natural-pdf 0.1.8__py3-none-any.whl → 0.1.9__py3-none-any.whl

natural_pdf/classification/manager.py

@@ -1,28 +1,35 @@
 import logging
 import time
 from datetime import datetime
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union, Tuple
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
+
+from PIL import Image
 
 # Use try-except for robustness if dependencies are missing
 try:
     import torch
-    from PIL import Image
-    from transformers import pipeline, AutoTokenizer, AutoModelForZeroShotImageClassification, AutoModelForSequenceClassification
+    from transformers import (
+        AutoModelForSequenceClassification,
+        AutoModelForZeroShotImageClassification,
+        AutoTokenizer,
+        pipeline,
+    )
+
     _CLASSIFICATION_AVAILABLE = True
 except ImportError:
     _CLASSIFICATION_AVAILABLE = False
     # Define dummy types for type hinting if imports fail
-    Image = type("Image", (), {})
     pipeline = object
     AutoTokenizer = object
     AutoModelForZeroShotImageClassification = object
     AutoModelForSequenceClassification = object
     torch = None
 
-# Import result classes
-from .results import ClassificationResult, CategoryScore
 from natural_pdf.utils.tqdm_utils import get_tqdm
 
+# Import result classes
+from .results import CategoryScore, ClassificationResult
+
 if TYPE_CHECKING:
     from transformers import Pipeline
 
@@ -34,8 +41,10 @@ _PIPELINE_CACHE: Dict[str, "Pipeline"] = {}
 _TOKENIZER_CACHE: Dict[str, Any] = {}
 _MODEL_CACHE: Dict[str, Any] = {}
 
+
 class ClassificationError(Exception):
     """Custom exception for classification errors."""
+
     pass
 
 
@@ -60,10 +69,12 @@ class ClassificationManager:
         if not _CLASSIFICATION_AVAILABLE:
             raise ImportError(
                 "Classification dependencies missing. "
-                "Install with: pip install \"natural-pdf[classification]\""
+                'Install with: pip install "natural-pdf[classification]"'
             )
 
-        self.pipelines: Dict[Tuple[str, str], "Pipeline"] = {} # Cache: (model_id, device) -> pipeline
+        self.pipelines: Dict[Tuple[str, str], "Pipeline"] = (
+            {}
+        )  # Cache: (model_id, device) -> pipeline
 
         self.device = default_device
         logger.info(f"ClassificationManager initialized on device: {self.device}")
@@ -76,7 +87,9 @@ class ClassificationManager:
         """Get or create a classification pipeline."""
         cache_key = f"{model_id}_{using}_{self.device}"
         if cache_key not in _PIPELINE_CACHE:
-            logger.info(f"Loading {using} classification pipeline for model '{model_id}' on device '{self.device}'...")
+            logger.info(
+                f"Loading {using} classification pipeline for model '{model_id}' on device '{self.device}'..."
+            )
             start_time = time.time()
             try:
                 task = (
@@ -84,16 +97,19 @@ class ClassificationManager:
                     if using == "text"
                     else "zero-shot-image-classification"
                 )
-                _PIPELINE_CACHE[cache_key] = pipeline(
-                    task,
-                    model=model_id,
-                    device=self.device
-                )
+                _PIPELINE_CACHE[cache_key] = pipeline(task, model=model_id, device=self.device)
                 end_time = time.time()
-                logger.info(f"Pipeline for '{model_id}' loaded in {end_time - start_time:.2f} seconds.")
+                logger.info(
+                    f"Pipeline for '{model_id}' loaded in {end_time - start_time:.2f} seconds."
+                )
             except Exception as e:
-                logger.error(f"Failed to load pipeline for model '{model_id}' (using: {using}): {e}", exc_info=True)
-                raise ClassificationError(f"Failed to load pipeline for model '{model_id}'. Ensure the model ID is correct and supports the {task} task.") from e
+                logger.error(
+                    f"Failed to load pipeline for model '{model_id}' (using: {using}): {e}",
+                    exc_info=True,
+                )
+                raise ClassificationError(
+                    f"Failed to load pipeline for model '{model_id}'. Ensure the model ID is correct and supports the {task} task."
+                ) from e
         return _PIPELINE_CACHE[cache_key]
 
     def infer_using(self, model_id: str, using: Optional[str] = None) -> str:
@@ -103,28 +119,44 @@ class ClassificationManager:
 
