{
 "cells": [
  {
   "cell_type": "code",
   "id": "initial_id",
   "metadata": {
    "collapsed": true,
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:10.176744Z",
     "start_time": "2025-04-14T03:41:08.344955Z"
    }
   },
   "source": [
    "import torch\n",
    "import numpy as np\n",
    "import cv2\n",
    "from decord import VideoReader, cpu"
   ],
   "outputs": [],
   "execution_count": 1
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:10.726870Z",
     "start_time": "2025-04-14T03:41:10.181751Z"
    }
   },
   "cell_type": "code",
   "source": [
    "print(torch.__version__)\n",
    "print(torch.cuda.is_available())"
   ],
   "id": "31d50e0f9e4ea204",
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2.6.0+cu124\n",
      "True\n"
     ]
    }
   ],
   "execution_count": 2
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:10.875039Z",
     "start_time": "2025-04-14T03:41:10.871492Z"
    }
   },
   "cell_type": "code",
   "source": [
    "def load_first_480_frames(video_path, resize=(224, 224)):\n",
    "    vr = VideoReader(video_path, ctx=cpu(0))\n",
    "    total_frames = len(vr)\n",
    "\n",
    "    if total_frames < 480:\n",
    "        raise ValueError(f\"{video_path},视频帧数不足 480 帧\")\n",
    "\n",
    "    indices = list(range(480))\n",
    "    frames = vr.get_batch(indices).asnumpy()\n",
    "\n",
    "    if resize:\n",
    "        frames = np.array([cv2.resize(f, resize) for f in frames])\n",
    "    return frames"
   ],
   "id": "e351cfd29d48331a",
   "outputs": [],
   "execution_count": 3
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:10.884466Z",
     "start_time": "2025-04-14T03:41:10.880516Z"
    }
   },
   "cell_type": "code",
   "source": [
    "def preprocess_for_slowfast(frames, num_frames=32, alpha=4):\n",
    "    total = len(frames)\n",
    "    indices = np.linspace(0, total - 1, num=num_frames, dtype=int)\n",
    "    frames = frames[indices]\n",
    "\n",
    "    frames = frames / 255.0\n",
    "    frames = (frames - [0.45, 0.45, 0.45]) / [0.225, 0.225, 0.225]\n",
    "    frames = frames.astype(np.float32)\n",
    "\n",
    "    frames = torch.from_numpy(frames).permute(3, 0, 1, 2).unsqueeze(0)\n",
    "\n",
    "    fast_pathway = frames\n",
    "    slow_pathway = frames[:, :, ::alpha, :, :]\n",
    "\n",
    "    return [slow_pathway, fast_pathway]"
   ],
   "id": "f8784f81a946176",
   "outputs": [],
   "execution_count": 4
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:10.893723Z",
     "start_time": "2025-04-14T03:41:10.890080Z"
    }
   },
   "cell_type": "code",
   "source": [
    "# 定义特征提取模型 (SlowFast Backbone)\n",
    "class SlowFastFeatureExtractor(torch.nn.Module):\n",
    "    def __init__(self):\n",
    "        super().__init__()\n",
    "        model = torch.hub.load(\"facebookresearch/pytorchvideo\", \"slowfast_r50\", pretrained=True)\n",
    "        # 移除分类头，仅保留backbone部分\n",
    "        self.blocks = model.blocks[:-1]  # 去掉最后的分类器 head\n",
    "        self.pool = torch.nn.AdaptiveAvgPool3d(1)  # 全局池化\n",
    "\n",
    "    def forward(self, x):\n",
    "        for block in self.blocks:\n",
    "            x = block(x)\n",
    "        x = self.pool(x)\n",
    "        x = torch.flatten(x, 1)  # [B, C]\n",
    "        x = torch.nn.functional.normalize(x, dim=1)  # 特征归一化\n",
    "        return x"
   ],
   "id": "6c52b60b3399c5b7",
   "outputs": [],
   "execution_count": 5
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:13.527271Z",
     "start_time": "2025-04-14T03:41:10.899167Z"
    }
   },
   "cell_type": "code",
   "source": [
    "\n",
    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
    "print(\"Loading SlowFast backbone for feature extraction...\")\n",
    "model = SlowFastFeatureExtractor().to(device).eval()\n",
    "\n",
    "\n",
    "def extract_video_feature(video_path):\n",
    "\n",
    "\n",
    "    frames = load_first_480_frames(video_path)\n",
    "    inputs = preprocess_for_slowfast(frames, num_frames=32, alpha=4)\n",
    "\n",
    "    with torch.no_grad():\n",
    "        inputs = [x.to(device) for x in inputs]\n",
    "        features = model(inputs)\n",
    "\n",
    "    features = features.cpu().numpy().squeeze()\n",
    "    return features"
   ],
   "id": "d841190cdd5ee920",
