Upload full dataset
fbff792 verified 7 months ago
26.2 kB
"""
Parallel VQA evaluation using Gemini API (google-genai SDK).
Usage:
    set GOOGLE_API_KEY=your_key
    python eval.py [--model MODEL] [--workers N] [--rpm N]
Results saved to eval_results.json (supports resume).
"""

import argparse
import json
import os
import re
import signal
import sys
import threading
import time
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor, as_completed
from datetime import datetime, timezone
from pathlib import Path

if sys.platform == "win32" and hasattr(sys.stdout, "reconfigure"):
    sys.stdout.reconfigure(encoding="utf-8")  # type: ignore[union-attr]

# ── Dependency check ─────────────────────────────────────────────────

try:
    from google import genai
    from google.genai import types as genai_types
except ImportError:
    print("google-genai SDK not found.")
    print("Install:  pip install google-genai")
    sys.exit(1)

# ── Config ───────────────────────────────────────────────────────────

BASE_DIR = Path(__file__).parent
CLIPS_DIR = BASE_DIR / "clips"
META_FULL = BASE_DIR / "meta_full.json"
RESULTS_FILE = BASE_DIR / "eval_results.json"

DEFAULT_MODEL = "gemini-3.1-pro-preview"
DEFAULT_WORKERS = 6
DEFAULT_RPM = 10
DEFAULT_THINKING_BUDGET = 8192


# ── Rate limiter ─────────────────────────────────────────────────────

class RateLimiter:
    """Simple per-minute rate limiter (thread-safe)."""

    def __init__(self, rpm: int):
        self._interval = 60.0 / rpm
        self._lock = threading.Lock()
        self._last = 0.0

    def wait(self):
        with self._lock:
            now = time.monotonic()
            gap = self._last + self._interval - now
            if gap > 0:
                time.sleep(gap)
            self._last = time.monotonic()


# ── Helpers ──────────────────────────────────────────────────────────

def build_prompt(question: str, choices: list[str]) -> str:
    opts = "\n".join(f"{chr(65 + i)}. {c}" for i, c in enumerate(choices))
    return (
        "Watch the video clip carefully, then answer the multiple-choice "
        "question.\n\n"
        f"Question: {question}\n\n"
        f"{opts}\n\n"
        "Respond with ONLY the letter (A, B, C, or D) of the correct answer."
    )


def parse_answer(text: str) -> int:
    """Parse model response -> answer index (0-3), or -1 if unparseable."""
    text = text.strip()
    if not text:
        return -1
    # 1) First character is a letter
    if text[0].upper() in "ABCD":
        return ord(text[0].upper()) - 65
    # 2) Pattern like "A." / "A)" / "A:"
    m = re.search(r"\b([A-Da-d])[.):\s]", text)
    if m:
        return ord(m.group(1).upper()) - 65
    # 3) Standalone letter
    m = re.search(r"\b([A-Da-d])\b", text)
    if m:
        return ord(m.group(1).upper()) - 65
    return -1


# ── Core evaluation ──────────────────────────────────────────────────

SAFETY = [
    genai_types.SafetySetting(
        category=genai_types.HarmCategory.HARM_CATEGORY_HARASSMENT,
        threshold=genai_types.HarmBlockThreshold.BLOCK_NONE,
    ),
    genai_types.SafetySetting(
        category=genai_types.HarmCategory.HARM_CATEGORY_HATE_SPEECH,
        threshold=genai_types.HarmBlockThreshold.BLOCK_NONE,
    ),
    genai_types.SafetySetting(
        category=genai_types.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
        threshold=genai_types.HarmBlockThreshold.BLOCK_NONE,
    ),
    genai_types.SafetySetting(
        category=genai_types.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
        threshold=genai_types.HarmBlockThreshold.BLOCK_NONE,
    ),
]


