def vizwiz_vqa_process_results()

in lmms_eval/tasks/vizwiz_vqa/utils.py [0:0]

• 25 lines of code
• 10 McCabe index (conditional complexity)

def vizwiz_vqa_process_results(doc, result):
    eval_ai_processor = EvalAIAnswerProcessor()
    assert len(result) == 1, f"The result should be a list of length 1, but got {len(result)}."
    resAns = eval_ai_processor(result[0])
    accuracy = 0

    if "answers" in doc and doc["answers"] is not None:
        gtAcc = []

        for i in range(len(doc["answers"])):
            doc["answers"][i] = eval_ai_processor(doc["answers"][i])

        for i in range(len(doc["answers"])):
            otherGTAns = [doc["answers"][j] for j in range(len(doc["answers"])) if i != j]
            matchingAns = [item for item in otherGTAns if item == resAns]
            acc = min(1, float(len(matchingAns)) / 3)
            gtAcc.append(acc)
        if gtAcc:
            accuracy = statistics.mean(gtAcc)
        else:
            accuracy = 0

    return {
        "exact_match": accuracy,
        "submission": {
            "image": f"{doc['question_id']}.jpg",
            "answer": resAns,
        },
    }