updated GARD_Search and Classify_Pipeline (#4)

- updated GARD_Search and Classify_Pipeline (8aaab5a851f02157ddbc2b2aa3551010b784825a)

epi_pipeline.py

@@ -6,7 +6,7 @@ from typing import List, Dict, Union, Optional, Set, Tuple
 
 ## This software/database is a "United States Government Work" under the terms of the United States Copyright Act. It was written as part of the author's official duties as United States Government employee and thus cannot be copyrighted. This software is freely available to the public for use. The National Center for Advancing Translational Science (NCATS) and the U.S. Government have not placed any restriction on its use or reproduction.  Although all reasonable efforts have been taken to ensure the accuracy and reliability of the software and data, the NCATS and the U.S. Government do not and cannot warrant the performance or results that may be obtained by using this software or data. The NCATS and the U.S.  Government disclaim all warranties, express or implied, including warranties of performance, merchantability or fitness for any particular purpose.  Please cite the authors in any work or product based on this material.
 
-# Written by William Kariampuzha @ NIH/NCATS. Adapted from code written by Jennifer John, et al. 
+# Written by William Kariampuzha @ NIH/NCATS. 
 # The transformer-based pipeline code has its own copyright notice under the Apache License. 
 # The code was compiled into a single python file to make adding additional features and importing into other modules easy.
 # Each section has its own import statements to facilitate clean code reuse, except for typing which applies to all.
@@ -91,7 +91,7 @@ def search_getAbs(searchterm_list:Union[List[str],List[int],str], maxResults:int
                 if pmid[0].isdigit():
                     pmids.add(pmid)
     
-    #Construct sets for filtering (right before adding abstract to pmid_abs
+    #Construct sets for filtering (right before adding abstract to pmid_abs)
     # The purpose of this is to do a second check of the abstracts, filters out any abstracts unrelated to the search terms
     #if filtering is 'lenient' or default
     if filtering !='none' or filtering !='strict':
@@ -220,70 +220,39 @@ def streamlit_getAbs(searchterm_list:Union[List[str],List[int],str], maxResults:
     
     return pmid_abs, (found, relevant)
 
-## Section: LSTM RNN Epi Classification Model (EpiClassify4GARD)
-import os
-os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
-from tensorflow.keras.preprocessing.sequence import pad_sequences
-from tensorflow.keras.preprocessing.text import tokenizer_from_json
-import tensorflow as tf
-import numpy as np
-import spacy
-import json
+## Section: Transformer based Epi Classification Model (EpiClassify4GARD)
 
+# Imports
+from transformers import AutoModelForSequenceClassification, BertTokenizer, BertConfig
 class Classify_Pipeline:
-    def __init__(self,model_name:str='LSTM_RNN_Model'):
-        #Load spaCy models
-        self.nlp = spacy.load('en_core_web_lg')
-        self.nlpSci = spacy.load("en_ner_bc5cdr_md")
-        self.nlpSci2 = spacy.load('en_ner_bionlp13cg_md')
-        # load the tokenizer
-        with open(model_name+'/tokenizer.json') as f:
-            self.classify_tokenizer = tokenizer_from_json(json.load(f))
-        #OLD Code - used pickle which is unsafe
-        #with open(model+'/tokenizer.pickle', 'rb') as handle:
-        #    import pickle
-        #    self.classify_tokenizer = pickle.load(handle)
-        # Defaults to load my_model_orphanet_final, the most up-to-date version of the classification model,
-        # but can also be run on any other tf.keras model
-    
-        # load the model
-        self.classify_model = tf.keras.models.load_model(model_name)
-        # for preprocessing
-        from nltk.corpus import stopwords
-        self.STOPWORDS = set(stopwords.words('english'))
-        # Modes 
-        self.max_length = 300
-        self.trunc_type = 'post'
-        self.padding_type = 'post'
-
+    def __init__(self, name_or_path_to_model_folder:str = "ncats/EpiClassify4GARD"):
+        #Initialize tokenizer and model
+        self.config = BertConfig.from_pretrained(name_or_path_to_model_folder)
+        self.tokenizer = BertTokenizer.from_pretrained(self.config._name_or_path, model_max_length=self.config.max_position_embeddings)
+        self.model = AutoModelForSequenceClassification.from_pretrained(name_or_path_to_model_folder,config=self.config)
+    
+    #Custom pipeline by WKariampuzha @NCATS (not Huggingface/Google/NVIDIA copyright)
     def __str__(self) -> str:
-        return "Instantiation: epi_classify = Classify_Pipeline(path_to_model_folder)" +"\n Calling: prob, isEpi = epi_classify(text) \n PubMed ID Predictions: abstracts, prob, isEpi = epi_classify.getPMIDPredictions(pmid)"
+        return "Instantiation: epi_classify = Classify_Pipeline(name_or_path_to_model_folder)" +"\n Calling: prob, isEpi = epi_classify(text) \n PubMed ID Predictions: abstracts, prob, isEpi = epi_classify.getPMIDPredictions(pmid)"
     
