[DEVX-828] Added image summarization in multimodal pipeline

.gitignore

@@ -51,6 +51,9 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 
+testing/*
+!testing/test.ipynb
+
 # Translations
 *.mo
 *.pot

clarifai_datautils/multimodal/pipeline/loaders.py

@@ -27,6 +27,7 @@ def __getitem__(self, index: int):
     meta.pop('coordinates', None)
     meta.pop('detection_class_prob', None)
     image_data = meta.pop('image_base64', None)
+    id = meta.get('input_id', None)
     if image_data is not None:
       # Ensure image_data is already bytes before encoding
       image_data = base64.b64decode(image_data)
@@ -39,7 +40,7 @@ def __getitem__(self, index: int):
       meta['type'] = 'table'
 
     return MultiModalFeatures(
-        text=text, image_bytes=image_data, labels=[self.pipeline_name], metadata=meta)
+        text=text, image_bytes=image_data, labels=[self.pipeline_name], metadata=meta, id=id)
 
   def __len__(self):
     return len(self.elements)
@@ -61,10 +62,13 @@ def task(self):
     return DATASET_UPLOAD_TASKS.TEXT_CLASSIFICATION  #TODO: Better dataset name in SDK
 
   def __getitem__(self, index: int):
+    id = self.elements[index].to_dict().get('element_id', None)
+    id = id[:48] if id is not None else None
     return TextFeatures(
         text=self.elements[index].text,
         labels=self.pipeline_name,
-        metadata=self.elements[index].metadata.to_dict())
+        metadata=self.elements[index].metadata.to_dict(),
+        id=id)
 
   def __len__(self):
     return len(self.elements)

clarifai_datautils/multimodal/pipeline/summarizer.py

@@ -0,0 +1,96 @@
+import base64
+import os
+import random
+from typing import List
+
+try:
+  from unstructured.documents.elements import CompositeElement, ElementMetadata, Image
+except ImportError:
+  raise ImportError(
+      "Could not import unstructured package. "
+      "Please install it with `pip install 'unstructured[pdf] @ git+https://github.com/clarifai/unstructured.git@support_clarifai_model'`."
+  )
+
+from clarifai.client.input import Inputs
+from clarifai.client.model import Model
+
+from .basetransform import BaseTransform
+
+
+class ImageSummarizer(BaseTransform):
+  """ Summarizes image elements. """
+
+  def __init__(self, model_url="https://clarifai.com/qwen/qwen-VL/models/qwen-VL-Chat"):
+    """Initializes an LlamaIndexWrapper object.
+
+    Args:
+        model_url (str): Model URL to use for summarization.
+
+    """
+    self.model_url = model_url
+    self.model = Model(url=model_url, pat=os.environ.get("CLARIFAI_PAT"))
+    self.summary_prompt = """You are an assistant tasked with summarizing images for retrieval. \
+        These summaries will be embedded and used to retrieve the raw image. \
+        Give a concise summary of the image that is well optimized for retrieval."""
+
+  def __call__(self, elements: List) -> List:
+    """Applies the transformation.
+
+    Args:
+        elements (List[str]): List of text elements.
+
+    Returns:
+        List of transformed text elements.
+
+    """
+    img_elements = []
+    for _, element in enumerate(elements):
+      element.metadata.update(
+          ElementMetadata.from_dict({
+              'is_original': True,
+              'input_id': f'{random.randint(1000000, 99999999)}'
+          }))
+      if isinstance(element, Image):
+        img_elements.append(element)
+    # new_elements = Parallel(n_jobs=len(elements))(delayed(self._summarize_image)(element) for element in img_elements)
+    new_elements = self._summarize_image(elements)
+    elements.extend(new_elements)
+    return elements
+
+  def _summarize_image(self, image_elements: List[Image]) -> List[CompositeElement]:
+    """Summarizes an image element.
+
+    Args:
+        image_element (Image): Image element to summarize.
+
+    Returns:
+        Summarized image element.
+
+    """
+    img_inputs = []
+    for element in image_elements:
+      if not isinstance(element, Image):
+        continue
+      new_input_id = "summarize_" + element.metadata.input_id
+      input_proto = Inputs.get_multimodal_input(
+          input_id=new_input_id,
+          image_bytes=base64.b64decode(element.metadata.image_base64),
+          raw_text=self.summary_prompt)
+      img_inputs.append(input_proto)
+    resp = self.model.predict(img_inputs)
+
+    new_elements = []
+    for i, element in enumerate(resp.outputs):
+      summary = ""
+      if image_elements[i].text:
+        summary = image_elements[i].text
+      summary = summary + " \n " + element.data.text.raw
+      eid = image_elements[i].metadata.input_id
+      meta_dict = {'source_input_id': eid, 'is_original': False}
+      comp_element = CompositeElement(
+          text=summary,
+          metadata=ElementMetadata.from_dict(meta_dict),
+          element_id="summarized_" + eid)
+      new_elements.append(comp_element)
+
+    return new_elements

tests/pipelines/test_multimodal_pipelines.py

@@ -66,3 +66,32 @@ def test_pipeline_run_loader(self,):
     assert elements.__class__.__name__ == 'MultiModalLoader'
     assert len(elements) == 14
     assert elements.elements[0].metadata.to_dict()['filename'] == 'Multimodal_sample_file.pdf'
+
+  def test_pipeline_summarize(self,):
+    """Tests for pipeline run with summarizer"""
+    from clarifai_datautils.multimodal import Pipeline
+    from clarifai_datautils.multimodal.pipeline.cleaners import Clean_extra_whitespace
+    from clarifai_datautils.multimodal.pipeline.PDF import PDFPartitionMultimodal
+    from clarifai_datautils.multimodal.pipeline.summarizer import ImageSummarizer
+
+    pipeline = Pipeline(
+        name='pipeline-1',
+        transformations=[
+            PDFPartitionMultimodal(chunking_strategy="by_title", max_characters=1024),
+            Clean_extra_whitespace(),
+            ImageSummarizer()
+        ])
+    elements = pipeline.run(files=PDF_FILE_PATH, loader=False)
+    assert len(elements) == 15
+    assert isinstance(elements, list)
+    assert elements[0].metadata.to_dict()['filename'] == 'Multimodal_sample_file.pdf'
+    assert elements[0].metadata.to_dict()['page_number'] == 1
+    assert elements[0].metadata.to_dict()['email_address'] == ['[email protected]']
+    assert elements[6].__class__.__name__ == 'Table'
+    assert elements[-2].__class__.__name__ == 'Image'
+    assert elements[-2].metadata.is_original is True
+    assert elements[-2].metadata.input_id is not None
+    id = elements[-2].metadata.input_id
+    assert elements[-1].__class__.__name__ == 'CompositeElement'
+    assert elements[-1].metadata.is_original is False
+    assert elements[-1].metadata.source_input_id == 'summarized_' + id