app.py

import os
import re
import json
import uuid
import random
from pathlib import Path
from datetime import datetime, timezone

import gradio as gr
from huggingface_hub import HfApi

# ----------------------------
# Config
# ----------------------------
STUDY_ROOT = Path("study")
CATEGORIES = ["search_t2v", "search_i2v", "opt_t2v", "opt_i2v"]

HF_TOKEN = os.getenv("HF_TOKEN")
RESULTS_REPO_ID = os.getenv("RESULTS_REPO_ID")

api = HfApi(token=HF_TOKEN) if HF_TOKEN else None

LOCAL_RESULTS_DIR = Path("local_results")
LOCAL_RESULTS_DIR.mkdir(exist_ok=True)

# ----------------------------
# Pair discovery
# ----------------------------
def normalize_stem(path: Path) -> str:
    """
    Match files like:
      0001_good.mp4
      0001_bad.mp4
    and return:
      0001
    """
    m = re.match(r"^(\d+)_(good|bad)$", path.stem)
    if m:
        return m.group(1)
    return path.stem


def build_pairs():
    all_pairs = []

    for category in CATEGORIES:
        good_dir = STUDY_ROOT / category / "good"
        bad_dir = STUDY_ROOT / category / "bad"

        good_files = (
            list(good_dir.glob("*.mp4"))
            + list(good_dir.glob("*.webm"))
            + list(good_dir.glob("*.ogg"))
        )
        bad_files = (
            list(bad_dir.glob("*.mp4"))
            + list(bad_dir.glob("*.webm"))
            + list(bad_dir.glob("*.ogg"))
        )

        good_map = {}
        for p in good_files:
            key = normalize_stem(p)
            good_map.setdefault(key, []).append(p)

        bad_map = {}
        for p in bad_files:
            key = normalize_stem(p)
            bad_map.setdefault(key, []).append(p)

        shared_keys = sorted(set(good_map.keys()) & set(bad_map.keys()))

        for key in shared_keys:
            goods = sorted(good_map[key])
            bads = sorted(bad_map[key])
            n = min(len(goods), len(bads))

            for i in range(n):
                all_pairs.append(
                    {
                        "category": category,
                        "pair_id": f"{category}::{key}::{i}",
                        "key": key,
                        "good_path": str(goods[i].resolve()),
                        "bad_path": str(bads[i].resolve()),
                        "good_file": goods[i].name,
                        "bad_file": bads[i].name,
                    }
                )

    return all_pairs


PAIRS = build_pairs()

if len(PAIRS) == 0:
    raise RuntimeError("No matched good/bad pairs found.")


# ----------------------------
# Video helper
# ----------------------------
def video_value(path_str: str):
    path = Path(path_str)
    if not path.exists():
        raise FileNotFoundError(f"Video file not found: {path}")
    return str(path.resolve())


# ----------------------------
# Sampling logic
# ----------------------------
def sample_pair(seen_pair_ids):
    seen_pair_ids = set(seen_pair_ids or [])
    unseen = [p for p in PAIRS if p["pair_id"] not in seen_pair_ids]
    pool = unseen if unseen else PAIRS

    pair = random.choice(pool)
    good_on_left = random.choice([True, False])

    if good_on_left:
        left_path = pair["good_path"]
        right_path = pair["bad_path"]
        left_label = "good"
        right_label = "bad"
    else:
        left_path = pair["bad_path"]
        right_path = pair["good_path"]
        left_label = "bad"
        right_label = "good"

    current = {
        **pair,
        "left_path": left_path,
        "right_path": right_path,
        "left_hidden_label": left_label,
        "right_hidden_label": right_label,
        "good_on_left": good_on_left,
    }
    return current


# ----------------------------
# Hub / persistence helpers
# ----------------------------
def local_save(record):
    local_name = LOCAL_RESULTS_DIR / f"{record['response_id']}.json"
    with open(local_name, "w", encoding="utf-8") as f:
        json.dump(record, f, ensure_ascii=False, indent=2)
    return str(local_name)


def check_hub_setup():
    """
    Returns a warning string if Hub saving is not configured correctly.
    Otherwise returns None.
    """
    if not RESULTS_REPO_ID:
        return "RESULTS_REPO_ID is not set. Responses will be saved locally only."

    if not HF_TOKEN:
        return "HF_TOKEN is not set. Responses will be saved locally only."

    try:
        api.repo_info(repo_id=RESULTS_REPO_ID, repo_type="dataset")
        return None
    except Exception as e:
        return f"Hub dataset check failed: {type(e).__name__}: {e}"


