File size: 8,625 Bytes
61f3f56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
import openai
import os
import argparse
import json
import ast
from multiprocessing.pool import Pool


def parse_args():
    parser = argparse.ArgumentParser(description="question-answer-generation-using-gpt-3")
    parser.add_argument("--pred_path", required=True, help="The path to file containing prediction.")
    parser.add_argument("--output_dir", required=True, help="The path to save annotation json files.")
    parser.add_argument("--output_json", required=True, help="The path to save annotation final combined json file.")
    parser.add_argument("--api_key", required=True, help="OpenAI API key.")
    parser.add_argument("--api_base", default="", type=str, help="OpenAI API base.")
    parser.add_argument("--num_tasks", required=True, type=int, help="Number of splits.")
    args = parser.parse_args()
    return args


def annotate(prediction_set, caption_files, output_dir, args):
    """
    Evaluates question and answer pairs using GPT-3 and
    returns a score for consistency.
    """
    # Set the OpenAI API key.
    openai.api_key = args.api_key
    if args.api_base is not None:
        openai.api_base = args.api_base
    for file in caption_files:
        key = file[:-5] # Strip file extension
        qa_set = prediction_set[key]
        question1 = qa_set['q1']
        question2 = qa_set['q2']
        answer = qa_set['a']
        pred1 = qa_set['pred1']
        pred2 = qa_set['pred2']
        try:
            # Compute the consistency score
            completion = openai.ChatCompletion.create(
                model="gpt-3.5-turbo",
                messages=[
                    {
                        "role": "system",
                        "content":
                            "You are an intelligent chatbot designed for evaluating the consistency of generative outputs for similar video-based question-answer pairs. "
                            "You will be given two very similar questions, a common answer common to both the questions and predicted answers for the two questions ."
                            "Your task is to compare the predicted answers for two very similar question, with a common correct answer and determine if they are consistent. Here's how you can accomplish the task:"
                            "------"
                            "##INSTRUCTIONS: "
                            "- Focus on the consistency between the two predicted answers and the correct answer. Both predicted answers should correspond to the correct answer and to each other, and should not contain any contradictions or significant differences in the conveyed information.\n"
                            "- Both predicted answers must be consistent with each other and the correct answer, in terms of the information they provide about the video content.\n"
                            "- Consider synonyms or paraphrases as valid matches, but only if they maintain the consistency in the conveyed information.\n"
                            "- Evaluate the consistency of the two predicted answers compared to the correct answer."
                    },
                    {
                        "role": "user",
                        "content":
                            "Please evaluate the following video-based question-answer pair:\n\n"
                            f"Question 1: {question1}\n"
                            f"Question 2: {question2}\n"
                            f"Correct Answer: {answer}\n"
                            f"Predicted Answer to Question 1: {pred1}\n"
                            f"Predicted Answer to Question 2: {pred2}\n\n"
                            "Provide your evaluation only as a consistency score where the consistency score is an integer value between 0 and 5, with 5 indicating the highest level of consistency. "
                            "Please generate the response in the form of a Python dictionary string with keys 'score', where its value is the consistency score in INTEGER, not STRING."
                            "DO NOT PROVIDE ANY OTHER OUTPUT TEXT OR EXPLANATION. Only provide the Python dictionary string. "
                            "For example, your response should look like this: {''score': 4.8}."
                    }
                ]
            )
            # Convert response to a Python dictionary.
            response_message = completion["choices"][0]["message"]["content"]
            response_dict = ast.literal_eval(response_message)
            result_qa_pair = [response_dict, qa_set]

            # Save the question-answer pairs to a json file.
            with open(f"{output_dir}/{key}.json", "w") as f:
                json.dump(result_qa_pair, f)

        except Exception as e:
            print(f"Error processing file '{key}': {e}")


def main():
    """
    Main function to control the flow of the program.
    """
    # Parse arguments.
    args = parse_args()

    file = open(args.pred_path)
    pred_contents = json.load(file)

    # Dictionary to store the count of occurrences for each video_id
    video_id_counts = {}
    new_pred_contents = []

    # Iterate through each sample in pred_contents
    for sample in pred_contents:
        video_id = sample['video_name']
        if video_id in video_id_counts:
            video_id_counts[video_id] += 1
        else:
            video_id_counts[video_id] = 0

        # Create a new sample with the modified key
        new_sample = sample
        new_sample['video_name'] = f"{video_id}_{video_id_counts[video_id]}"
        new_pred_contents.append(new_sample)

    # Generating list of id's and corresponding files
    id_list = [x['video_name'] for x in new_pred_contents]
    caption_files = [f"{id}.json" for id in id_list]

    output_dir = args.output_dir
    # Generate output directory if not exists.
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)

    # Preparing dictionary of question-answer sets
    prediction_set = {}
    for sample in new_pred_contents:
        id = sample['video_name']
        question1 = sample['Q1']
        question2 = sample['Q1']
        answer = sample['A']
        pred1 = sample['pred1']
        pred2 = sample['pred2']
        qa_set = {"q1": question1, "q2": question2, "a": answer, "pred1": pred1, "pred2": pred2}
        prediction_set[id] = qa_set

    # Set the OpenAI API key.
    openai.api_key = args.api_key
    num_tasks = args.num_tasks

    # While loop to ensure that all captions are processed.
    while True:
        try:
            # Files that have not been processed yet.
            completed_files = os.listdir(output_dir)
            print(f"completed_files: {len(completed_files)}")

            # Files that have not been processed yet.
            incomplete_files = [f for f in caption_files if f not in completed_files]
            print(f"incomplete_files: {len(incomplete_files)}")

            # Break the loop when there are no incomplete files
            if len(incomplete_files) == 0:
                break
            if len(incomplete_files) <= num_tasks:
                num_tasks = 1

            # Split tasks into parts.
            part_len = len(incomplete_files) // num_tasks
            all_parts = [incomplete_files[i:i + part_len] for i in range(0, len(incomplete_files), part_len)]
            task_args = [(prediction_set, part, args.output_dir, args) for part in all_parts]

            # Use a pool of workers to process the files in parallel.
            with Pool() as pool:
                pool.starmap(annotate, task_args)

        except Exception as e: 
            print(f"Error: {e}")

    # Combine all the processed files into one
    combined_contents = {}
    json_path = args.output_json

    # Iterate through json files
    for file_name in os.listdir(output_dir):
        if file_name.endswith(".json"):
            file_path = os.path.join(output_dir, file_name)
            with open(file_path, "r") as json_file:
                content = json.load(json_file)
                combined_contents[file_name[:-5]] = content

    # Write combined content to a json file
    with open(json_path, "w") as json_file:
        json.dump(combined_contents, json_file)
    print("All evaluation completed!")

    # Calculate average score
    score_sum = 0
    count = 0
    for key, result in combined_contents.items():
        count += 1
        score_match = result[0]['score']
        score = int(score_match)
        score_sum += score
    average_score = score_sum / count

    print("Average score for consistency:", average_score)


if __name__ == "__main__":
    main()