def _build_item_meta(item: dict, idx: int) -> dict:
    """Extract all metadata fields common to both success and error results."""
    tags = item.get("tag", [])
    correct = item.get("answer", -1)
    choices = item.get("choices", [])
    return {
        "idx": idx,
        # ── identification ──
        "task": item.get("task", ""),
        "scenario": item.get("scenario", ""),
        "source_video": item.get("source_video", ""),
        "clip": item.get("clip", ""),
        "duration": item.get("duration", []),
        "perspective": tags[0] if len(tags) >= 1 else "",
        "question_type": tags[1] if len(tags) >= 2 else "",
        "tag": tags,
        # ── question content ──
        "question": item.get("question", ""),
        "choices": choices,
        # ── ground truth ──
        "correct_answer": correct,
        "correct_letter": chr(65 + correct) if 0 <= correct < len(choices) else "?",
        "correct_text": choices[correct] if 0 <= correct < len(choices) else "",
    }


def _extract_thinking_and_response(resp) -> tuple[str, str, str]:
    """Separate thinking/response from a Gemini response.
    Returns (thinking, response, diagnosis).
    diagnosis is empty on success, otherwise explains the failure.
    """
    thinking_parts: list[str] = []
    response_parts: list[str] = []

    # No candidates at all → blocked before generation
    if not resp.candidates:
        reason = "no candidates"
        if hasattr(resp, "prompt_feedback") and resp.prompt_feedback:
            reason += f" | prompt_feedback={resp.prompt_feedback}"
        return "", "", reason

    cand = resp.candidates[0]

    # Candidate exists but finish_reason indicates a block
    finish = getattr(cand, "finish_reason", None)
    if finish and str(finish) not in ("STOP", "FinishReason.STOP", "1", "MAX_TOKENS",
                                       "FinishReason.MAX_TOKENS", "2"):
        safety_ratings = getattr(cand, "safety_ratings", None)
        return "", "", f"finish_reason={finish} | safety={safety_ratings}"

    # Candidate content might be None
    content = getattr(cand, "content", None)
    if content is None:
        return "", "", "candidate.content is None"

    parts = getattr(content, "parts", None)
    if not parts:
        return "", "", "candidate.content.parts is empty"

    for part in parts:
        text = getattr(part, "text", None) or ""
        if not text:
            continue
        if hasattr(part, "thought") and part.thought:
            thinking_parts.append(text)
        else:
            response_parts.append(text)

    thinking = "\n".join(thinking_parts)
    response = "\n".join(response_parts)

    if not response.strip():
        # Model only produced thinking, no actual answer
        return thinking, "", "only thinking returned, no answer part"

    return thinking, response, ""


def evaluate_one(
    client: genai.Client,
    model_name: str,
    item: dict,
    idx: int,
    limiter: RateLimiter,
    thinking_budget: int = 0,
    max_retries: int = 3,
) -> dict:
    """Upload clip -> generate -> parse -> cleanup.  Returns result dict."""
    clip_path = BASE_DIR / item["clip"]
    prompt = build_prompt(item["question"], item["choices"])

    meta = _build_item_meta(item, idx)
    meta["prompt"] = prompt  # always record what we sent

    if not clip_path.exists():
        return {**meta, "error": f"clip not found: {clip_path.name}"}

    last_err = ""

    # Build generation config (with optional thinking)
    gen_kwargs: dict = {
        "temperature": 0,
        "max_output_tokens": 64,
        "safety_settings": SAFETY,
    }
    if thinking_budget > 0:
        gen_kwargs["thinking_config"] = genai_types.ThinkingConfig(
            thinking_budget=thinking_budget,
        )

    for attempt in range(max_retries):
        uploaded = None
        try:
            limiter.wait()

            # Upload
            uploaded = client.files.upload(
                file=str(clip_path),
                config=genai_types.UploadFileConfig(mime_type="video/mp4"),
            )

            # Poll until processing completes
            for _ in range(60):
                if uploaded.state and uploaded.state.name != "PROCESSING":
                    break
                time.sleep(2)
                uploaded = client.files.get(name=uploaded.name or "")

            if uploaded.state and uploaded.state.name == "FAILED":
                last_err = "file processing failed"
                continue

