src/streamlit_app.py

import streamlit as st
import json
import os
import random
import yaml
import uuid
from datetime import datetime
from filelock import FileLock
from collections import defaultdict
from huggingface_hub import HfApi, login
DATASET_REPO_ID = "Alexhe101/video_ranking_results"  # 你的 Dataset 仓库名
HF_TOKEN = os.environ.get("HF_TOKEN")                # 从 Secret 读取 Token
if HF_TOKEN:
    try:
        login(token=HF_TOKEN)
        api = HfApi()
    except Exception as e:
        st.warning(f"HF Login Failed: {e}")
from huggingface_hub import snapshot_download
DATA_ROOT = "./web_data_new"
JSON_PATH = os.path.join(DATA_ROOT, "dataset.json")
LOG_FILE = "final_eval_log.txt"
LOCK_FILE = "final_eval_log.txt.lock"

# 采样配置
BATCH_SIZE_PER_SCENE = 2
MAX_SCENES = 5
DATA_ROOT = "./web_data_new" # 本地存储路径（保持不变）

DATA_SOURCE_REPO = "Alexhe101/video_eval_data" # 刚才创建的 Dataset 名字
if not os.path.exists(DATA_ROOT):
    st.info(f"正在从 Dataset ({DATA_SOURCE_REPO}) 下载评测视频，请稍候...")
    try:
        snapshot_download(
            repo_id=DATA_SOURCE_REPO,
            repo_type="dataset",
            local_dir=DATA_ROOT,
            token=os.environ.get("HF_TOKEN") # 如果Dataset是Private的，需要Token
        )
        st.success("数据下载完成！")
        st.rerun() # 刷新页面以加载数据
    except Exception as e:
        st.error(f"数据下载失败: {e}")
        st.stop()

# ================= 配置区域 =================
st.set_page_config(layout="wide", page_title="Video Eval Platform")


# 检查目录是否存在，不存在则下载
# --- 评分标准 ---
PHYSICAL_RUBRIC = """
### ⚛️ 物理评分标准 (Physical Score)
- **5 (Perfect)**: 物理交互完美，重力、碰撞、接触点真实。
- **4 (Good)**: 物理规律基本正确，轻微瑕疵不影响理解。
- **3 (Fair)**: 有明显漂浮或穿模，但动作逻辑连贯。
- **2 (Poor)**: 严重物理错误（物体瞬移、穿透）。
- **1 (Fail)**: 完全崩坏，不符合物理规律。
"""

TASK_RUBRIC = """
### ✅ 子目标判定标准 (Subgoal Criteria)
勾选某个子目标 (Subgoal) 需同时满足：
1. **动作执行**: 视频中明确展示了该步骤。
2. **物理达标**: 该动作片段的物理质量 **≥ 4 (Good)**。
   *(如果动作发生了但穿模严重，请勿勾选)*
"""

# ================= 工具函数 =================

@st.cache_data
def load_full_data():
    if not os.path.exists(JSON_PATH):
        return []
    with open(JSON_PATH, 'r') as f:
        return json.load(f)

def get_session_user():
    if 'user_id' not in st.session_state:
        st.session_state['user_id'] = f"u_{str(uuid.uuid4())[:8]}"
    return st.session_state['user_id']

def parse_yaml_content(yaml_str):
    try:
        clean_str = yaml_str.replace("```yaml", "").replace("```", "").strip()
        data = yaml.safe_load(clean_str)
        # 兼容不同拼写 (intention vs intension)
        intent = data.get('intention') or data.get('intension') or 'Unknown'
        return intent, data.get('subgoals', [])
    except:
        return "Unknown", []

def save_log(record):
    lock = FileLock(LOCK_FILE)
    try:
        # 1. 先保存到本地 (原逻辑)
        with lock.acquire(timeout=5):
            with open(LOG_FILE, "a", encoding='utf-8') as f:
                f.write(json.dumps(record, ensure_ascii=False) + "\n")
        