         # Simple inference based on common model names
         normalized_model_id = model_id.lower()
-        if "clip" in normalized_model_id or "vit" in normalized_model_id or "siglip" in normalized_model_id:
-             logger.debug(f"Inferred using='vision' for model '{model_id}'")
-             return "vision"
-        if "bart" in normalized_model_id or "bert" in normalized_model_id or "mnli" in normalized_model_id or "xnli" in normalized_model_id or "deberta" in normalized_model_id:
-             logger.debug(f"Inferred using='text' for model '{model_id}'")
-             return "text"
+        if (
+            "clip" in normalized_model_id
+            or "vit" in normalized_model_id
+            or "siglip" in normalized_model_id
+        ):
+            logger.debug(f"Inferred using='vision' for model '{model_id}'")
+            return "vision"
+        if (
+            "bart" in normalized_model_id
+            or "bert" in normalized_model_id
+            or "mnli" in normalized_model_id
+            or "xnli" in normalized_model_id
+            or "deberta" in normalized_model_id
+        ):
+            logger.debug(f"Inferred using='text' for model '{model_id}'")
+            return "text"
 
         # Fallback or raise error? Let's try loading text first, then vision.
-        logger.warning(f"Could not reliably infer mode for '{model_id}'. Trying text, then vision pipeline loading.")
+        logger.warning(
+            f"Could not reliably infer mode for '{model_id}'. Trying text, then vision pipeline loading."
+        )
         try:
             self._get_pipeline(model_id, "text")
             logger.info(f"Successfully loaded '{model_id}' as a text model.")
             return "text"
         except Exception:
-             logger.warning(f"Failed to load '{model_id}' as text model. Trying vision.")
-             try:
-                 self._get_pipeline(model_id, "vision")
-                 logger.info(f"Successfully loaded '{model_id}' as a vision model.")
-                 return "vision"
-             except Exception as e_vision:
-                 logger.error(f"Failed to load '{model_id}' as either text or vision model.", exc_info=True)
-                 raise ClassificationError(f"Cannot determine mode for model '{model_id}'. Please specify `using='text'` or `using='vision'`. Error: {e_vision}")
+            logger.warning(f"Failed to load '{model_id}' as text model. Trying vision.")
+            try:
+                self._get_pipeline(model_id, "vision")
+                logger.info(f"Successfully loaded '{model_id}' as a vision model.")
+                return "vision"
+            except Exception as e_vision:
+                logger.error(
+                    f"Failed to load '{model_id}' as either text or vision model.", exc_info=True
+                )
+                raise ClassificationError(
+                    f"Cannot determine mode for model '{model_id}'. Please specify `using='text'` or `using='vision'`. Error: {e_vision}"
+                )
 
     def classify_item(
         self,
@@ -134,90 +166,109 @@ class ClassificationManager:
         using: Optional[str] = None,
         min_confidence: float = 0.0,
         multi_label: bool = False,
-        **kwargs
-    ) -> ClassificationResult: # Return ClassificationResult
+        **kwargs,
+    ) -> ClassificationResult:  # Return ClassificationResult
         """Classifies a single item (text or image)."""
 
         # Determine model and engine type
         effective_using = using
         if model_id is None:
-             # Try inferring based on content type
-             if isinstance(item_content, str):
-                 effective_using = "text"
-                 model_id = self.DEFAULT_TEXT_MODEL
-             elif isinstance(item_content, Image.Image):
-                 effective_using = "vision"
-                 model_id = self.DEFAULT_VISION_MODEL
-             else:
-                 raise TypeError(f"Unsupported item_content type: {type(item_content)}")
+            # Try inferring based on content type
+            if isinstance(item_content, str):
+                effective_using = "text"
+                model_id = self.DEFAULT_TEXT_MODEL
+            elif isinstance(item_content, Image.Image):
+                effective_using = "vision"
+                model_id = self.DEFAULT_VISION_MODEL
+            else:
+                raise TypeError(f"Unsupported item_content type: {type(item_content)}")
         else:
-             # Infer engine type if not given
-             effective_using = self.infer_using(model_id, using)
-             # Set default model if needed (though should usually be provided if engine known)
-             if model_id is None:
-                  model_id = self.DEFAULT_TEXT_MODEL if effective_using == "text" else self.DEFAULT_VISION_MODEL
+            # Infer engine type if not given
+            effective_using = self.infer_using(model_id, using)
+            # Set default model if needed (though should usually be provided if engine known)
+            if model_id is None:
+                model_id = (
+                    self.DEFAULT_TEXT_MODEL
+                    if effective_using == "text"
+                    else self.DEFAULT_VISION_MODEL
+                )
 
         if not categories:
-             raise ValueError("Categories list cannot be empty.")
+            raise ValueError("Categories list cannot be empty.")
 
