worldmodelbench / evaluation.py
Fang Yunhao
Update.
c0380af
raw
history blame
13.7 kB
from dataclasses import dataclass, field
from enum import Enum
from pathlib import Path
from typing import Dict, List, Optional, Union
import logging
import os
from rich.console import Console
from rich.table import Table
from rich.panel import Panel
# from rich.progress import track
# from rich import print as rprint
from rich.progress import Progress, BarColumn, TimeRemainingColumn
import numpy as np
from mmengine import load, dump
from collections import defaultdict
class EvaluationType(Enum):
INSTRUCTION = "instruction"
PHYSICAL_LAWS = "physical_laws"
COMMON_SENSE = "common_sense"
def get_default_prompt_templates() -> Dict[str, str]:
"""Factory function for default prompt templates."""
return {
EvaluationType.INSTRUCTION.value: """
Evaluate if this video follows the instruction: '{instruction}'.
Use the following scoring criteria:
- 0: The video does not follow the instruction at all.
- 1: The video includes the correct object but performs the wrong action, or vice versa.
- 2: The video follows the instruction and shows a tendency toward the intended goal.
- 3: The video follows the instruction precisely and successfully achieves the goal.
Let's analyze step-by-step and conclude with 'Score: [score]'.
""".strip(),
EvaluationType.PHYSICAL_LAWS.value: """
Watch the video and determine if it shows any '{physical_laws}'
Let's think step-by-step and conclude with "Yes" or "No".
""".strip(),
EvaluationType.COMMON_SENSE.value: """
Does the video exhibit '{common_sense}'?
Let's think step-by-step and conclude with "Yes" or "No".
""".strip(),
}
def get_default_question_pool() -> Dict[str, Optional[List[str]]]:
"""Factory function for default question pool."""
return {
EvaluationType.INSTRUCTION.value: None,
EvaluationType.PHYSICAL_LAWS.value: [
"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 irregularly.",
"Violation of Fluid Constitutive Law: Liquids flow in an unnatural manner.",
"Violation of Non-physical Penetration: Objects unnaturally pass through each other.",
"Violation of Gravity: Objects behave inconsistently with gravity.",
],
EvaluationType.COMMON_SENSE.value: [
"Poor Aesthetics: Visually unappealing or low-quality content.",
"Temporal Inconsistency: Noticeable flickering or abrupt changes.",
],
}
@dataclass
class EvaluationConfig:
"""Configuration for evaluation prompts and scoring criteria."""
PROMPT_TEMPLATES: Dict[str, str] = field(default_factory=get_default_prompt_templates)
QUESTION_POOL: Dict[str, Optional[List[str]]] = field(default_factory=get_default_question_pool)
class ResultsPrinter:
"""Handles formatted output of evaluation results."""
def __init__(self):
self.console = Console()
def print_header(self, text: str):
"""Print a styled header."""
self.console.print(f"\n[bold blue]{text}[/bold blue]")
def print_score(self, category: str, score: float, indent: int = 0):
"""Print a score with proper formatting."""
indent_str = " " * indent
self.console.print(f"{indent_str}[cyan]{category}:[/cyan] [yellow]{score:.2f}[/yellow]")
def create_results_table(self, category: str, scores: Dict[str, float]) -> Table:
"""Create a rich table for displaying results."""
table = Table(title=f"{category} Results", show_header=True, header_style="bold magenta")
table.add_column("Metric", style="cyan")
table.add_column("Score", justify="right", style="yellow")
for metric, score in scores.items():
table.add_row(metric, f"{score:.2f}")
return table
def print_summary_panel(self, total_score: float, num_categories: int):
"""Print a panel with summary information."""
panel = Panel(
f"[bold green]Total Score: {total_score:.2f}[/bold green]\n",
# f"[blue]Average per category: {total_score/num_categories:.2f}[/blue]",
title="Evaluation Summary",
border_style="green"
)
self.console.print(panel)
class WorldModelEvaluator:
"""Evaluates world model benchmark videos using VILA model."""
def __init__(self, judge_path: str, video_dir: str, config: EvaluationConfig):
self.judge = self._load_judge(judge_path)
self.video_dir = Path(video_dir)
self.config = config
self.logger = logging.getLogger(__name__)
self.printer = ResultsPrinter()
@staticmethod
def _load_judge(judge_path: str):
"""Load the VILA judge model."""
import llava
return llava.load(judge_path)
def _load_video(self, video_name: str) -> Optional['llava.Video']:
"""Load a video file for evaluation."""
video_path = self.video_dir / f"{video_name}.mp4"
if not video_path.exists():
self.logger.warning(f"Video not found: {video_path}")
return None
import llava
return llava.Video(str(video_path))
def evaluate_video(self, video: 'llava.Video', prompt: str, cot: bool = True) -> str:
"""Generate evaluation content for a video."""
if not cot:
prompt = 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"
)
return self.judge.generate_content([video, prompt])
def process_results(self, preds: Dict, accs: defaultdict) -> float:
"""Process and print evaluation results with rich formatting."""