        # 2. 新增：同步上传到 Hugging Face (静默上传，不打扰用户)
        if HF_TOKEN:
            api.upload_file(
                path_or_fileobj=LOG_FILE,
                path_in_repo="final_eval_log.txt",  # 在 Dataset 里的文件名
                repo_id=DATASET_REPO_ID,
                repo_type="dataset",
                commit_message=f"Sync data: {record.get('case_id', 'unknown')}"
            )
            print("Cloud sync success.") 
            
    except Exception as e:
        st.error(f"Save/Sync failed: {e}")
def get_my_batch(all_data):
    if 'my_batch' not in st.session_state:
        # 分层采样逻辑
        scene_map = defaultdict(list)
        for item in all_data:
            parts = item['case_id'].split('_')
            scene_name = parts[0] if len(parts) > 1 else "misc"
            scene_map[scene_name].append(item)
        
        available = list(scene_map.keys())
        random.shuffle(available)
        
        selected = []
        for s in available[:MAX_SCENES]:
            items = scene_map[s]
            cnt = min(len(items), BATCH_SIZE_PER_SCENE)
            selected.extend(random.sample(items, cnt))
            
        st.session_state['my_batch'] = selected
        st.session_state['current_index'] = 0
    return st.session_state['my_batch']

# ================= 主界面逻辑 =================

user_id = get_session_user()
full_data = load_full_data()
my_batch = get_my_batch(full_data)
curr_idx = st.session_state.get('current_index', 0)

# --- 侧边栏: 进度 & Rubric ---
with st.sidebar:
    st.title("📹 视频评估系统")
    st.write(f"User: `{user_id}`")
    
    total = len(my_batch)
    st.progress(curr_idx / total if total > 0 else 0)
    st.write(f"当前进度: {curr_idx} / {total}")
    
    st.divider()
    st.markdown(PHYSICAL_RUBRIC) 
    st.divider()
    st.markdown(TASK_RUBRIC)

# --- 完成判断 ---
if curr_idx >= len(my_batch):
    st.balloons()
    st.success("🎉 所有任务已完成！")
    if st.button("开始新的一组 (New Batch)"):
        del st.session_state['my_batch']
        del st.session_state['current_index']
        st.rerun()
    st.stop()

# --- 当前任务 ---
current_case = my_batch[curr_idx]
c_id = current_case['case_id']
videos = current_case['videos']
yaml_text = current_case['yaml_text']
intention, subgoals = parse_yaml_content(yaml_text)

# 随机化顺序 (Blind Test)
if "curr_case_id_final" not in st.session_state or st.session_state["curr_case_id_final"] != c_id:
    st.session_state["curr_case_id_final"] = c_id
    methods = list(videos.keys())
    random.shuffle(methods)
    st.session_state["curr_methods_order"] = methods

methods_order = st.session_state["curr_methods_order"]
labels = ["A", "B", "C", "D"]

# --- 页面顶部信息 ---
st.subheader(f"📌 Case: {c_id}")
st.markdown(f"**🎯 Goal (Intention):** `{intention}`")

# --- 视频展示与打分 ---
col1, col2 = st.columns(2, gap="large")

# 辅助函数：渲染单个视频块
def render_video_block(col, idx):
    method_name = methods_order[idx]
    label = labels[idx]
    video_path = os.path.join(DATA_ROOT, videos[method_name])
    
    with col:
        st.markdown(f"#### 📺 Video {label}")
        if os.path.exists(video_path):
            st.video(video_path, autoplay=True, loop=True, muted=True)
        else:
            st.warning("Video missing")

        # 1. 物理评分 (1-5)
        st.caption("1. Physical Score (1-5)")
        st.radio(
            f"phy_score_{label}", 
            [1, 2, 3, 4, 5], 
            index=None, 
            horizontal=True,
            key=f"score_{c_id}_{method_name}", 
            label_visibility="collapsed"
        )
        
