Update utils/vulnerability_classifier.py

utils/vulnerability_classifier.py

@@ -1,152 +1,307 @@
-from typing import List, Tuple
+from haystack.nodes.base import BaseComponent
+from haystack.schema import Document
+from haystack.nodes import ImageToTextConverter, PDFToTextConverter
+from haystack.nodes import TextConverter, DocxToTextConverter, PreProcessor
+from pdf2image import convert_from_path
+from typing import Callable, Dict, List, Optional, Text, Tuple, Union
 from typing_extensions import Literal
-import logging
 import pandas as pd
-from pandas import DataFrame, Series
-from utils.config import getconfig
-from utils.preprocessing import processingpipeline
+import logging
+import re
+import string
+from haystack.pipelines import Pipeline
 import streamlit as st
-from transformers import pipeline
-from setfit import SetFitModel
-
-label_dict= {0: 'Agricultural communities',
- 1: 'Children',
- 2: 'Coastal communities',
- 3: 'Ethnic, racial or other minorities',
- 4: 'Fishery communities',
- 5: 'Informal sector workers',
- 6: 'Members of indigenous and local communities',
- 7: 'Migrants and displaced persons',
- 8: 'Older persons',
- 9: 'Other',
- 10: 'Persons living in poverty',
- 11: 'Persons with disabilities',
- 12: 'Persons with pre-existing health conditions',
- 13: 'Residents of drought-prone regions',
- 14: 'Rural populations',
- 15: 'Sexual minorities (LGBTQI+)',
- 16: 'Urban populations',
- 17: 'Women and other genders'}
-
-def get_vulnerability_labels(preds):
 
+@st.cache_data
+def useOCR(file_path: str)-> Text:
     """
-    Function that takes the numerical predictions as an input and returns a list of the labels.
+    Converts image pdfs into text, Using the Farm-haystack[OCR]
     
-    """
+    Params
+    ----------
+    file_path: file_path of uploade file, returned by add_upload function in 
+    uploadAndExample.py
     
-    # Get label names
-    preds_list = preds.tolist()
+    Returns the text file as string.
+    """
+    # we need pdf file to be first converted into image file
+    # this will create each page as image file
+    images = convert_from_path(pdf_path = file_path)
+    list_ = []
+    # save image file in cache and read them one by one to pass it to OCR
+    for i, pdf in enumerate(images):
+        # Save pages as images in the pdf
+        pdf.save(f'PDF\image_converted_{i+1}.png', 'PNG')
+        list_.append(f'PDF\image_converted_{i+1}.png')
     
-    # Get the name of the group where the prediction is equal to "1"
-    result = []
+    converter = ImageToTextConverter(remove_numeric_tables=True,
+                                      valid_languages=["eng"])
+    # placeholder to collect  the text from each page
+    placeholder = []
+    for file in list_:
+        document = converter.convert(
+                            file_path=file, meta=None,
+                            )[0]
+
+        text = document.content
+        placeholder.append(text)
+    # join the text from each page by page separator
+    text = '\x0c'.join(placeholder)
+    return text
+
+
+
+class FileConverter(BaseComponent):
+    """
+    Wrapper class to convert uploaded document into text by calling appropriate 
+    Converter class, will use internally haystack PDFToTextOCR in case of image 
+    pdf. Cannot use the FileClassifier from haystack as its doesnt has any 
+    label/output class for image.
+    1. https://haystack.deepset.ai/pipeline_nodes/custom-nodes
+    2. https://docs.haystack.deepset.ai/docs/file_converters
+    3. https://github.com/deepset-ai/haystack/tree/main/haystack/nodes/file_converter
+    4. https://docs.haystack.deepset.ai/reference/file-converters-api
+    """
+
+    outgoing_edges = 1
+
+    def run(self, file_name: str , file_path: str, encoding: Optional[str]=None,
+            id_hash_keys: Optional[List[str]] = None,
+            ) -> Tuple[dict,str]:
+        """ this is required method to invoke the component in 
+            the pipeline implementation. 
+            
+        Params
+        ----------
+        file_name: name of file
+        file_path: file_path of uploade file, returned by add_upload function in 
+                    uploadAndExample.py
+        
+        See the links provided in Class docstring/description to see other params
+        
+        Return
+        ---------
+        output: dictionary, with key as identifier and value could be anything 
+                we need to return. In this case its the List of Hasyatck Document
+        
+        output_1: As there is only one outgoing edge, we pass 'output_1' string
+        """
+        try:
+            if file_name.endswith('.pdf'):
+                converter = PDFToTextConverter(remove_numeric_tables=True)
+            if file_name.endswith('.txt'):
+                converter = TextConverter(remove_numeric_tables=True)
+            if file_name.endswith('.docx'):
+                converter = DocxToTextConverter() 
+        except Exception as e:
+            logging.error(e)
+            return 
+
+
+
+        documents = []
+
+        document = converter.convert(
+                      file_path=file_path, meta=None, 
+                      encoding=encoding, id_hash_keys=id_hash_keys
+                      )[0]
+
+        text = document.content
 
