get_results.py

import json
import sys
import numpy as np
from utils import *
from tqdm import tqdm
from pydantic import BaseModel
from openai import OpenAI
from typing import List
import multiprocessing as mp
from functools import partial
from glob import glob


class AnswerScore(BaseModel):
    score: int

data_functions = {
    # medical
    "medicalmmt": medicalmmt_eval,
    "medicalMMMU": medicalMMMU_eval,
    "medicalMMMUPro": medicalMMMU_eval,

    # cultural
    "celebvqa": celebvqa_eval,
    "foodvqa": foodvqa_eval,
    "countriesvqa": countriesvqa_eval,

    # agro
    "agrovqa": agrovqa_eval,

    # chart and diagrams
    "iconqa": iconqa_eval,
    "chartqa": chartqa_eval,
    "diagramsMMMU": diagramsMMMU_eval,
    "diagramsvqa": diagramsvqa_eval,
    "tablesvqa": tablesvqa_eval,

    # video
    "culturevideovqa": culturevideovqa_eval,
    "videomme": videomme_eval,

    # ocr
    "ocrisi": ocrisi_eval,
    "khatt": khatt_eval,
    "isidocvqa": isidocvqa_eval,
    "patddocvqa": patddocvqa_eval,
    "patsocr": patsocr_eval,
    "evarest": evarest_eval,
    "historicalbooks": historicalbooks_eval,
    "arabicocr": arabicocr_eval,
    
    # vqa
    "mme": mme_eval,
    "mmbench": mmbench_eval,
    "vqammt": vqammt_eval,
    "seed": seed_eval,
    "mmmu": mmmu_eval,
    "countbench": countbench_eval,
    "hallucinationmmt": hallucinationmmt_eval,
    "pope": pope_eval,
    "scienceqa": scienceqa_eval,
    "examsv": examsv_eval,
    "gqa": gqa_eval,
    "vizwiz": vizwiz_eval,
    "infographicsvqa": infographicsvqa_eval,
    "blink": blink_eval,
    "realworldqa": realworldqa_eval,
    "mutliimagemmt": mutliimagemmt_eval,
    "muribench": muribench_eval,
    "objectcoco": objectcoco_eval,  
}

fuzz = {
    "mtvqa": mtvqa_user_prompt,
    "geochat": geochat_user_prompt,
    "mathvista": mathvista_user_prompt,
    "vqav2": vqav2_user_prompt,
}

medical_data = ["medicalmmt", "medicalMMMU", "medicalMMMUPro"]
medical_results = {}

cultural_data = ["celebvqa", "foodvqa", "countriesvqa"]
cultural_results = {}

agro_data = ["agrovqa"]
agro_results = {}

charts_data = ["iconqa", "chartqa", "diagramsMMMU", "diagramsvqa", "tablesvqa"]
charts_results = {}

remotesensing_data = ["geochat"]
remotesensing_results = {}

video_data = ["culturevideovqa", "videomme"]
video_results = {}

ocr_data = ["ocrisi", "khatt", "isidocvqa", "patddocvqa", "patsocr", "mtvqa", "evarest", "historicalbooks", "arabicocr"]
ocr_results = {}

vqa_data = ["mme", "mmbench", "vqammt", "seed", "mmmu", "countbench", "hallucinationmmt", "pope", "mathvista", "scienceqa", "examsv", "gqa", "vizwiz", "vqav2", "infographicsvqa", "blink", "realworldqa", "mutliimagemmt", "muribench", "objectcoco"]
vqa_results = {}


def eval_gpt(row, user_prompt):
    client = OpenAI()
    question = row['question'].split("\n")[0]
    pred = row['pred_answer']
    pred = pred.split("assistant\n")[-1].strip()
    gt = row['answer']

    messages = [
        {
            "role": "system",
            "content": fuzz_eval_system_prompt,
        },
        {
            "role": "user",
            "content": user_prompt.format(question=question, pred=pred, gt=gt)
        },
    ]