        # 2. Subgoals (直接展示列表)
        st.caption("2. Subgoals & Completion")
        if subgoals:
            st.markdown(
                "<small style='color: #FF4B4B;'>"
                "⚠️ 若动作伴随严重缺陷（如严重穿模、物体幻觉等），请勿勾选。”"
                "</small>", 
                unsafe_allow_html=True
            )
            # 使用 expander 稍微收纳一下，防止占用太多空间，默认展开
            with st.expander("Subgoals Checklist", expanded=True):
                for i, sg in enumerate(subgoals):
                    st.checkbox(sg, key=f"sub_{c_id}_{method_name}_{i}")
        else:
            st.caption("No subgoals defined in YAML.")

# 渲染上半部分 (A, B)
render_video_block(col1, 0)
render_video_block(col2, 1)

st.divider()

# 渲染下半部分 (C, D)
col3, col4 = st.columns(2, gap="large")
render_video_block(col3, 2)
render_video_block(col4, 3)

st.divider()

# --- 4. 整体对比 (Best & Worst) ---
st.markdown("### 🏆 Overall Comparison")
st.markdown("请基于整体质量（物理 + 意图完成度）选出最好和最差的视频。")

bw_col1, bw_col2 = st.columns(2)
with bw_col1:
    best_choice = st.radio("🌟 Best Video", labels, horizontal=True, key=f"best_{c_id}")
with bw_col2:
    worst_choice = st.radio("💩 Worst Video", labels, horizontal=True, key=f"worst_{c_id}")

st.write("")
st.divider()

# --- 5. 异常上报 ---
is_case_error = st.checkbox("🚫 无法标注 (Case Error): 首帧设置不合理或任务无法完成", key=f"error_{c_id}")

st.write("")

# --- 提交按钮 ---
if st.button("🚀 提交 (Submit & Next)", type="primary", use_container_width=True):
    
    # 优先处理异常上报
    if is_case_error:
        final_record = {
            "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
            "user": user_id,
            "case_id": c_id,
            "is_error": True,
            "error_reason": "User reported impossible setting",
            "bws": {
                "order": methods_order
            }
        }
        save_log(final_record)
        st.warning("已标记为异常 Case，正在切换下一个...")
        st.session_state['current_index'] += 1
        st.rerun()

    else:
        # 1. 验证数据完整性 (正常流程)
        errors = []
        
        if not best_choice or not worst_choice:
            errors.append("请选择 Best 和 Worst 视频！")
        elif best_choice == worst_choice:
            errors.append("Best 和 Worst 不能是同一个视频！")
            
        # 验证每个视频的评分
        results = {}
        for m in methods_order:
            score = st.session_state.get(f"score_{c_id}_{m}")
            # 移除 is_succ 的获取
            
            # 收集选中的 Subgoals
            completed_subs = []
            if subgoals:
                for i, sg in enumerate(subgoals):
                    if st.session_state.get(f"sub_{c_id}_{m}_{i}", False):
                        completed_subs.append(sg)
            
            if score is None:
                errors.append(f"请为 {m} (Video {labels[methods_order.index(m)]}) 打物理分！")
                
            results[m] = {
                "physical_score": score,
                # "success": is_succ, # 已移除
                "completed_subgoals": completed_subs, # 完成的具体子目标
                "subgoal_rate": len(completed_subs)/len(subgoals) if len(subgoals)>0 else 0.0
            }

        if errors:
            for e in errors:
                st.error(e)
        else:
            # 2. 构造数据
            real_best = methods_order[labels.index(best_choice)]
            real_worst = methods_order[labels.index(worst_choice)]
            
            final_record = {
                "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
                "user": user_id,
                "case_id": c_id,
                "is_error": False, # 正常Case
                "details": results, 
                "bws": {
                    "best": real_best,
                    "worst": real_worst,
                    "order": methods_order
                }
            }
            
            # 3. 保存
            save_log(final_record)
            st.success("保存成功！")
            
            # 4. 切换下一个
            st.session_state['current_index'] += 1
            st.rerun()