            # Generate
            resp = client.models.generate_content(
                model=model_name,
                contents=[uploaded, prompt],
                config=genai_types.GenerateContentConfig(**gen_kwargs),
            )

            thinking_text, resp_text, diag = _extract_thinking_and_response(resp)

            if diag or not resp_text.strip():
                last_err = diag or "empty response"
                continue

            pred = parse_answer(resp_text)
            choices = item.get("choices", [])
            valid = 0 <= pred < len(choices)

            return {
                **meta,
                # ── model I/O ──
                "thinking": thinking_text,
                "response": resp_text.strip(),
                # ── model prediction ──
                "predicted": pred if valid else -1,
                "predicted_letter": chr(65 + pred) if valid else "?",
                "predicted_text": choices[pred] if valid else "",
                # ── evaluation ──
                "is_correct": (pred == item.get("answer", -1)) if valid else False,
                "missing": not valid,  # True = unparseable / bad format
            }

        except Exception as e:
            last_err = str(e)
            if attempt < max_retries - 1:
                time.sleep(2 ** attempt * 5)
        finally:
            if uploaded:
                try:
                    client.files.delete(name=uploaded.name or "")
                except Exception:
                    pass

    return {**meta, "error": last_err}


# ── Result I/O ───────────────────────────────────────────────────────

def load_existing() -> dict[int, dict]:
    if not RESULTS_FILE.exists():
        return {}
    with open(RESULTS_FILE, "r", encoding="utf-8") as f:
        data = json.load(f)
    return {r["idx"]: r for r in data.get("items", [])}


def save_results(items: list[dict], model_name: str):
    items_sorted = sorted(items, key=lambda x: x["idx"])

    errors = [r for r in items_sorted if "error" in r]
    missing = [r for r in items_sorted if "error" not in r and r.get("missing")]
    scored = [r for r in items_sorted if "error" not in r and not r.get("missing")]
    correct = sum(1 for r in scored if r.get("is_correct"))
    total_scored = len(scored)

    # per-task (only scored items)
    per_task: dict[str, dict] = {}
    for r in scored:
        t = r["task"]
        if t not in per_task:
            per_task[t] = {"total": 0, "correct": 0, "accuracy": 0.0}
        per_task[t]["total"] += 1
        if r["is_correct"]:
            per_task[t]["correct"] += 1
    for v in per_task.values():
        v["accuracy"] = round(v["correct"] / v["total"], 4) if v["total"] else 0.0

    # per-tag (only scored items)
    per_tag: dict[str, dict] = {}
    for r in scored:
        tags = r.get("tag", [])
        if len(tags) >= 2:
            tag = tags[1]
            if tag not in per_tag:
                per_tag[tag] = {"total": 0, "correct": 0, "accuracy": 0.0}
            per_tag[tag]["total"] += 1
            if r["is_correct"]:
                per_tag[tag]["correct"] += 1
    for v in per_tag.values():
        v["accuracy"] = round(v["correct"] / v["total"], 4) if v["total"] else 0.0

    output = {
        "model": model_name,
        "timestamp": datetime.now(timezone.utc).isoformat(),
        "summary": {
            "submitted": len(items_sorted),
            "scored": total_scored,
            "missing": len(missing),
            "errors": len(errors),
            "correct": correct,
            "accuracy": round(correct / total_scored, 4) if total_scored else 0,
        },
        "per_task": dict(per_task),
        "per_tag": dict(per_tag),
        "items": items_sorted,
    }

    with open(RESULTS_FILE, "w", encoding="utf-8") as f:
        json.dump(output, f, ensure_ascii=False, indent=2)


# ── Display ──────────────────────────────────────────────────────────

def print_summary(items: list[dict], model_name: str):
    try:
        from rich.console import Console  # noqa: F401
        _rich_summary(items, model_name)
    except ImportError:
        _ascii_summary(items, model_name)


def _collect_stats(items):
    """Collect all accuracy breakdowns from result items."""
    errors = [r for r in items if "error" in r]
    missing = [r for r in items if "error" not in r and r.get("missing")]
    evaluated = [r for r in items if "error" not in r and not r.get("missing")]
    correct = sum(1 for r in evaluated if r["is_correct"])
    total = len(evaluated)

    def _acc_map(key_fn):
        m: dict[str, list[int]] = {}
        for r in evaluated:
            k = key_fn(r)
            if k is None:
                continue
            if k not in m:
                m[k] = [0, 0]
            m[k][1] += 1
            if r["is_correct"]:
                m[k][0] += 1
        return m

    per_task = _acc_map(lambda r: r.get("task", ""))
    per_qtype = _acc_map(lambda r: r.get("question_type") or (r.get("tag", [None, None])[1] if len(r.get("tag", [])) >= 2 else None))
    per_persp = _acc_map(lambda r: r.get("perspective") or (r.get("tag", [None])[0] if len(r.get("tag", [])) >= 1 else None))