     def __call__(self, abstract:str) -> Tuple[float,bool]:
         return self.getTextPredictions(abstract)
     
     def getTextPredictions(self, abstract:str) -> Tuple[float,bool]:
         if len(abstract)>5:
-            # remove stopwords
-            for word in self.STOPWORDS:
-                token = ' ' + word + ' '
-                abstract = abstract.replace(token, ' ')
-                abstract = abstract.replace('  ', ' ')
-
-            # preprocess abstract
-            abstract_standard = [self.standardizeAbstract(self.standardizeSciTerms(abstract))]
-            sequence = self.classify_tokenizer.texts_to_sequences(abstract_standard)
-            padded = pad_sequences(sequence, maxlen=self.max_length, padding=self.padding_type, truncating=self.trunc_type)
-
-            y_pred1 = self.classify_model.predict(padded) # generate prediction
-            y_pred = np.argmax(y_pred1, axis=1) # get binary prediction
-
-            prob = y_pred1[0][1]
-            if y_pred == 1:
-                isEpi = True
-            else:
-                isEpi = False
-
+            # input_ids
+            input_ids = self.tokenizer(text=abstract, max_length=self.config.max_position_embeddings,padding="max_length",truncation=True,return_tensors='pt')
+            if len(input_ids)>self.config.max_position_embeddings:
+              raise InputError(f"Token Embeddings of size {input_ids} exceed length for maximum model embedding input {self.config.max_position_embeddings}.")
+              #split into sentences?
+            # softmax output is a Torch Tensor with two classes [[vector_False_class,vector_True_class]]
+            output = self.model(**input_ids)
+            # True = 1, False = 0
+            isEpi = bool(output.logits.argmax().item())
+            # softmax output is a Torch Tensor with two classes [[prob_is_False,prob_is_True]]
+            prob_tensor = output.logits.softmax(dim=-1)
+            # We only want to return the probability that it is true
+            prob = prob_tensor.data[0][1].item()
             return prob, isEpi
         else:
             return 0.0, False
@@ -292,36 +261,6 @@ class Classify_Pipeline:
         abstract = PMID_getAb(pmid)
         prob, isEpi = self.getTextPredictions(abstract)
         return abstract, prob, isEpi
-    
-    # Standardize the abstract by replacing all named entities with their entity label.
-    # Eg. 3 patients reported at a clinic in England --> CARDINAL patients reported at a clinic in GPE
-    # expects the spaCy model en_core_web_lg as input
-    def standardizeAbstract(self, abstract:str) -> str:
-        doc = self.nlp(abstract)
-        newAbstract = abstract
-        for e in reversed(doc.ents):
-            if e.label_ in {'PERCENT','CARDINAL','GPE','LOC','DATE','TIME','QUANTITY','ORDINAL'}:
-                start = e.start_char
-                end = start + len(e.text)
-                newAbstract = newAbstract[:start] + e.label_ + newAbstract[end:]
-        return newAbstract
-
-    # Same as above but replaces biomedical named entities from scispaCy models
-    # Expects as input en_ner_bc5cdr_md and en_ner_bionlp13cg_md
-    def standardizeSciTerms(self, abstract:str) -> str:
-        doc = self.nlpSci(abstract)
-        newAbstract = abstract
-        for e in reversed(doc.ents):
-            start = e.start_char
-            end = start + len(e.text)
-            newAbstract = newAbstract[:start] + e.label_ + newAbstract[end:]
-
-        doc = self.nlpSci2(newAbstract)
-        for e in reversed(doc.ents):
-            start = e.start_char
-            end = start + len(e.text)
-            newAbstract = newAbstract[:start] + e.label_ + newAbstract[end:]
-        return newAbstract
         
