sinapsis-huggingface 0.2.8__tar.gz → 0.2.9__tar.gz

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sinapsis-huggingface
-Version: 0.2.8
+Version: 0.2.9
 Summary: Package for HuggingFace-based templates
 Author-email: SinapsisAI <dev@sinapsis.tech>
 Project-URL: Homepage, https://sinapsis.tech

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sinapsis-huggingface
-Version: 0.2.8
+Version: 0.2.9
 Summary: Package for HuggingFace-based templates
 Author-email: SinapsisAI <dev@sinapsis.tech>
 Project-URL: Homepage, https://sinapsis.tech

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_embeddings/src/sinapsis_huggingface_embeddings/templates/speaker_embedding_from_audio.py

@@ -3,6 +3,7 @@
 import io
 from typing import Literal
 
+import numpy as np
 import soundfile as sf
 import torch
 import torch.nn.functional as F
@@ -89,7 +90,7 @@ class SpeakerEmbeddingFromAudio(Template):
         return EncoderClassifier.from_hparams(source=self.attributes.model_path, savedir=self.attributes.data_cache_dir)
 
     @staticmethod
-    def _postprocess_speaker_embedding(speaker_embedding: torch.Tensor) -> list[float]:
+    def _postprocess_speaker_embedding(speaker_embedding: torch.Tensor) -> np.ndarray:
         """Normalize and convert the speaker embedding tensor into a list of floats.
 
         Args:
@@ -99,7 +100,7 @@ class SpeakerEmbeddingFromAudio(Template):
             list[float]: A normalized and flattened embedding as a list of floats.
         """
         speaker_embedding = F.normalize(speaker_embedding, dim=2)
-        speaker_embedding_list: list[float] = speaker_embedding.squeeze().tolist()
+        speaker_embedding_list: np.ndarray = speaker_embedding.detach().numpy().squeeze()
         return speaker_embedding_list
 
     @staticmethod
@@ -158,14 +159,13 @@ class SpeakerEmbeddingFromAudio(Template):
 
         if len(container.audios) == 1:
             for packet in packets:
-                packet.embedding = embeddings[0]
+                packet.embedding = [embeddings[0]]
         elif len(container.audios) == len(packets):
             for packet, embedding in zip(packets, embeddings):
-                packet.embedding = embedding
+                packet.embedding = [embedding]
         else:
             raise ValueError(
                 "Mismatch between the number of audio packets and target packets. "
                 "Ensure either a single audio or matching numbers of audios and target packets."
             )
-
         return container

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/base_transformers.py

@@ -1,5 +1,6 @@
 # -*- coding: utf-8 -*-
 
+import gc
 import random
 from abc import abstractmethod
 from typing import Any, Literal
@@ -64,6 +65,14 @@ class TransformersBase(Template):
         super().__init__(attributes)
         self._TORCH_DTYPE = {"float16": torch.float16, "float32": torch.float32}
         self.task: str | None = None
+        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._set_seed()
 
     def setup_pipeline(self) -> None:
@@ -147,7 +156,25 @@ class TransformersBase(Template):
         return transformed_data_container
 
     def reset_state(self, template_name: str | None = None) -> None:
-        if self.attributes.device == "cuda":
+        """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:
+            if self.pipeline.model is not None:
+                self.pipeline.model.to("cpu")
+            del self.pipeline
+
+        if hasattr(self, "processor"):
+            del self.processor
+
+        gc.collect()
+
+        if torch.cuda.is_available():
             torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
-        super().reset_state(template_name)
+
+        self.initialize()
+        self.logger.info(f"Reset template instance `{self.instance_name}`")

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/image_to_text_transformers.py

@@ -3,7 +3,7 @@
 import numpy as np
 from PIL import Image
 from sinapsis_core.data_containers.data_packet import DataContainer, TextPacket
-from sinapsis_core.template_base.base_models import OutputTypes, TemplateAttributeType
+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
@@ -14,8 +14,7 @@ ImageToTextTransformersUIProperties.tags.extend([Tags.IMAGE, Tags.TEXT, Tags.IMA
 
 
 class ImageToTextTransformers(TransformersBase):
-    """
-    ImageToTextTransformers template to generate text from an image.
+    """ImageToTextTransformers template to generate text from an image.
 