         pipeline_instance = self._get_pipeline(model_id, effective_using)
         timestamp = datetime.now()
-        parameters = { # Store parameters used for this run
-            'categories': categories,
-            'model_id': model_id,
-            'using': effective_using,
-            'min_confidence': min_confidence,
-            'multi_label': multi_label,
-            **kwargs
+        parameters = {  # Store parameters used for this run
+            "categories": categories,
+            "model_id": model_id,
+            "using": effective_using,
+            "min_confidence": min_confidence,
+            "multi_label": multi_label,
+            **kwargs,
         }
 
-        logger.debug(f"Classifying content (type: {type(item_content).__name__}) with model '{model_id}'")
+        logger.debug(
+            f"Classifying content (type: {type(item_content).__name__}) with model '{model_id}'"
+        )
         try:
             # Handle potential kwargs for specific pipelines if needed
             # The zero-shot pipelines expect `candidate_labels`
-            result_raw = pipeline_instance(item_content, candidate_labels=categories, multi_label=multi_label, **kwargs)
+            result_raw = pipeline_instance(
+                item_content, candidate_labels=categories, multi_label=multi_label, **kwargs
+            )
             logger.debug(f"Raw pipeline result: {result_raw}")
 
-            # --- Process raw result into ClassificationResult --- # 
+            # --- Process raw result into ClassificationResult --- #
             scores_list: List[CategoryScore] = []
-            
+
             # Handle text pipeline format (dict with 'labels' and 'scores')
-            if isinstance(result_raw, dict) and 'labels' in result_raw and 'scores' in result_raw:
-                for label, score_val in zip(result_raw['labels'], result_raw['scores']):
-                     if score_val >= min_confidence:
-                          try:
-                              scores_list.append(CategoryScore(label=label, confidence=score_val))
-                          except (ValueError, TypeError) as score_err:
-                               logger.warning(f"Skipping invalid score from text pipeline: label='{label}', score={score_val}. Error: {score_err}")
+            if isinstance(result_raw, dict) and "labels" in result_raw and "scores" in result_raw:
+                for label, score_val in zip(result_raw["labels"], result_raw["scores"]):
+                    if score_val >= min_confidence:
+                        try:
+                            scores_list.append(CategoryScore(label=label, confidence=score_val))
+                        except (ValueError, TypeError) as score_err:
+                            logger.warning(
+                                f"Skipping invalid score from text pipeline: label='{label}', score={score_val}. Error: {score_err}"
+                            )
             # Handle vision pipeline format (list of dicts with 'label' and 'score')
-            elif isinstance(result_raw, list) and all(isinstance(item, dict) and 'label' in item and 'score' in item for item in result_raw):
-                 for item in result_raw:
-                      score_val = item['score']
-                      label = item['label']
-                      if score_val >= min_confidence:
-                           try:
-                               scores_list.append(CategoryScore(label=label, confidence=score_val))
-                           except (ValueError, TypeError) as score_err:
-                                logger.warning(f"Skipping invalid score from vision pipeline: label='{label}', score={score_val}. Error: {score_err}")
+            elif isinstance(result_raw, list) and all(
+                isinstance(item, dict) and "label" in item and "score" in item
+                for item in result_raw
+            ):
+                for item in result_raw:
+                    score_val = item["score"]
+                    label = item["label"]
+                    if score_val >= min_confidence:
+                        try:
+                            scores_list.append(CategoryScore(label=label, confidence=score_val))
+                        except (ValueError, TypeError) as score_err:
+                            logger.warning(
+                                f"Skipping invalid score from vision pipeline: label='{label}', score={score_val}. Error: {score_err}"
+                            )
             else:
-                 logger.warning(f"Unexpected raw result format from pipeline for model '{model_id}': {type(result_raw)}. Cannot extract scores.")
-                 # Return empty result?
-                 # scores_list = []
+                logger.warning(
+                    f"Unexpected raw result format from pipeline for model '{model_id}': {type(result_raw)}. Cannot extract scores."
+                )
+                # Return empty result?
+                # scores_list = []
 
             return ClassificationResult(
                 model_id=model_id,
                 using=effective_using,
                 timestamp=timestamp,
                 parameters=parameters,
-                scores=scores_list
+                scores=scores_list,
             )
             # --- End Processing --- #
 