 ## Section: GARD SEARCH
 # can identify rare diseases in text using the GARD dictionary from neo4j
@@ -331,11 +270,18 @@ class GARD_Search:
     def __init__(self):
         import json, codecs
         #These are opened locally so that garbage collection removes them from memory
-        with codecs.open('gard-id-name-synonyms.json', 'r', 'utf-8-sig') as f:
-            diseases = json.load(f)
+        try:
+            with codecs.open('gard-id-name-synonyms.json', 'r', 'utf-8-sig') as f:
+                diseases = json.load(f)
+        except:
+            r = requests.get('https://raw.githubusercontent.com/ncats/epi4GARD/master/EpiExtract4GARD/gard-id-name-synonyms.json')
+            diseases = json.loads(r.content)
+        
         from nltk.corpus import stopwords
         STOPWORDS = set(stopwords.words('english'))
         
+        #This should be a list of all GARD IDs for purposes like random choice for testing
+        GARD_id_list = [entry['gard_id'] for entry in diseases]
         #keys are going to be disease names, values are going to be the GARD ID, set up this way bc dictionaries are faster lookup than lists
         GARD_dict = {}
         #Find out what the length of the longest disease name sequence is, of all names and synonyms. This is used by get_diseases
@@ -356,6 +302,7 @@ class GARD_Search:
                             GARD_dict[s] = entry['gard_id']
                             max_length = max(max_length,len(s.split()))
                             
+        self.GARD_id_list = GARD_id_list
         self.GARD_dict = GARD_dict
         self.max_length = max_length
     
@@ -444,6 +391,12 @@ class GARD_Search:
         print("SEARCH TERM DID NOT MATCH TO GARD DICTIONARY. SEARCHING BY USER INPUT")
         return [searchterm]
 
+    # Return a random GARD_ID Search Term list
+    def random_disease(self) -> List[str]:
+        import random
+        gard_id = random.choice(self.GARD_id_list)
+        return self.autosearch(gard_id)
+
 ## Section: BioBERT-based epidemiology NER Model (EpiExtract4GARD)
 from nltk import tokenize as nltk_tokenize
 from dataclasses import dataclass
@@ -455,6 +408,7 @@ import re
 from transformers import BertConfig, AutoModelForTokenClassification, BertTokenizer, Trainer
 from unidecode import unidecode
 from collections import OrderedDict
+import json
 import pandas as pd
 from more_itertools import pairwise
 
@@ -785,7 +739,7 @@ class NER_Pipeline:
     
 # Unattached function -- not a method
 # move this to the NER_pipeline as a method??
-#This ensures that there is a standardized ordering of df columns while ensuring dynamics with multiple models. This is used by search_term_extraction.
+# This ensures that there is a standardized ordering of df columns while ensuring dynamics with multiple models. This is used by search_term_extraction.
 def order_labels(entity_classes:Union[Set[str],List[str]]) -> List[str]:
     ordered_labels = []
     label_order = ['DIS','ABRV','EPI','STAT','LOC','DATE','SEX','ETHN']
@@ -800,7 +754,7 @@ def order_labels(entity_classes:Union[Set[str],List[str]]) -> List[str]:
 ## This section combines all of the previous code into pipelines so that usage of these models and search functions are easy to implement in apps.
 