     This template uses a Hugging Face Transformers pipeline to generate textual descriptions
     from input images.
@@ -41,8 +40,13 @@ class ImageToTextTransformers(TransformersBase):
     GENERATED_TEXT_KEY = "generated_text"
     UIProperties = ImageToTextTransformersUIProperties
 
-    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.
+        """
+        super().initialize()
         self.task = "image-to-text"
         self.setup_pipeline()
 

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/pali_gemma/pali_gemma_base.py

@@ -54,6 +54,14 @@ class PaliGemmaBase(Template):
 
     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.model = self._setup_model()
         self.processor = self._setup_processor()
 
@@ -105,16 +113,24 @@ class PaliGemmaBase(Template):
         """
 
     def reset_state(self, template_name: str | None = None) -> None:
-        with torch.no_grad():
-            for child in self.model.children():
-                child.cpu()
-                del child
-                gc.collect()
+        """Releases the model 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"):
+            self.model.to("cpu")
             del self.model
 
+        if hasattr(self, "processor"):
             del self.processor
 
+        gc.collect()
+
+        if torch.cuda.is_available():
             torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
 
-        super().reset_state(template_name)
+        self.initialize()
+        self.logger.info(f"Reset template instance `{self.instance_name}`")

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/pali_gemma/pali_gemma_detection.py

@@ -2,7 +2,6 @@
 from dataclasses import dataclass
 
 from sinapsis_core.data_containers.annotations import ImageAnnotations
-from sinapsis_core.template_base.base_models import TemplateAttributeType
 from sinapsis_huggingface_transformers.templates.pali_gemma.pali_gemma_inference import (
     PaliGemmaInference,
     PaliGemmaInferenceAttributes,
@@ -66,9 +65,13 @@ class PaliGemmaDetection(PaliGemmaInference):
     AttributesBaseModel = PaliGemmaDetectionAttributes
     KEYS = PaliGemmaDetectionKeys
 
-    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.
+        """
+        super().initialize()
         objects_str = self.initialize_objects_str()
         self.prompt = self.KEYS.detection_prompt.format(objects_str)
 

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/pali_gemma/pali_gemma_inference.py

@@ -3,7 +3,6 @@ import numpy as np
 import torch
 from sinapsis_core.data_containers.annotations import BoundingBox, ImageAnnotations
 from sinapsis_core.data_containers.data_packet import DataContainer, ImagePacket
-from sinapsis_core.template_base.base_models import TemplateAttributeType
 from sinapsis_data_visualization.helpers.detection_utils import bbox_xyxy_to_xywh
 from sinapsis_huggingface_transformers.helpers.tags import Tags
 from sinapsis_huggingface_transformers.templates.pali_gemma.pali_gemma_base import (
@@ -67,8 +66,13 @@ class PaliGemmaInference(PaliGemmaBase):
     INPUT_IDS = "input_ids"
     UIProperties = PaliGemmaInferenceUIProperties
 
-    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.
+        """
+        super().initialize()
         self.model = self.model.eval()
         self.prompt = self.attributes.prompt
 

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/speech_to_text_transformers.py

@@ -2,7 +2,7 @@
 
 import numpy as np
 from sinapsis_core.data_containers.data_packet import DataContainer, TextPacket
-from sinapsis_core.template_base.base_models import OutputTypes, TemplateAttributeType
+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
@@ -15,8 +15,7 @@ SpeechToTextTransformersUIProperties.tags.extend(
 
 
 class SpeechToTextTransformers(TransformersBase):
-    """
-    Template to perform speech-to-text actions
+    """Template to perform speech-to-text actions
     using the HuggingFace module through the 'transformers' architecture.
 
     The template takes an Audio from the DataContainer and uses a speech-recognition
@@ -44,8 +43,13 @@ class SpeechToTextTransformers(TransformersBase):
     TEXT_KEY = "text"
     UIProperties = SpeechToTextTransformersUIProperties
 
-    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.
+        """
+        super().initialize()
         self.task = "automatic-speech-recognition"
         self.setup_pipeline()
 

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/summarization_transformers.py

@@ -1,7 +1,7 @@
 # -*- coding: utf-8 -*-
 
 from sinapsis_core.data_containers.data_packet import DataContainer
-from sinapsis_core.template_base.base_models import OutputTypes, TemplateAttributeType
+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
@@ -12,12 +12,10 @@ SummarizationTransformersUIProperties.tags.extend([Tags.SUMMARIZATION, Tags.TEXT
 
 
 class SummarizationTransformers(TransformersBase):
-    """
-    Template for text summarization using a Hugging Face Transformers pipeline.
+    """Template for text summarization using a Hugging Face Transformers pipeline.
 