         except Exception as e:
-             logger.error(f"Classification failed for model '{model_id}': {e}", exc_info=True)
-             # Return an empty result object on failure?
-             # return ClassificationResult(model_id=model_id, engine_type=engine_type, timestamp=timestamp, parameters=parameters, scores=[])
-             raise ClassificationError(f"Classification failed using model '{model_id}'. Error: {e}") from e
+            logger.error(f"Classification failed for model '{model_id}': {e}", exc_info=True)
+            # Return an empty result object on failure?
+            # return ClassificationResult(model_id=model_id, engine_type=engine_type, timestamp=timestamp, parameters=parameters, scores=[])
+            raise ClassificationError(
+                f"Classification failed using model '{model_id}'. Error: {e}"
+            ) from e
 
     def classify_batch(
         self,
@@ -229,45 +280,51 @@ class ClassificationManager:
         multi_label: bool = False,
         batch_size: int = 8,
         progress_bar: bool = True,
-        **kwargs
-    ) -> List[ClassificationResult]: # Return list of ClassificationResult
+        **kwargs,
+    ) -> List[ClassificationResult]:  # Return list of ClassificationResult
         """Classifies a batch of items (text or image) using the pipeline's batching."""
         if not item_contents:
-             return []
+            return []
 
         # Determine model and engine type (assuming uniform type in batch)
         first_item = item_contents[0]
         effective_using = using
         if model_id is None:
-             if isinstance(first_item, str):
-                 effective_using = "text"
-                 model_id = self.DEFAULT_TEXT_MODEL
-             elif isinstance(first_item, Image.Image):
-                 effective_using = "vision"
-                 model_id = self.DEFAULT_VISION_MODEL
-             else:
-                 raise TypeError(f"Unsupported item_content type in batch: {type(first_item)}")
+            if isinstance(first_item, str):
+                effective_using = "text"
+                model_id = self.DEFAULT_TEXT_MODEL
+            elif isinstance(first_item, Image.Image):
+                effective_using = "vision"
+                model_id = self.DEFAULT_VISION_MODEL
+            else:
+                raise TypeError(f"Unsupported item_content type in batch: {type(first_item)}")
         else:
-             effective_using = self.infer_using(model_id, using)
-             if model_id is None:
-                  model_id = self.DEFAULT_TEXT_MODEL if effective_using == "text" else self.DEFAULT_VISION_MODEL
+            effective_using = self.infer_using(model_id, using)
+            if model_id is None:
+                model_id = (
+                    self.DEFAULT_TEXT_MODEL
+                    if effective_using == "text"
+                    else self.DEFAULT_VISION_MODEL
+                )
 
         if not categories:
-             raise ValueError("Categories list cannot be empty.")
+            raise ValueError("Categories list cannot be empty.")
 
         pipeline_instance = self._get_pipeline(model_id, effective_using)
-        timestamp = datetime.now() # Single timestamp for the batch run
-        parameters = { # Parameters for the whole batch
-            'categories': categories,
-            'model_id': model_id,
-            'using': effective_using,
-            'min_confidence': min_confidence,
-            'multi_label': multi_label,
-            'batch_size': batch_size,
-            **kwargs
+        timestamp = datetime.now()  # Single timestamp for the batch run
+        parameters = {  # Parameters for the whole batch
+            "categories": categories,
+            "model_id": model_id,
+            "using": effective_using,
+            "min_confidence": min_confidence,
+            "multi_label": multi_label,
+            "batch_size": batch_size,
+            **kwargs,
         }
 
-        logger.info(f"Classifying batch of {len(item_contents)} items with model '{model_id}' (batch size: {batch_size})")
+        logger.info(
+            f"Classifying batch of {len(item_contents)} items with model '{model_id}' (batch size: {batch_size})"
+        )
         batch_results_list: List[ClassificationResult] = []
 
         try:
@@ -277,67 +334,89 @@ class ClassificationManager:
                 candidate_labels=categories,
                 multi_label=multi_label,
                 batch_size=batch_size,
-                **kwargs
+                **kwargs,
             )
-            
+
             # Wrap with tqdm for progress if requested
             total_items = len(item_contents)
             if progress_bar:
-                 # Get the appropriate tqdm class
-                 tqdm_class = get_tqdm()
-                 results_iterator = tqdm_class(
-                      results_iterator, 
-                      total=total_items, 
-                      desc=f"Classifying batch ({model_id})",
-                      leave=False # Don't leave progress bar hanging
-                 )
+                # Get the appropriate tqdm class
+                tqdm_class = get_tqdm()
+                results_iterator = tqdm_class(
+                    results_iterator,
+                    total=total_items,
+                    desc=f"Classifying batch ({model_id})",
+                    leave=False,  # Don't leave progress bar hanging
+                )
 