-    for sublist in preds_list:
-        names = [label_dict[key] for key, value in enumerate(sublist) if value == 1]
-        result.append(names)
+        # in case of scanned/images only PDF the content might contain only
+        # the page separator (\f or \x0c). We check if is so and use
+        # use the OCR to get the text.
+        filtered = re.sub(r'\x0c', '', text)
 
-    return result
+        if filtered == "":
+            logging.info("Using OCR")
+            text = useOCR(file_path)
+            
+        documents.append(Document(content=text, 
+                              meta={"name": file_name}, 
+                              id_hash_keys=id_hash_keys))
+
+        
+        
+        logging.info('file conversion succesful')
+        output = {'documents': documents}
+        return output, 'output_1'
+
+    def run_batch():
+        """
+        we dont have requirement to process the multiple files in one go
+        therefore nothing here, however to use the custom node we need to have
+        this method for the class.
+        """
         
-@st.cache_resource
-def load_vulnerabilityClassifier(config_file:str = None, classifier_name:str = None):
+        return
+
+
+def basic(s:str, remove_punc:bool = False):
+
     """
-    loads the document classifier using haystack, where the name/path of model
-    in HF-hub as string is used to fetch the model object.Either configfile or 
-    model should be passed.
-    1. https://docs.haystack.deepset.ai/reference/document-classifier-api
-    2. https://docs.haystack.deepset.ai/docs/document_classifier
+    Performs basic cleaning of text.
     Params
-    --------
-    config_file: config file path from which to read the model name
-    classifier_name: if modelname is passed, it takes a priority if not \
-    found then will look for configfile, else raise error.
-    Return: document classifier model
+    ----------
+    s: string to be processed
+    removePunc: to remove all Punctuation including ',' and '.' or not
+    
+    Returns: processed string: see comments in the source code for more info
     """
+    
+    # Remove URLs
+    s = re.sub(r'^https?:\/\/.*[\r\n]*', ' ', s, flags=re.MULTILINE)
+    s = re.sub(r"http\S+", " ", s)
+
+    # Remove new line characters
+    s = re.sub('\n', ' ', s) 
 
-    # If no classifier given
+    # Remove punctuations
+    if remove_punc == True:
+      translator = str.maketrans(' ', ' ', string.punctuation) 
+      s = s.translate(translator)
+    # Remove distracting single quotes and dotted pattern
+    s = re.sub("\'", " ", s)
+    s = s.replace("..","") 
     
-    if not classifier_name:
-        if not config_file:
-            logging.warning("Pass either model name or config file")
-            return
-        else:
-            config = getconfig(config_file)
-            classifier_name = config.get('vulnerability','MODEL')
+    return s.strip()
+
+
+def paraLengthCheck(paraList, max_len = 100):
+    """
+    There are cases where preprocessor cannot respect word limit, when using 
+    respect sentence boundary flag due to missing sentence boundaries.
+    Therefore we run one more round of split here for those paragraphs
     