    # per scenario (grouped under task)
    per_scenario: dict[str, dict[str, list[int]]] = {}
    for r in evaluated:
        t = r.get("task", "")
        s = r.get("scenario", "")
        if t not in per_scenario:
            per_scenario[t] = {}
        if s not in per_scenario[t]:
            per_scenario[t][s] = [0, 0]
        per_scenario[t][s][1] += 1
        if r["is_correct"]:
            per_scenario[t][s][0] += 1

    return {
        "evaluated": evaluated, "errors": errors, "missing": missing,
        "correct": correct, "total": total,
        "per_task": per_task, "per_qtype": per_qtype,
        "per_persp": per_persp, "per_scenario": per_scenario,
    }


def _rich_summary(items, model_name):
    from rich.console import Console
    from rich.table import Table
    from rich import box

    console = Console()
    s = _collect_stats(items)
    total, correct = s["total"], s["correct"]
    acc = correct / total if total else 0
    bar_w = 20

    def _bar(c, t, color="green"):
        if t == 0:
            return ""
        ratio = c / t
        filled = int(ratio * bar_w)
        return f"[{color}]{'#' * filled}[/][bright_black]{'.' * (bar_w - filled)}[/]"

    # ── Header ──
    console.print()
    console.rule(f"[bold cyan]Evaluation Results - {model_name}[/]")
    console.print()

    hdr = Table(box=None, show_header=False, padding=(0, 2))
    hdr.add_column(style="white")
    hdr.add_column(style="bold")
    hdr.add_row("Model", f"[cyan]{model_name}[/]")
    hdr.add_row("Submitted / Scored / Missing / Errors",
                f"{len(items)} / {total} / {len(s['missing'])} / {len(s['errors'])}")
    hdr.add_row("Overall Accuracy", f"[bold green]{acc:.1%}[/]  ({correct}/{total})")
    console.print(hdr)
    console.print()

    # ── Per Task + Scenario (nested) ──
    tbl = Table(
        title="Accuracy by Task & Scenario",
        box=box.ROUNDED, border_style="bright_black", show_lines=False, padding=(0, 1),
    )
    tbl.add_column("Task", style="cyan bold", min_width=8)
    tbl.add_column("Scenario", style="green", min_width=26)
    tbl.add_column("Correct", justify="right", style="white", min_width=8)
    tbl.add_column("Total", justify="right", style="white", min_width=6)
    tbl.add_column("Accuracy", justify="right", style="yellow bold", min_width=8)
    tbl.add_column("", min_width=bar_w + 2)

    for task in sorted(s["per_scenario"]):
        scenarios = s["per_scenario"][task]
        first = True
        for sc in sorted(scenarios):
            c, t = scenarios[sc]
            a = c / t if t else 0
            tbl.add_row(
                task if first else "", sc, str(c), str(t),
                f"{a:.1%}", _bar(c, t),
            )
            first = False
        # task subtotal
        tc, tt = s["per_task"].get(task, [0, 0])
        ta = tc / tt if tt else 0
        tbl.add_row(
            f"[bold]{task}[/]", "[dim]-------- Subtotal --------[/]",
            f"[bold]{tc}[/]", f"[bold]{tt}[/]",
            f"[bold]{ta:.1%}[/]", _bar(tc, tt, "yellow"),
            style="on grey11",
        )

    # grand total
    tbl.add_row(
        "[bold white]TOTAL", "",
        f"[bold white]{correct}", f"[bold white]{total}",
        f"[bold white]{acc:.1%}", _bar(correct, total, "cyan"),
        style="on grey19",
    )
    console.print(tbl)