HUB_WARNING = check_hub_setup()


def save_response(record):
    """
    Always save locally first.
    Then try to upload to the dataset repo.
    Never raise an exception to the UI callback.
    """
    local_path = local_save(record)

    if not RESULTS_REPO_ID or not HF_TOKEN or api is None:
        return {
            "ok": True,
            "saved_local": True,
            "saved_hub": False,
            "message": f"Saved locally to {local_path}. Hub upload is not configured."
        }

    tmp_name = f"/tmp/{record['response_id']}.json"
    with open(tmp_name, "w", encoding="utf-8") as f:
        json.dump(record, f, ensure_ascii=False, indent=2)

    remote_path = f"responses/{record['timestamp'][:10]}/{record['response_id']}.json"

    try:
        api.upload_file(
            path_or_fileobj=tmp_name,
            path_in_repo=remote_path,
            repo_id=RESULTS_REPO_ID,
            repo_type="dataset",
        )
        return {
            "ok": True,
            "saved_local": True,
            "saved_hub": True,
            "message": f"Saved to dataset repo: {remote_path}"
        }
    except Exception as e:
        return {
            "ok": False,
            "saved_local": True,
            "saved_hub": False,
            "message": f"Saved locally to {local_path}, but Hub upload failed: {type(e).__name__}: {e}"
        }


# ----------------------------
# Gradio callbacks
# ----------------------------
def start_session():
    participant_id = str(uuid.uuid4())
    seen_pair_ids = []
    current = sample_pair(seen_pair_ids)

    status_msg = "Study loaded. Watch both videos and answer the questions below."
    if HUB_WARNING:
        status_msg += f"\n\nWarning: {HUB_WARNING}"

    return (
        video_value(current["left_path"]),
        video_value(current["right_path"]),
        f"Participant ID: {participant_id}",
        participant_id,
        seen_pair_ids,
        current,
        None,
        None,
        status_msg,
        gr.update(visible=False),  # hide Study study
        gr.update(visible=True),   # show Submit and continue
    )


def submit_and_next(
    plausible_answer,
    quality_answer,
    participant_id,
    seen_pair_ids,
    current
):
    if current is None:
        return (
            None, None, seen_pair_ids, current,
            "No current pair loaded."
        )

    if plausible_answer is None or quality_answer is None:
        return (
            video_value(current["left_path"]),
            video_value(current["right_path"]),
            seen_pair_ids,
            current,
            "Please answer both questions before continuing."
        )

    timestamp = datetime.now(timezone.utc).isoformat()

    record = {
        "response_id": str(uuid.uuid4()),
        "timestamp": timestamp,
        "participant_id": participant_id,
        "category": current["category"],
        "pair_id": current["pair_id"],
        "pair_key": current["key"],
        "left_video": current["left_path"],
        "right_video": current["right_path"],
        "good_video": current["good_path"],
        "bad_video": current["bad_path"],
        "good_on_left": current["good_on_left"],
        "left_hidden_label": current["left_hidden_label"],
        "right_hidden_label": current["right_hidden_label"],
        "physical_plausibility_answer": plausible_answer,
        "visual_quality_answer": quality_answer,
    }

    save_result = save_response(record)

    seen_pair_ids = list(set(seen_pair_ids + [current["pair_id"]]))
    next_pair = sample_pair(seen_pair_ids)

    status_msg = save_result["message"]
    if not save_result["saved_hub"]:
        status_msg += "\n\nYour response was still saved locally."

    return (
        video_value(next_pair["left_path"]),
        video_value(next_pair["right_path"]),
        seen_pair_ids,
        next_pair,
        status_msg
    )


# ----------------------------
# UI
# ----------------------------
with gr.Blocks(title="Human Study") as demo:
    gr.Markdown(
        """
        # Human Study

        Please answer as many video pairs as you can.

        You do not need to finish the full study. Each response is saved automatically when you click **Submit and continue**, so you may stop at any time.

        ### Physical plausibility
        - Watch both videos and choose the one that is relatively better in terms of physical realism.
        - Sometimes both videos may be imperfect, but please select the one with better physics than the other.
        - Pay attention to object motion, interactions, and overall realism.
        - Please avoid choosing **No preference** for physical plausibility as much as possible.

        ### Visual quality
        - Choose the video with better visual quality.
        - Focus on overall visual appearance and rendering quality.
        """
    )

    participant_label = gr.Markdown()

    with gr.Row():
        video_left = gr.Video(
            label="Left",
            interactive=False,
            format="mp4",
        )
        video_right = gr.Video(
            label="Right",
            interactive=False,
            format="mp4",
        )

    plausible = gr.Radio(
        choices=["Left", "Right", "No preference"],
        label="Which video is more physically plausible?"
    )

    quality = gr.Radio(
        choices=["Left", "Right", "No preference"],
        label="Which video has better visual quality?"
    )

    status = gr.Markdown()

    start_btn = gr.Button("Study study", visible=True)
    next_btn = gr.Button("Submit and continue", visible=False)

    participant_id_state = gr.State()
    seen_pair_ids_state = gr.State([])
    current_pair_state = gr.State()

    start_btn.click(
        fn=start_session,
        outputs=[
            video_left,
            video_right,
            participant_label,
            participant_id_state,
            seen_pair_ids_state,
            current_pair_state,
            plausible,
            quality,
            status,
            start_btn,
            next_btn,
        ]
    )

    next_btn.click(
        fn=submit_and_next,
        inputs=[
            plausible,
            quality,
            participant_id_state,
            seen_pair_ids_state,
            current_pair_state
        ],
        outputs=[
            video_left,
            video_right,
            seen_pair_ids_state,
            current_pair_state,
            status
        ]
    ).then(
        lambda: (None, None),
        outputs=[plausible, quality]
    )

demo.launch()