 # Given a search term and max results to return, this will acquire PubMed IDs and Title+Abstracts and Classify them as epidemiological.
-# results = search_term_extraction(search_term, maxResults, filering, GARD_dict, classify_model_vars)
+# results = search_term_extraction(search_term, maxResults, filering, GARD_Search, Classify_Pipeline)
 #Returns a Pandas dataframe   
 def search_term_classification(search_term:Union[int,str], maxResults:int, 
                                filtering:str, rd_identify:GARD_Search, #for abstract search & filtering 
@@ -855,6 +809,11 @@ def API_search_classification(search_term:Union[int,str], maxResults:int,
         
     return results
 
+def API_PMID_classification(pmid:Union[int,str], epi_classify:Classify_Pipeline) ->  Dict[str,str]: 
+    text = PMID_getAb(pmid)
+    epi_prob, isEpi = epi_classify(text)
+    return {'PMID':pmid,'ABSTRACT':text, 'EPI_PROB':str(epi_prob), 'IsEpi':isEpi}
+
 def API_text_classification(text:str,epi_classify:Classify_Pipeline) ->  Dict[str,str]: 
     epi_prob, isEpi = epi_classify(text)
     return {'ABSTRACT':text, 'EPI_PROB':str(epi_prob), 'IsEpi':isEpi}
@@ -900,7 +859,7 @@ def search_term_extraction(search_term:Union[int,str], maxResults:int, filtering
     print(len(results),'abstracts classified as epidemiological.')
     return results.sort_values('EPI_PROB', ascending=False)
     
-#Returns a Pandas dataframe                                                                                                          
+#Returns a Pandas dataframe 
 def streamlit_extraction(search_term:Union[int,str], maxResults:int, filtering:str, #for abstract search
                            epi_ner:NER_Pipeline, #for biobert extraction 
                            GARD_Search:GARD_Search, extract_diseases:bool, #for disease extraction
@@ -1021,7 +980,7 @@ def API_text_extraction(text:str, #Text to be extracted
     else:
         json_output = ['ABSTRACT']+ordered_labels
     
-    results = {'entries':[]} 
+    extraction = dict()
     #Do the extraction
     if extract_diseases:
         extraction = epi_ner(text, GARD_Search)
@@ -1031,15 +990,17 @@ def API_text_extraction(text:str, #Text to be extracted
     if extraction:
         #Re-order the dictionary into desired JSON output
         extraction = OrderedDict([(term, extraction[term]) for term in json_output if term in extraction.keys()])
-        results['entries'].append(extraction)
+    else:
+        #This may return JSONs of different length than above
+        extraction = OrderedDict([(term, []) for term in json_output])
         
-    return results
+    return extraction
 
 def API_text_classification_extraction(text:str, #Text to be extracted
                            epi_ner:NER_Pipeline, #for biobert extraction 
                            GARD_Search:GARD_Search, extract_diseases:bool, #for disease extraction
                            epi_classify:Classify_Pipeline) ->  Dict[str,str]:
-                                                                                                                  
+
     #Format of Output
     ordered_labels = order_labels(epi_ner.labels)
     if extract_diseases:
@@ -1061,7 +1022,11 @@ def API_text_classification_extraction(text:str, #Text to be extracted
         
         #Re-order the dictionary into desired JSON output
         output = OrderedDict([(term, extraction[term]) for term in json_output if term in extraction.keys()])
-        return output
+    else:
+        #This may return JSONs of different length than above
+        output = OrderedDict([(term, []) for term in json_output])
+        
+    return output
 
 ## Section: Deprecated Functions
 import requests
@@ -1148,4 +1113,4 @@ def search_EBI_API(searchterm_list:Union[List[str],str], maxResults:int) -> Dict
                         pmids_abs[pmid] = titles[0]+' '+abstracts[0]
                         i+=1
     
-    return pmids_abs
+    return pmids_abs