in container.texts:
-            summarized_text = self.pipeline(text_packet.content, **self.attributes.inference_kwargs)[0][
-                self.SUMMARY_TEXT_KEY
-            ]
-
-            text_packet.content = summarized_text
+            results = self.pipeline(
+                text_packet.content, **self.attributes.inference_kwargs.model_dump(exclude_none=True)
+            )
+            if results:
+                summarized_text = results[0].get(self.SUMMARY_TEXT_KEY)
+                if summarized_text:
+                    text_packet.content = summarized_text
         return container

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/text_to_speech_transformers.py

@@ -1,5 +1,4 @@
 # -*- coding: utf-8 -*-
-
 import numpy as np
 import torch
 from sinapsis_core.data_containers.data_packet import AudioPacket, DataContainer, TextPacket
@@ -65,6 +64,7 @@ class TextToSpeechTransformers(TransformersBase):
     """
 
     AttributesBaseModel = TextToSpeechAttributes
+    SAMPLE_RATE_KEY = "sampling_rate"
     UIProperties = TextToSpeechTransformersUIProperties
 
     def initialize(self) -> None:
@@ -94,7 +94,7 @@ class TextToSpeechTransformers(TransformersBase):
             else {}
         )
         output = self.pipeline("Fetching sampling rate.", forward_params=forward_params)
-        sample_rate = output.get("sampling_rate", 16000)
+        sample_rate = output.get(self.SAMPLE_RATE_KEY, 16000)
 
         return sample_rate
 
@@ -132,7 +132,9 @@ class TextToSpeechTransformers(TransformersBase):
             else {}
         )
         for chunk in sentences:
-            output = self.pipeline(chunk, forward_params=forward_params, **self.attributes.inference_kwargs)
+            output = self.pipeline(
+                chunk, forward_params=forward_params, **self.attributes.inference_kwargs.model_dump(exclude_none=True)
+            )
             total_audio.append(output["audio"][0] if output["audio"].ndim == 2 else output["audio"])
         if total_audio:
             total_audio = np.concatenate(total_audio)

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/translation_transformers.py

@@ -1,11 +1,12 @@
 # -*- coding: utf-8 -*-
 
-
+from pydantic import Field
 from sinapsis_core.data_containers.data_packet import DataContainer
 from sinapsis_core.template_base.base_models import OutputTypes
 
 from sinapsis_huggingface_transformers.helpers.tags import Tags
 from sinapsis_huggingface_transformers.templates.base_transformers import (
+    BaseInferenceKwargs,
     TransformersBase,
     TransformersBaseAttributes,
 )
@@ -15,14 +16,31 @@ TranslationTransformersUIProperties.output_type = OutputTypes.TEXT
 TranslationTransformersUIProperties.tags.extend([Tags.LANGUAGE, Tags.TRANSLATION])
 
 
+class TranslationInferenceKwargs(BaseInferenceKwargs):
+    """Specific keyword arguments for the translation pipeline.
+
+    Attributes:
+        return_text (bool | None): Whether or not to include the decoded texts in the outputs.
+        return_tensors (bool | None): Whether or not to include the tensors of predictions.
+        clean_up_tokenization_spaces (bool | None): Whether or not to clean up the potential extra spaces.
+    """
+
+    return_text: bool | None = True
+    return_tensors: bool | None = False
+    clean_up_tokenization_spaces: bool | None = False
+
+
 class TranslationTransformersAttributes(TransformersBaseAttributes):
     """Attributes for the transformers pipeline translation task.
 
     Attributes:
+        inference_kwargs: Task-specific parameters for the transaltion pipeline,
+            such as `clean_up_tokenization_spaces`.
         source_language (str): The language code of the source language (e.g., "en" for English).
         target_language (str): The language code of the target language (e.g., "fr" for French).
     """
 
+    inference_kwargs: TranslationInferenceKwargs = Field(default_factory=TranslationInferenceKwargs)
     source_language: str
     target_language: str
 
@@ -79,8 +97,11 @@ class TranslationTransformers(TransformersBase):
             DataContainer: DataContainer including the translated text.
         """
         for text_packet in container.texts:
-            translated_text = self.pipeline(text_packet.content, **self.attributes.inference_kwargs)[0][
-                self.TRANSLATION_TEXT_KEY
-            ]
-            text_packet.content = translated_text
+            results = self.pipeline(
+                text_packet.content, **self.attributes.inference_kwargs.model_dump(exclude_none=True)
+            )
+            if results:
+                translated_text = results[0].get(self.TRANSLATION_TEXT_KEY)
+                if translated_text:
+                    text_packet.content = translated_text
         return container

{sinapsis_huggingface-0.2.10 → sinapsis_huggingface-0.2.12}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sinapsis-huggingface"
-version = "0.2.10"
+version = "0.2.12"
 description = "Package for HuggingFace-based templates"
 authors = [
     {name = "SinapsisAI", email = "dev@sinapsis.tech"},