num_insts = len(preds)
total_score = 0
category_mapping = {
2: [("framewise", "temporal")],
5: [("newton", "mass", "fluid", "penetration", "gravity")]
}
for category, scores in accs.items():
self.printer.print_header(f"{category.replace('_', ' ').title()} Details")
num_sub = len(scores) // num_insts
if num_sub == 1:
overall = np.mean(scores)
self.printer.print_score("Overall", overall)
total_score += overall
elif num_sub in category_mapping:
sub_scores = {}
for i, sub in enumerate(category_mapping[num_sub][0]):
sub_mean = np.mean(scores[i::num_sub])
sub_scores[sub.title()] = sub_mean
# Create and display results table
table = self.printer.create_results_table(
category.replace('_', ' ').title(),
sub_scores
)
self.printer.console.print(table)
overall = np.sum(list(sub_scores.values()))
self.printer.print_score("Overall", overall, indent=2)
total_score += overall
else:
raise ValueError(f"Unexpected number of subcategories: {num_sub}")
self.printer.print_summary_panel(total_score, len(accs))
return total_score
def save_results(results: Dict, save_path: str):
"""Save evaluation results to a file."""
dump(results, save_path, indent=4)
Console().print(f"[green]Results saved to: {save_path}[/green]")
class RichLogHandler(logging.Handler):
"""Custom logging handler that uses Rich for formatting."""
def __init__(self):
super().__init__()
self.console = Console()
def emit(self, record):
try:
msg = self.format(record)
style = "bold red" if record.levelno >= logging.WARNING else "blue"
self.console.print(f"[{style}]{msg}[/{style}]")
except Exception:
self.handleError(record)
def main():
import argparse
parser = argparse.ArgumentParser(description="Evaluate World Model Benchmark")
parser.add_argument("--judge", type=str, required=True, help="Path to judge model checkpoint")
parser.add_argument("--video_dir", type=str, required=True, help="Path to generated video directory")
parser.add_argument("--model_name", type=str, required=True, help="Tested model name")
parser.add_argument("--save_name", type=str, default="worldmodelbench_results", help="Path to save evaluation results")
parser.add_argument("--cot", action="store_true", help="Enable Chain-of-Thought output")
parser.add_argument("--no-save", action="store_true", help="Disable saving results")
args = parser.parse_args()
# Setup logging with custom Rich handler
logging.basicConfig(
level=logging.INFO,
format="%(message)s",
handlers=[RichLogHandler()]
)
logger = logging.getLogger(__name__)
# Initialize evaluator
config = EvaluationConfig()
evaluator = WorldModelEvaluator(args.judge, args.video_dir, config)
printer = ResultsPrinter()
# Load validation set with status message
printer.console.print("[bold]Loading validation set...[/bold]")
validation_set = load("./worldmodelbench.json")
# Check for existing results
save_path = f"{args.save_name}_cot.json" if args.cot else f"{args.save_name}.json"
if os.path.exists(save_path):
printer.console.print("[bold yellow]Loading existing results...[/bold yellow]")
results = load(save_path)
try:
preds, accs = results["preds"], results["accs"]
except KeyError:
raise KeyError("Expected keys not found in results file")
else:
printer.console.print("[bold green]Starting new evaluation...[/bold green]")
preds = {}
accs = defaultdict(list)
# Create a single progress instance for all operations
with Progress(
"[progress.description]{task.description}",
BarColumn(),
"[progress.percentage]{task.percentage:>3.0f}%",
TimeRemainingColumn(),
console=printer.console
) as progress:
# Main task for video processing
video_task = progress.add_task("Processing videos", total=len(validation_set))
for vid, v_i in enumerate(validation_set):
video_name = Path(v_i["first_frame"]).stem
video = evaluator._load_video(video_name)
if not video:
progress.advance(video_task)
continue
# Evaluation task
eval_task = progress.add_task(
f"Evaluating {video_name}",
total=len(EvaluationType)
)
for eval_type in EvaluationType:
preds_i = []
prompt_template = config.PROMPT_TEMPLATES[eval_type.value]
questions = config.QUESTION_POOL[eval_type.value]
if questions:
accs_i = []
# Questions task
question_task = progress.add_task(
f"Processing {eval_type.value} questions",
total=len(questions)
)
for question in questions:
format_kwargs = {
f"{eval_type.value}": question.lower()
}
prompt = prompt_template.format(**format_kwargs)
pred = evaluator.evaluate_video(video, prompt, args.cot)
preds_i.append(pred)
accs_i.append("no" in pred.lower())
progress.advance(question_task)
progress.remove_task(question_task)
accs[eval_type.value].extend(accs_i)
else:
prompt = prompt_template.format(instruction=v_i["text_instruction"])
pred = evaluator.evaluate_video(video, prompt, args.cot)
preds_i.append(pred)
try:
score = float(pred.split(":")[-1].strip(" ."))
except ValueError:
logger.warning(f"Could not parse score from prediction: {pred}")
score = 0
accs[eval_type.value].append(score)
if video_name not in preds:
preds[video_name] = {}
preds[video_name][eval_type.value] = preds_i
progress.advance(eval_task)
progress.remove_task(eval_task)
progress.advance(video_task)
# Save results if requested
if not args.no_save:
results = {"model_name": args.model_name, "preds": preds, "accs": accs}
save_results(results, save_path)
# Process and display results
printer.console.print("\n[bold]Final Evaluation Results[/bold]")
total_score = evaluator.process_results(preds, accs)
if __name__ == "__main__":
main()