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Loading SlowFast backbone for feature extraction...\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using cache found in C:\\Users\\zikai/.cache\\torch\\hub\\facebookresearch_pytorchvideo_main\n"
     ]
    }
   ],
   "execution_count": 6
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:13.545015Z",
     "start_time": "2025-04-14T03:41:13.541939Z"
    }
   },
   "cell_type": "code",
   "source": [
    "def getFeature(video_path = r\"D:\\DESKTOP\\2025\\44\\a1\\dataset\\org\\1.mp4\"):\n",
    "\n",
    "    feature = extract_video_feature(video_path)\n",
    "    # print(\"Video feature shape (for copyright):\", feature.shape)\n",
    "    return feature"
   ],
   "id": "dce706080dfba5b6",
   "outputs": [],
   "execution_count": 7
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:41:13.607891Z",
     "start_time": "2025-04-14T03:41:13.554129Z"
    }
   },
   "cell_type": "code",
   "source": [
    "import sqlite3\n",
    "import io\n",
    "import os\n",
    "import tqdm\n",
    "\n",
    "# 注册适配器与转换器：numpy数组 <-> BLOB\n",
    "def adapt_array(arr):\n",
    "    out = io.BytesIO()\n",
    "    np.save(out, arr)\n",
    "    out.seek(0)\n",
    "    return sqlite3.Binary(out.read())\n",
    "\n",
    "def convert_array(text):\n",
    "    out = io.BytesIO(text)\n",
    "    out.seek(0)\n",
    "    return np.load(out)\n",
    "\n",
    "# 注册自定义类型处理\n",
    "sqlite3.register_adapter(np.ndarray, adapt_array)\n",
    "sqlite3.register_converter(\"array\", convert_array)\n",
    "\n",
    "# 创建数据库连接（带类型检测）\n",
    "conn = sqlite3.connect(\"slowfast.db\", detect_types=sqlite3.PARSE_DECLTYPES)\n",
    "cursor = conn.cursor()\n",
    "\n",
    "# 创建表\n",
    "cursor.execute(\"\"\"\n",
    "CREATE TABLE IF NOT EXISTS data (\n",
    "    id INTEGER PRIMARY KEY,\n",
    "    array BLOB,\n",
    "    group_path TEXT,\n",
    "    full_path TEXT\n",
    ")\n",
    "\"\"\")\n",
    "\n",
    "\n",
    "def add_to_db(array_to_store,group_path,full_path):\n",
    "    cursor.execute(\"INSERT INTO data (array,group_path,full_path) VALUES (?,?,?)\", (array_to_store,group_path,full_path,))\n",
    "    conn.commit()\n",
    "\n",
    "# # 读取数组\n",
    "# cursor.execute(\"SELECT array FROM data WHERE id=1\")\n",
    "# fetched_array = cursor.fetchone()[0]\n",
    "#\n",
    "# print(\"原始数组:\\n\", array_to_store)\n",
    "# print(\"读取的数组:\\n\", fetched_array)\n",
    "\n",
    "folder_path = r\"D:\\DESKTOP\\2025\\44\\a1\\dataset\"\n",
    "\n",
    "all_files = []\n",
    "names = [str(x)+\".mp4\" for x in range(10)]\n",
    "print(names)\n",
    "\n",
    "\n",
    "# 遍历文件夹\n",
    "for group_path in os.listdir(folder_path):\n",
    "    full_path = os.path.join(folder_path, group_path)\n",
    "    if os.path.isdir(full_path):\n",
    "        # 遍历子文件夹\n",
    "        for video_path in os.listdir(full_path):\n",
    "            if os.path.basename(video_path) in names:\n",
    "                video_full_path = os.path.join(full_path, video_path)\n",
    "                if os.path.isfile(video_full_path):\n",
    "                    # 处理视频文件\n",
    "                    all_files.append((group_path,video_full_path))\n",
    "print(len(all_files))\n"
   ],
   "id": "dcf5af026d672b41",
   "outputs": [],
   "execution_count": 8
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:42:00.157934Z",
     "start_time": "2025-04-14T03:41:13.633220Z"
    }
   },
   "cell_type": "code",
   "source": [
    "for group_path, video_full_path in tqdm.tqdm(all_files):\n",
    "    # 读取视频特征\n",
    "    feature = getFeature(video_full_path)\n",
    "    # 将特征存储到数据库\n",
    "    add_to_db(feature,group_path,video_full_path)\n",
    "    # print(f\"已处理并存储: {video_full_path}\")\n",
    "\n",
    "conn.close()\n"
   ],
   "id": "63eed21338e8f26c",
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 70/70 [00:46<00:00,  1.50it/s]\n"
     ]
    }
   ],
   "execution_count": 10
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:42:00.170765Z",
     "start_time": "2025-04-14T03:42:00.167606Z"
    }
   },
   "cell_type": "code",
   "source": "",
   "id": "d8a84ff72bc12b33",
   "outputs": [],
   "execution_count": 11
  },
  {
   "metadata": {
    "ExecuteTime": {
     "end_time": "2025-04-14T03:42:00.184252Z",
     "start_time": "2025-04-14T03:42:00.181880Z"
    }
   },
   "cell_type": "code",
   "source": "",
   "id": "7343d8b3fcea327c",
   "outputs": [],
   "execution_count": null
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}