    # ── Per Question Type ──
    tbl2 = Table(
        title="Accuracy by Question Type",
        box=box.ROUNDED, border_style="bright_black", padding=(0, 1),
    )
    tbl2.add_column("Type", style="magenta bold", min_width=12)
    tbl2.add_column("Correct", justify="right", style="white", min_width=8)
    tbl2.add_column("Total", justify="right", style="white", min_width=6)
    tbl2.add_column("Accuracy", justify="right", style="yellow bold", min_width=8)
    tbl2.add_column("", min_width=bar_w + 2)

    for tag in sorted(s["per_qtype"]):
        c, t = s["per_qtype"][tag]
        tbl2.add_row(tag, str(c), str(t), f"{c / t:.1%}" if t else "-", _bar(c, t))

    console.print()
    console.print(tbl2)

    # ── Per Perspective ──
    tbl3 = Table(
        title="Accuracy by Perspective",
        box=box.ROUNDED, border_style="bright_black", padding=(0, 1),
    )
    tbl3.add_column("Perspective", style="cyan bold", min_width=12)
    tbl3.add_column("Correct", justify="right", style="white", min_width=8)
    tbl3.add_column("Total", justify="right", style="white", min_width=6)
    tbl3.add_column("Accuracy", justify="right", style="yellow bold", min_width=8)
    tbl3.add_column("", min_width=bar_w + 2)

    for p in sorted(s["per_persp"]):
        c, t = s["per_persp"][p]
        tbl3.add_row(p, str(c), str(t), f"{c / t:.1%}" if t else "-", _bar(c, t))

    console.print()
    console.print(tbl3)

    # ── Missing / Error list ──
    problems = s["missing"] + s["errors"]
    if problems:
        console.print()
        tbl4 = Table(
            title=f"Missing & Errors ({len(problems)} items)",
            box=box.ROUNDED, border_style="red", padding=(0, 1),
        )
        tbl4.add_column("#", style="white", min_width=4)
        tbl4.add_column("Task", style="cyan", min_width=6)
        tbl4.add_column("Scenario", style="green", min_width=20)
        tbl4.add_column("Clip", style="white", min_width=30)
        tbl4.add_column("Reason", style="red", min_width=30)

        for r in sorted(problems, key=lambda x: x.get("idx", 0)):
            if "error" in r:
                reason = f"[red]API Error: {r['error'][:60]}[/]"
            else:
                reason = f"[yellow]Bad format: {r.get('response', '')[:60]}[/]"
            tbl4.add_row(
                str(r.get("idx", "?")),
                r.get("task", ""),
                r.get("scenario", ""),
                r.get("clip", ""),
                reason,
            )
        console.print(tbl4)

    console.print()


def _ascii_summary(items, model_name):
    s = _collect_stats(items)
    total, correct = s["total"], s["correct"]
    acc = correct / total if total else 0

    print(f"\n{'=' * 60}")
    print(f"  Model: {model_name}")
    print(f"  Submitted: {len(items)}  Scored: {total}  "
          f"Missing: {len(s['missing'])}  Errors: {len(s['errors'])}")
    print(f"  Overall Accuracy: {acc:.1%}  ({correct}/{total})")
    print(f"{'=' * 60}")

    print("\n  Per Task:")
    for t in sorted(s["per_task"]):
        c, n = s["per_task"][t]
        print(f"    {t:<8}  {c:>3}/{n:<3}  {c/n:.1%}" if n else f"    {t:<8}  -")

    print("\n  Per Question Type:")
    for t in sorted(s["per_qtype"]):
        c, n = s["per_qtype"][t]
        print(f"    {t:<12}  {c:>3}/{n:<3}  {c/n:.1%}" if n else f"    {t:<12}  -")

    print("\n  Per Perspective:")
    for p in sorted(s["per_persp"]):
        c, n = s["per_persp"][p]
        print(f"    {p:<12}  {c:>3}/{n:<3}  {c/n:.1%}" if n else f"    {p:<12}  -")

    print()


# ── Main ─────────────────────────────────────────────────────────────

def main():
    parser = argparse.ArgumentParser(description="Evaluate VQA dataset with Gemini")
    parser.add_argument("--model", default=os.environ.get("GEMINI_MODEL", DEFAULT_MODEL),
                        help=f"Model name (default: {DEFAULT_MODEL})")
    parser.add_argument("--workers", type=int,
                        default=int(os.environ.get("EVAL_WORKERS", DEFAULT_WORKERS)),
                        help=f"Parallel workers (default: {DEFAULT_WORKERS})")
    parser.add_argument("--rpm", type=int,
                        default=int(os.environ.get("EVAL_RPM", DEFAULT_RPM)),
                        help=f"Max requests per minute (default: {DEFAULT_RPM})")
    parser.add_argument("--thinking-budget", type=int,
                        default=int(os.environ.get("EVAL_THINKING", DEFAULT_THINKING_BUDGET)),
                        help=f"Thinking token budget (0=off, default: {DEFAULT_THINKING_BUDGET})")
    parser.add_argument("--no-resume", action="store_true",
                        help="Ignore previous results, start fresh")
    args = parser.parse_args()

