evaluation.py

import os, llava, argparse
import numpy as np
from mmengine import load, dump
from tqdm import tqdm
from collections import defaultdict


PROMPT_TEMPLATES = {
    "instruction": "Evaluate if this video follows the instruction: '{instruction}'. Use the following scoring criteria:\n\n- 0: The video does not follow the instruction at all.\n- 1: The video includes the correct object but performs the wrong action, or vice versa.\n- 2: The video follows the instruction and shows a tendency toward the intended action but does not fully achieve the goal.\n- 3: The video follows the instruction precisely and successfully achieves the intended goal.\n\nLet's analyze step-by-step and conclude with 'Score: [score]'.",
    "physical_laws": 'Watch the video and determine if it shows any \'{physical_laws}\' Let\'s think step-by-step and conclude with "Yes" or "No".',
    "commonsense": 'Does the video exhibit \'{commonsense}\'? Let\'s think step-by-step and conclude with "Yes" or "No".',
}

QUESTION_POOL = {
    "instruction": None,
    "physical_laws": [
        "Violation of Newton's Law: Objects move without any external force.",
        "Violation of the Law of Conservation of Mass or Solid Constitutive Law: Objects deform or distort irregularly.",
        "Violation of Fluid Constitutive Law: Liquids flow in an unnatural or irregular manner.",
        "Violation of Non-physical Penetration: Objects unnaturally pass through each other.",
        "Violation of Gravity: Objects behave inconsistently with gravity, such as floating in the air.",
    ],
    "common_sense": [
        "Poor Aesthetics: Visually unappealing or low-quality content.",
        "Temporal Inconsistency: Noticeable flickering, choppy motion, or abrupt appearance/disappearance of irrelevant objects.",
    ],
}

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Script for evaluating the WorldModelBenchmark.")
    parser.add_argument(
        "--judge",
        type=str,
        help="Path to judge model checkpoint.",
    )
    parser.add_argument(
        "--video_dir",
        type=str,
        help="Path to the generated video directory.",
    )
    parser.add_argument(
        "--save_name",
        type=str,
        help="Path to save evaluation results.",
    )
    parser.add_argument("--cot", action="store_true", help="Enable or disable Chain-of-Thought output.")
    args = parser.parse_args()

    validation_set = load("./worldmodelbench.json")
    if args.cot:
        args.save_name += "_cot"
    results = None
    if os.path.exists(args.save_name):
        results = load(args.save_name)
        try:
            preds = results["preds"]
            accs = results["accs"]
        except:
            raise "Expected keys are not found in the results."
    else:
        model = llava.load(args.judge)

        preds = dict()
        accs = defaultdict(list)
        for vid, v_i in tqdm(enumerate(validation_set), total=len(validation_set)):
            ## Load video
            video_name = v_i["first_frame"].split("/")[-1].split(".")[0]
            video = os.path.join(args.video_dir, video_name + ".mp4")
            video = llava.Video(video)
            ## Traverse criterions
            for k in ["instruction", "physical_laws", "commonsense"]:
                preds_i = []
                prompt_template = PROMPT_TEMPLATES[k]
                qs = QUESTION_POOL[k]
                if qs is not None:
                    accs_i = []
                    for q in qs:
                        if k == "physical_laws":
                            text_prompt = prompt_template.format(physical_laws=k.lower())
                        else:
                            text_prompt = prompt_template.format(commonsense=k.lower())
                        if not args.cot:
                            text_prompt = text_prompt.replace(
                                "Let's think step-by-step and conclude with", "Answer with"
                            ).replace("Let's analyze step-by-step and conclude with", "Answer with")
                        pred = model.generate_content([video, text_prompt])
                        preds_i.append(pred)
                        ## Always ask for violations, so a "No" is preferred!
                        accs_i.append("no" in pred.lower())
                    accs[k].append(np.mean(accs_i))
                else:
                    text_prompt = prompt_template.format(instruction=v_i["text_instruction"])
                    if not args.cot:
                        text_prompt = text_prompt.replace(
                            "Let's think step-by-step and conclude with", "Answer with"
                        ).replace("Let's analyze step-by-step and conclude with", "Answer with")
                    pred = model.generate_content([video, text_prompt])
                    preds_i.append(pred)
                    try:
                        score = float(pred.split(":")[-1].strip(" ."))
                    except:
                        score = 0
                    accs[k].append(score / 3)
                if video_name not in preds:
                    preds[video_name] = dict()
                preds[video_name][k] = preds_i
    ## Print results
    for k, v in accs.items():
        if isinstance(v, list):
            print(f"{k} accuracy: {np.mean(v) * 100}%.")
        else:
            print(f"{k} accuracy: {v}%.")
    ## Save results
    if results is None:
        results = {"preds": preds, "accs": accs}
        dump(results, f"./{args.save_name}.json", indent=4)