    completion = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=messages,
        max_tokens=300,
        tools=[
            {
                "type": "function",
                "function": {
                    "name": "answer_score",
                    "description": "Provide a [0, 1] score to the semantic similarity between two sentences",
                    "parameters": AnswerScore.model_json_schema(),
                },
            }
        ],
        tool_choice={"type": "function", "function": {"name": "answer_score"}},
    )

    vqa_answer = AnswerScore.model_validate_json(
        completion.choices[0].message.tool_calls[0].function.arguments
    )
    return {
        'index': row['index'],
        'question': question,
        'pred_answer': pred,
        'answer': gt,
        'evaluation': vqa_answer.score
    }

def process_chunk(user_prompt, chunk):
    d = []
    for row in chunk:
        try:
            d.append(eval_gpt(row, user_prompt))
        except Exception as e:
            print("ERROR", e)
            continue
    return d 

def fuzz_eval(user_prompt, data):
    num_cores = mp.cpu_count()
    chunk_size = len(data) // num_cores
    chunks = [data[i:i + chunk_size] for i in range(0, len(data), chunk_size)]
    pool = mp.Pool(num_cores)
    results = []
    process_chunk_f = partial(process_chunk, user_prompt)
    with tqdm(total=len(data)) as pbar:
        for chunk_result in pool.imap_unordered(process_chunk_f, chunks):
            results.extend(chunk_result)
            pbar.update(len(chunk_result))

    pool.close()
    pool.join()

    correct_count = sum(1 for item in results if item['evaluation'] == 1)
    total_count = len(results)
    return round(correct_count * 100 / total_count, 2)


MODEL = "peacock"
files = glob(f"results/{MODEL}_*.json")
for file in files:
    name = file.split(f"_")[-1].replace(".json", "")
    print(name)
    with open(file, "r") as f:
        data = json.load(f)
    if len(data) == 0: continue
    
    accuracy = 0
    if name in fuzz:
        accuracy = fuzz_eval(fuzz[name], data)
    else:
        tot = 0
        for r in data:
            tot += data_functions[name](r["pred_answer"], r["answer"])
        accuracy = round(tot * 100 / len(data), 2)
        print(f"{name}: {tot} / {len(data)} -> {accuracy:.2f}")

    if name in medical_data:
        medical_results[name] = accuracy
    elif name in cultural_data:
        cultural_results[name] = accuracy
    elif name in agro_data:
        agro_results[name] = accuracy
    elif name in charts_data:
        charts_results[name] = accuracy
    elif name in remotesensing_data:
        remotesensing_results[name] = accuracy
    elif name in video_data:
        video_results[name] = accuracy
    elif name in ocr_data:
        ocr_results[name] = accuracy
    elif name in vqa_data:
        vqa_results[name] = accuracy


from pprint import pprint
print("\nMedical Results")
pprint(medical_results)
if len(medical_results) > 0:
    print("Medical average:", round(sum(list(medical_results.values())) / len(medical_results), 2))

print("\ncultural Results")
pprint(cultural_results)
if len(cultural_results) > 0:
    print("cultural average:", round(sum(list(cultural_results.values())) / len(cultural_results), 2))

print("\nagro Results")
pprint(agro_results)
if len(agro_results) > 0:
    print("agro average:", round(sum(list(agro_results.values())) / len(agro_results), 2))

print("\ncharts Results")
pprint(charts_results)
if len(charts_results) > 0:
    print("charts average:", round(sum(list(charts_results.values())) / len(charts_results), 2))

print("\nremotesensing Results")
pprint(remotesensing_results)
if len(remotesensing_results) > 0:
    print("remotesensing average:", round(sum(list(remotesensing_results.values())) / len(remotesensing_results), 2))

print("\nvideo Results")
pprint(video_results)
if len(video_results) > 0:
    print("video average:", round(sum(list(video_results.values())) / len(video_results), 2))

print("\nocr Results")
pprint(ocr_results)
if len(ocr_results) > 0:
    print("ocr average:", round(sum(list(ocr_results.values())) / len(ocr_results), 2))

print("\nvqa Results")
pprint(vqa_results)
if len(vqa_results) > 0:
    print("vqa average:", round(sum(list(vqa_results.values())) / len(vqa_results), 2))