Create finetuning_evaluation.py

my_model/tabs/finetuning_evaluation.py

@@ -0,0 +1,88 @@
+import pandas as pd
+from fuzzywuzzy import fuzz
+from collections import Counter
+from nltk.stem import PorterStemmer
+from ast import literal_eval
+from typing import Union, List
+
+class KBVQAEvaluator:
+    def __init__(self, data_path: str, use_fuzzy: bool = False):
+        """
+        Initialize the VQA Processor with the dataset and configuration settings.
+        """
+        self.stemmer = PorterStemmer()
+        self.df = pd.read_excel(data_path)
+        self.scores_df = pd.read_excel(data_path, sheet_name="Scores")
+        self.df = pd.read_excel(data_path, sheet_name="Main Data")
+        self.use_fuzzy = use_fuzzy
+        self.vqa_scores = {}
+        self.exact_match_scores = {}
+
+    def stem_answers(self, answers: Union[str, List[str]]) -> Union[str, List[str]]:
+        """
+        Apply Porter Stemmer to either a single string or a list of strings.
+        """
+        if isinstance(answers, list):
+            return [" ".join(self.stemmer.stem(word.strip()) for word in answer.split()) for answer in answers]
+        else:
+            words = answers.split()
+            return " ".join(self.stemmer.stem(word.strip()) for word in words)
+
+    def calculate_vqa_score(self, ground_truths, model_answer):
+        """
+        Calculate VQA score based on the number of matching answers, with optional fuzzy matching.
+        """
+        if self.use_fuzzy:
+            fuzzy_matches = sum(fuzz.partial_ratio(model_answer, gt) >= 80 for gt in ground_truths)
+            return min(fuzzy_matches / 3, 1)
+        else:
+            count = Counter(ground_truths)
+            return min(count.get(model_answer, 0) / 3, 1)
+
+    def calculate_exact_match_score(self, ground_truths, model_answer):
+        """
+        Calculate Exact Match score, with optional fuzzy matching.
+        """
+        if self.use_fuzzy:
+            return int(any(fuzz.partial_ratio(model_answer, gt) >= 80 for gt in ground_truths))
+        else:
+            return int(model_answer in ground_truths)
+
+    def evaluate(self):
+        """
+        Process the DataFrame: stem answers, calculate scores, and store results.
+        """
+        self.df['raw_answers_stemmed'] = self.df['raw_answers'].apply(literal_eval).apply(self.stem_answers)
+        model_configurations = ['caption+detic', 'caption+yolov5', 'only_caption', 'only_detic', 'only_yolov5']
+        model_names = ['13b', '7b']
+        
+        for name in model_names:
+            for config in model_configurations:
+                full_config = f'{name}_{config}'
+                self.df[f'{full_config}_stemmed'] = self.df[full_config].apply(self.stem_answers)
+                
+                self.df[f'vqa_score_{full_config}'] = self.df.apply(lambda x: self.calculate_vqa_score(x['raw_answers_stemmed'], x[f'{full_config}_stemmed']), axis=1)
+                self.df[f'exact_match_score_{full_config}'] = self.df.apply(lambda x: self.calculate_exact_match_score(x['raw_answers_stemmed'], x[f'{full_config}_stemmed']), axis=1)
+                
+                self.vqa_scores[full_config] = round(self.df[f'vqa_score_{full_config}'].mean()*100, 2)
+                self.exact_match_scores[full_config] = round(self.df[f'exact_match_score_{full_config}'].mean()*100, 2)
+                
+    def save_results(self):
+        # Create a DataFrame for the scores
+        scores_data = {
+            'Model Configuration': list(self.vqa_scores.keys()),
+            'VQA Score': list(self.vqa_scores.values()),
+            'Exact Match Score': list(self.exact_match_scores.values())
+        }
+        scores_df = pd.DataFrame(scores_data)
+
+        # Saving the scores DataFrame to an Excel file
+        with pd.ExcelWriter('evaluation_results_final.xlsx', engine='openpyxl', mode='w') as writer:
+            scores_df.to_excel(writer, sheet_name='Scores', index=False)
+
+def run_evaluator(self):
+    evaluator = KBVQAEvaluator('evaluation_results_final.xlsx', use_fuzzy=False)
+    evaluator.evaluate()
+    st.table(self.vqa_scores)
+    print(evaluator.exact_match_scores)
+    evaluator.save_results()