    # API key: .env -> environment variable
    from dotenv import load_dotenv
    load_dotenv(BASE_DIR / ".env")

    api_key = os.environ.get("GOOGLE_API_KEY", "")
    if not api_key:
        print("[ERROR] GOOGLE_API_KEY not found. Set it in .env or environment.")
        sys.exit(1)

    client = genai.Client(api_key=api_key)

    # Load dataset
    if not META_FULL.exists():
        print(f"[ERROR] {META_FULL} not found - run clip.py first")
        sys.exit(1)
    with open(META_FULL, "r", encoding="utf-8") as f:
        vqa_list = json.load(f)
    print(f"Loaded {len(vqa_list)} VQA items")

    # Resume
    existing = {} if args.no_resume else load_existing()
    if existing:
        print(f"Resuming: {len(existing)} already evaluated")

    pending = [(i, item) for i, item in enumerate(vqa_list) if i not in existing]
    if not pending:
        print("All items already evaluated.")
        all_items = list(existing.values())
        print_summary(all_items, args.model)
        return

    think_label = f"thinking={args.thinking_budget}" if args.thinking_budget else "thinking=off"
    print(f"Pending: {len(pending)}  |  Model: {args.model}  |  "
          f"Workers: {args.workers}  |  RPM: {args.rpm}  |  {think_label}")

    limiter = RateLimiter(args.rpm)

    results = list(existing.values())
    lock = threading.Lock()
    done = 0
    ok_count = sum(1 for r in results if "error" not in r and r.get("is_correct"))
    eval_count = sum(1 for r in results if "error" not in r)
    err_count = sum(1 for r in results if "error" in r)

    # ── Ctrl+C: save checkpoint and force exit ──
    stop_flag = threading.Event()

    def _on_sigint(sig, frame):
        if stop_flag.is_set():
            # Second Ctrl+C: force kill
            print("\n\nForce quit.")
            os._exit(1)
        stop_flag.set()
        print("\n\nCtrl+C received - finishing current items then stopping...")

    signal.signal(signal.SIGINT, _on_sigint)

    t0 = time.monotonic()
    executor = ThreadPoolExecutor(max_workers=args.workers)
    futures = {
        executor.submit(
            evaluate_one, client, args.model, item, idx, limiter,
            thinking_budget=args.thinking_budget,
        ): idx
        for idx, item in pending
    }

    try:
        for future in as_completed(futures):
            if stop_flag.is_set():
                # Cancel remaining futures
                for f in futures:
                    f.cancel()
                break

            result = future.result()
            with lock:
                results.append(result)
                done += 1
                if "error" not in result and not result.get("missing"):
                    eval_count += 1
                    if result.get("is_correct"):
                        ok_count += 1
                elif "error" in result:
                    err_count += 1

                # Progress
                elapsed = time.monotonic() - t0
                speed = done / elapsed if elapsed > 0 else 0
                eta = (len(pending) - done) / speed if speed > 0 else 0
                cur_acc = ok_count / eval_count if eval_count else 0
                status = "OK" if "error" not in result else f"ERR: {result.get('error', '')[:40]}"
                print(
                    f"\r[{done}/{len(pending)}] "
                    f"acc={cur_acc:.1%} err={err_count} "
                    f"eta={eta:.0f}s  | {result.get('clip', ''):40s} {status}",
                    end="", flush=True,
                )

                # Realtime checkpoint
                save_results(results, args.model)
    finally:
        executor.shutdown(wait=False, cancel_futures=True)

    print()
    save_results(results, args.model)
    print(f"\nResults saved to {RESULTS_FILE}  ({len(results)} items)")
    print_summary(results, args.model)


if __name__ == "__main__":
    main()