     This class provides a reusable framework for summarizing text using a pre-trained
-    Hugging Face model. The `max_length` and `min_length` attributes control the length
-    of the generated summaries.
+    Hugging Face model.
 
     Usage example:
 
@@ -35,14 +33,22 @@ class SummarizationTransformers(TransformersBase):
         model_cache_dir: /path/to/cache/dir
         device: 'cuda'
         torch_dtype: float16
+        inference_kwargs:
+            min_length: 5
+            max_length: 20
 
     """
 
     SUMMARY_TEXT_KEY = "summary_text"
     UIProperties = SummarizationTransformersUIProperties
 
-    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.
+        """
+        super().initialize()
         self.task = "summarization"
         self.setup_pipeline()
 
@@ -57,12 +63,9 @@ class SummarizationTransformers(TransformersBase):
             DataContainer: DataContainer including the summarized text.
         """
         for text_packet in container.texts:
-            original_text = text_packet.content
-            summarized_text = self.pipeline(
-                original_text,
-                max_length=self.attributes.max_length,
-                min_length=self.attributes.min_length,
-            )[0][self.SUMMARY_TEXT_KEY]
+            summarized_text = self.pipeline(text_packet.content, **self.attributes.inference_kwargs)[0][
+                self.SUMMARY_TEXT_KEY
+            ]
 
             text_packet.content = summarized_text
         return container

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/text_to_speech_transformers.py

@@ -3,7 +3,7 @@
 import numpy as np
 import torch
 from sinapsis_core.data_containers.data_packet import AudioPacket, DataContainer, TextPacket
-from sinapsis_core.template_base.base_models import OutputTypes, TemplateAttributeType
+from sinapsis_core.template_base.base_models import OutputTypes
 
 from sinapsis_huggingface_transformers.helpers import sentences_to_n_words, split_text_into_sentences
 from sinapsis_huggingface_transformers.helpers.tags import Tags
@@ -37,8 +37,7 @@ class TextToSpeechAttributes(TransformersBaseAttributes):
 
 
 class TextToSpeechTransformers(TransformersBase):
-    """
-    The template generates an audio from a prompt that is passed
+    """The template generates an audio from a prompt that is passed
     through the text packet in the DataContainer.
     It uses the transformers architecture and a HuggingFace model to
     produce the audio. Finally, it sends the audio through the DataContainer
@@ -68,8 +67,13 @@ class TextToSpeechTransformers(TransformersBase):
     AttributesBaseModel = TextToSpeechAttributes
     UIProperties = TextToSpeechTransformersUIProperties
 
-    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.
+        """
+        super().initialize()
         self.task = "text-to-speech"
         self.setup_pipeline()
         self.sample_rate = self._get_sample_rate()

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/packages/sinapsis_huggingface_transformers/src/sinapsis_huggingface_transformers/templates/translation_transformers.py

@@ -2,7 +2,7 @@
 
 
 from sinapsis_core.data_containers.data_packet import DataContainer
-from sinapsis_core.template_base.base_models import OutputTypes, TemplateAttributeType
+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 (
@@ -28,8 +28,7 @@ class TranslationTransformersAttributes(TransformersBaseAttributes):
 
 
 class TranslationTransformers(TransformersBase):
-    """
-    Template for text translation using a Hugging Face Transformers pipeline.
+    """Template for text translation using a Hugging Face Transformers pipeline.
 
     This class provides a reusable framework for translating text from one language
     to another using a pre-trained Hugging Face model. The source and target languages
@@ -59,8 +58,13 @@ class TranslationTransformers(TransformersBase):
     TRANSLATION_TEXT_KEY = "translation_text"
     UIProperties = TranslationTransformersUIProperties
 
-    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.
+        """
+        super().initialize()
         self.task = f"translation_{self.attributes.source_language}_to_{self.attributes.target_language}"
         self.setup_pipeline()
 

{sinapsis_huggingface-0.2.8 → sinapsis_huggingface-0.2.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sinapsis-huggingface"
-version = "0.2.8"
+version = "0.2.9"
 description = "Package for HuggingFace-based templates"
 authors = [
     {name = "SinapsisAI", email = "dev@sinapsis.tech"},