{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import av\n",
    "import torch\n",
    "import numpy as np\n",
    "from transformers import SiglipImageProcessor\n",
    "from torch.utils.data import Dataset, DataLoader\n",
    "import pandas as pd\n",
    "import os\n",
    "from transformers import SiglipModel\n",
    "from tqdm import tqdm\n",
    "\n",
    "import json\n",
    "\n",
    "DEVICE= torch.device(\"cuda:1\" if torch.cuda.is_available() else \"cpu\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def read_video_pyav(container, indices):\n",
    "    '''\n",
    "    Decode the video with PyAV decoder.\n",
    "    Args:\n",
    "        container (`av.container.input.InputContainer`): PyAV container.\n",
    "        indices (`List[int]`): List of frame indices to decode.\n",
    "    Returns:\n",
    "        result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3).\n",
    "    '''\n",
    "    frames = []\n",
    "    container.seek(0)\n",
    "    start_index = indices[0]\n",
    "    end_index = indices[-1]\n",
    "    for i, frame in enumerate(container.decode(video=0)):\n",
    "        if i > end_index:\n",
    "            break\n",
    "        if i >= start_index and i in indices:\n",
    "            frames.append(frame)\n",
    "    return np.stack([x.to_ndarray(format=\"rgb24\") for x in frames])\n",
    "\n",
    "\n",
    "def sample_frame_indices(clip_len, frame_sample_rate, seg_len, num_segment, chunk_nb):\n",
    "    '''\n",
    "    Sample a given number of frame indices from the video.\n",
    "    Args:\n",
    "        clip_len (`int`): Total number of frames to sample.\n",
    "        frame_sample_rate (`int`): Sample every n-th frame.\n",
    "        seg_len (`int`): Maximum allowed index of sample's last frame.\n",
    "    Returns:\n",
    "        indices (`List[int]`): List of sampled frame indices\n",
    "    '''\n",
    "    \n",
    "    end_idx = seg_len - 1\n",
    "    start_idx = 0\n",
    "    indices = np.linspace(start_idx, end_idx, num=num_segment * clip_len)[chunk_nb*clip_len:(chunk_nb+1)*clip_len]\n",
    "    indices = np.clip(indices, 0, seg_len - 1).astype(np.int64)\n",
    "    return indices"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "class VideoClsDataset(Dataset):\n",
    "    def __init__(self, anno_path, prefix, processor, num_segment) -> None:\n",
    "        super().__init__()\n",
    "        cleaned = pd.read_csv(anno_path, header=None, delimiter=' ')\n",
    "        self.samples = list(cleaned.values[:, 0])\n",
    "        self.labels = list(cleaned.values[:, 1])\n",
    "        self.num_segment = num_segment\n",
    "        \n",
    "        self.label_array = list()\n",
    "        self.dataset_samples = list()\n",
    "        for i in range(self.num_segment):\n",
    "            for sample, label in zip(self.samples, self.labels):\n",
    "                self.label_array.append((i, label))\n",
    "                self.dataset_samples.append(sample)\n",
    "            \n",
    "        self.prefix = prefix\n",
    "        self.processor = processor\n",
    "    def __len__(self):\n",
    "        return len(self.label_array)\n",
    "    def __getitem__(self, index):\n",
    "        video_file = os.path.join(self.prefix, self.dataset_samples[index])\n",
    "        chunk_nb, label = self.label_array[index]\n",
    "        container = av.open(video_file)\n",
    "        indices = sample_frame_indices(clip_len=16, frame_sample_rate=1, seg_len=container.streams.video[0].frames, num_segment = self.num_segment, chunk_nb=chunk_nb)\n",
    "        video = read_video_pyav(container, indices)\n",
    "        inputs = self.processor.preprocess(list(video), return_tensors='pt')\n",
    "        inputs.update(labels=label)\n",