-    logging.info("Loading vulnerability classifier")
+    Params
+    ---------------
+    paraList : list of paragraphs/text
+    max_len : max length to be respected by sentences which bypassed 
+              preprocessor strategy
+              
+    """
+    new_para_list = []
+    for passage in paraList:
+        # check if para exceeds words limit
+        if len(passage.content.split()) > max_len:
+          # we might need few iterations example if para = 512 tokens
+          # we need to iterate 5 times to reduce para to size limit of '100'
+            iterations = int(len(passage.content.split())/max_len)
+            for i in range(iterations):
+                temp  = " ".join(passage.content.split()[max_len*i:max_len*(i+1)])
+                new_para_list.append((temp,passage.meta['page']))
+            temp  = " ".join(passage.content.split()[max_len*(i+1):])
+            new_para_list.append((temp,passage.meta['page']))
+        else:
+            # paragraphs which dont need any splitting
+            new_para_list.append((passage.content, passage.meta['page']))
     
-    # we are using the pipeline as the model is multilabel and DocumentClassifier 
-    # from Haystack doesnt support multilabel
-    # in pipeline we use 'sigmoid' to explicitly tell pipeline to make it multilabel
-    # if not then it will automatically use softmax, which is not a desired thing.
-    # doc_classifier = TransformersDocumentClassifier(
-    #                     model_name_or_path=classifier_name,
-    #                     task="text-classification",
-    #                     top_k = None)
-
-    # Download model from HF Hub
-    doc_classifier = SetFitModel.from_pretrained("leavoigt/vulnerability_multilabel")
+    logging.info("New paragraphs length {}".format(len(new_para_list)))
+    return new_para_list
+
+class UdfPreProcessor(BaseComponent):
+    """
+    class to preprocess the document returned by FileConverter. It will check
+    for splitting strategy and splits the document by word or sentences and then
+    synthetically create the paragraphs. 
+    1. https://docs.haystack.deepset.ai/docs/preprocessor
+    2. https://docs.haystack.deepset.ai/reference/preprocessor-api
+    3. https://github.com/deepset-ai/haystack/tree/main/haystack/nodes/preprocessor
+    """
+    outgoing_edges = 1
+
+    def run(self, documents:List[Document], remove_punc:bool=False, apply_clean = True,
+            split_by: Literal["sentence", "word"] = 'sentence',
+            split_length:int = 2, split_respect_sentence_boundary:bool = False,
+            split_overlap:int = 0):
+
+        """ this is required method to invoke the component in 
+        the pipeline implementation. 
+            
+        Params
+        ----------
+        documents: documents from the output dictionary returned by Fileconverter
+        remove_punc: to remove all Punctuation including ',' and '.' or not
+        split_by: document splitting strategy either as word or sentence
+        split_length: when synthetically creating the paragrpahs from document,
+                      it defines the length of paragraph.
+        split_respect_sentence_boundary: Used when using 'word' strategy for 
+        splititng of text.
+        split_overlap: Number of words or sentences that overlap when creating
+        the paragraphs. This is done as one sentence or 'some words' make sense
+        when  read in together with others. Therefore the overlap is used.
+        
+        Return
+        ---------
+        output: dictionary, with key as identifier and value could be anything 
+                we need to return. In this case the output will contain 4 objects
+                the paragraphs text list as List, Haystack document, Dataframe and 
+                one raw text file.
+        
+        output_1: As there is only one outgoing edge, we pass 'output_1' string
+      
+        """
+        
+        if split_by == 'sentence':
+            split_respect_sentence_boundary = False
     
-    # doc_classifier = pipeline("text-classification", 
-    #                         model=classifier_name, 
-    #                         return_all_scores=True, 
-    #                         function_to_apply= "sigmoid")
+        else:
+            split_respect_sentence_boundary = split_respect_sentence_boundary
+      
+        preprocessor = PreProcessor(
+            clean_empty_lines=True,
+            clean_whitespace=True,
+            clean_header_footer=True,
+            split_by=split_by,
+            split_length=split_length,
+            split_respect_sentence_boundary= split_respect_sentence_boundary,
+            split_overlap=split_overlap,
 