             for raw_result in results_iterator:
-                # --- Process each raw result (which corresponds to ONE input item) --- # 
+                # --- Process each raw result (which corresponds to ONE input item) --- #
                 scores_list: List[CategoryScore] = []
                 try:
                     # Check for text format (dict with 'labels' and 'scores')
-                    if isinstance(raw_result, dict) and 'labels' in raw_result and 'scores' in raw_result:
-                        for label, score_val in zip(raw_result['labels'], raw_result['scores']):
-                             if score_val >= min_confidence:
-                                 try:
-                                     scores_list.append(CategoryScore(label=label, confidence=score_val))
-                                 except (ValueError, TypeError) as score_err:
-                                      logger.warning(f"Skipping invalid score from text pipeline batch: label='{label}', score={score_val}. Error: {score_err}")
+                    if (
+                        isinstance(raw_result, dict)
+                        and "labels" in raw_result
+                        and "scores" in raw_result
+                    ):
+                        for label, score_val in zip(raw_result["labels"], raw_result["scores"]):
+                            if score_val >= min_confidence:
+                                try:
+                                    scores_list.append(
+                                        CategoryScore(label=label, confidence=score_val)
+                                    )
+                                except (ValueError, TypeError) as score_err:
+                                    logger.warning(
+                                        f"Skipping invalid score from text pipeline batch: label='{label}', score={score_val}. Error: {score_err}"
+                                    )
                     # Check for vision format (list of dicts with 'label' and 'score')
                     elif isinstance(raw_result, list):
-                         for item in raw_result:
-                              try:
-                                  score_val = item['score']
-                                  label = item['label']
-                                  if score_val >= min_confidence:
-                                      scores_list.append(CategoryScore(label=label, confidence=score_val))
-                              except (KeyError, ValueError, TypeError) as item_err:
-                                   logger.warning(f"Skipping invalid item in vision result list from batch: {item}. Error: {item_err}")
+                        for item in raw_result:
+                            try:
+                                score_val = item["score"]
+                                label = item["label"]
+                                if score_val >= min_confidence:
+                                    scores_list.append(
+                                        CategoryScore(label=label, confidence=score_val)
+                                    )
+                            except (KeyError, ValueError, TypeError) as item_err:
+                                logger.warning(
+                                    f"Skipping invalid item in vision result list from batch: {item}. Error: {item_err}"
+                                )
                     else:
-                         logger.warning(f"Unexpected raw result format in batch item from model '{model_id}': {type(raw_result)}. Cannot extract scores.")
-                
+                        logger.warning(
+                            f"Unexpected raw result format in batch item from model '{model_id}': {type(raw_result)}. Cannot extract scores."
+                        )
+
                 except Exception as proc_err:
-                     logger.error(f"Error processing result item in batch: {proc_err}", exc_info=True)
-                     # scores_list remains empty for this item
+                    logger.error(
+                        f"Error processing result item in batch: {proc_err}", exc_info=True
+                    )
+                    # scores_list remains empty for this item
 
                 # Append result object for this item
-                batch_results_list.append(ClassificationResult(
-                     model_id=model_id,
-                     using=effective_using,
-                     timestamp=timestamp, # Use same timestamp for batch
-                     parameters=parameters, # Use same params for batch
-                     scores=scores_list
-                ))
+                batch_results_list.append(
+                    ClassificationResult(
+                        model_id=model_id,
+                        using=effective_using,
+                        timestamp=timestamp,  # Use same timestamp for batch
+                        parameters=parameters,  # Use same params for batch
+                        scores=scores_list,
+                    )
+                )
                 # --- End Processing --- #
 
             if len(batch_results_list) != total_items:
-                 logger.warning(f"Batch classification returned {len(batch_results_list)} results, but expected {total_items}. Results might be incomplete or misaligned.")
+                logger.warning(
+                    f"Batch classification returned {len(batch_results_list)} results, but expected {total_items}. Results might be incomplete or misaligned."
+                )
 
             return batch_results_list
 
         except Exception as e:
-             logger.error(f"Batch classification failed for model '{model_id}': {e}", exc_info=True)
-             # Return list of empty results?
-             # return [ClassificationResult(model_id=model_id, s=engine_type, timestamp=timestamp, parameters=parameters, scores=[]) for _ in item_contents]
-             raise ClassificationError(f"Batch classification failed using model '{model_id}'. Error: {e}") from e
+            logger.error(f"Batch classification failed for model '{model_id}': {e}", exc_info=True)
+            # Return list of empty results?
+            # return [ClassificationResult(model_id=model_id, s=engine_type, timestamp=timestamp, parameters=parameters, scores=[]) for _ in item_contents]
+            raise ClassificationError(
+                f"Batch classification failed using model '{model_id}'. Error: {e}"
+            ) from e