-    return doc_classifier
+            # will add page number only in case of PDF not for text/docx file.
+            add_page_number=True
+            )
+        
+        for i in documents:
+            # # basic cleaning before passing it to preprocessor.
+            # i = basic(i)
+            docs_processed = preprocessor.process([i])
+            if apply_clean:
+                for item in docs_processed:
+                    item.content = basic(item.content, remove_punc= remove_punc)
+            else:
+                pass
 
+        df = pd.DataFrame(docs_processed)
+        all_text = " ".join(df.content.to_list())
+        para_list = df.content.to_list()
+        logging.info('document split into {} paragraphs'.format(len(para_list)))
+        output = {'documents': docs_processed,
+                  'dataframe': df,
+                  'text': all_text,
+                  'paraList': para_list
+                 }
+        return output, "output_1"
+    def run_batch():
+        """
+            we dont have requirement to process the multiple files in one go
+            therefore nothing here, however to use the custom node we need to have
+            this method for the class.
+        """
+        return
 
-@st.cache_data
-def vulnerability_classification(haystack_doc:pd.DataFrame,
-                        threshold:float = 0.5, 
-                        classifier_model:pipeline= None
-                        )->Tuple[DataFrame,Series]:
+def processingpipeline():
     """
-    Text-Classification on the list of texts provided. Classifier provides the 
-    most appropriate label for each text. these labels are in terms of if text 
-    reference a group in a vulnerable situation.
-    ---------
-    haystack_doc: List of haystack Documents. The output of Preprocessing Pipeline 
-    contains the list of paragraphs in different format,here the list of 
-    Haystack Documents is used.
-    threshold: threshold value for the model to keep the results from classifier
-    classifiermodel: you can pass the classifier model directly,which takes priority
-    however if not then looks for model in streamlit session.
-    In case of streamlit avoid passing the model directly.
-    Returns
-    ----------
-    df: Dataframe with two columns['SDG:int', 'text']
-    x: Series object with the unique SDG covered in the document uploaded and 
-    the number of times it is covered/discussed/count_of_paragraphs. 
+    Returns the preprocessing pipeline. Will use FileConverter and UdfPreProcesor 
+    from utils.preprocessing
     """
-    logging.info("Working on vulnerability Identification")
-    haystack_doc['Vulnerability Label'] = 'NA'
-    # haystack_doc['PA_check'] = haystack_doc['Policy-Action Label'].apply(lambda x: True if len(x) != 0 else False)
-                            
-    # df1 = haystack_doc[haystack_doc['PA_check'] == True]
-    # df = haystack_doc[haystack_doc['PA_check'] == False]
-    if not classifier_model:
-        classifier_model = st.session_state['vulnerability_classifier']
-    
-        predictions = classifier_model(list(haystack_doc.text))
 
-        
+    preprocessing_pipeline = Pipeline()
+    file_converter = FileConverter()
+    custom_preprocessor = UdfPreProcessor()
 
-        pred_labels = get_vulnerability_labels(predictions)
-      
-        haystack_doc['Vulnerability Label'] = pred_labels
-    #   placeholder = {}
-    #   for j in range(len(temp)):
-    #     placeholder[temp[j]['label']] = temp[j]['score']
-    #   list_.append(placeholder)
-    # labels_ = [{**list_[l]} for l in range(len(predictions))]
-    # truth_df = DataFrame.from_dict(labels_)
-    # truth_df = truth_df.round(2)
-    # truth_df = truth_df.astype(float) >= threshold
-    # truth_df = truth_df.astype(str)
-    # categories = list(truth_df.columns)
-    # truth_df['Vulnerability Label'] = truth_df.apply(lambda x: {i if x[i]=='True' else 
-    #                                           None for i in categories}, axis=1)
-    # truth_df['Vulnerability Label'] = truth_df.apply(lambda x: list(x['Vulnerability Label'] 
-    #                                                         -{None}),axis=1)
-    # haystack_doc['Vulnerability Label'] = list(truth_df['Vulnerability Label'])
-    return haystack_doc
+    preprocessing_pipeline.add_node(component=file_converter, 
+                                    name="FileConverter", inputs=["File"])
+    preprocessing_pipeline.add_node(component = custom_preprocessor, 
+                            name ='UdfPreProcessor', inputs=["FileConverter"])
+
+    return preprocessing_pipeline