    "        return inputs"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = SiglipModel.from_pretrained(\"google/siglip-so400m-patch14-384\") #google/siglip-so400m-patch14-384\n",
    "model.to(DEVICE)\n",
    "preprocess = SiglipImageProcessor.from_pretrained(\"google/siglip-so400m-patch14-384\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "@torch.no_grad()\n",
    "def extract_feature(data_loader, model):\n",
    "    model.eval()\n",
    "    final_result = dict()\n",
    "    for batch in tqdm(data_loader):\n",
    "        for k in batch.keys():\n",
    "            batch[k] = batch[k].to(DEVICE, non_blocking=True)\n",
    "        B, T, C, H, W = batch['pixel_values'].shape\n",
    "        batch['pixel_values'] = batch['pixel_values'].flatten(0,1)\n",
    "        target = batch.pop('labels')\n",
    "        output =  model.get_image_features(**batch)\n",
    "        output =  output.reshape(B, T, output.shape[-1])\n",
    "        for id, feat in zip(target, output):\n",
    "            id = id.item()\n",
    "            feat = feat.cpu().numpy()\n",
    "            if final_result.get(id, None) is None:\n",
    "                final_result[id] = []\n",
    "            final_result[id].append(feat)\n",
    "        break\n",
    "    for k, v in final_result.items():\n",
    "        final_result[k] = np.concatenate(final_result[k], axis=0)\n",
    "    #np.save(file + \"_embedding\", np.array(list(final_result.values())))\n",
    "    #np.save(file + \"_labels\", np.array(list(final_result.keys())))\n",
    "    return np.array(list(final_result.values())), np.array(list(final_result.keys()))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# FOR DVSC23"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "DATASET_NAME = \"DVSC23\"\n",
    "DATA_PATH_QUERY=f'annotations/{DATASET_NAME}/queries/test.csv'\n",
    "DATA_PATH_DATABASE=f'annotations/{DATASET_NAME}/database/test.csv'\n",
    "\n",
    "QUERY_PREFIX=f'/path/to/prefix_path_of_query'\n",
    "DATABASE_PREFIX=f'/path/to/prefix_path_of_database'\n",
    "\n",
    "QUERY_MAP_PATH = f\"annotations/{DATASET_NAME}/queries/class_map.json\"\n",
    "DATABASE_MAP_PATH = f\"annotations/{DATASET_NAME}/database/class_map.json\"\n",
    "\n",
    "OUTPUT_DIR_QUERY=f'{DATASET_NAME}/queries'\n",
    "OUTPUT_DIR_DATABASE=f'{DATASET_NAME}/database'\n",
    "os.makedirs(OUTPUT_DIR_QUERY, exist_ok=True)\n",
    "os.makedirs(OUTPUT_DIR_DATABASE, exist_ok=True)\n",
    "\n",
    "NUM_SEGMENTS = 5\n",
    "BATCH_SIZE = 4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def normalize(embeddings):\n",
    "    #embeddings = (embeddings - embeddings.mean(0, keepdims=True)) / (embeddings.std(0, keepdims=True) + 1e-6)\n",
    "    embeddings /= np.linalg.norm(embeddings, axis=-1, keepdims=True) + 1e-6\n",
    "    embeddings = (embeddings - embeddings.mean(0, keepdims=True)) / (embeddings.std(0, keepdims=True) + 1e-6)\n",
    "    return embeddings\n",
    "\n",
    "def save_features(embeddings, ids, save_path, video_id_map):\n",
    "    \n",
    "    def flatten(data, ids):\n",
    "        T = data.shape[1]\n",
    "        return data.reshape(-1, data.shape[-1]), np.tile(ids[:, None],(1, T)).flatten()\n",
    "    embeddings, ids = flatten(embeddings, ids)\n",
    "\n",
    "\n",
    "    video_ids = [video_id_map[str(id)] for id in ids]  # Can also be str: \"Q20000\", ...\n",
    "    sort_ids = np.argsort(video_ids).tolist()\n",
    "    np.savez(\n",
    "        save_path,\n",
    "        video_ids=np.sort(video_ids).tolist(),\n",
    "        timestamps=[[0, 1]]*len(video_ids),\n",
    "        features=embeddings[sort_ids].astype(np.float32),\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#extract feature for query in DVSC23\n",
    "query_dataset = VideoClsDataset(DATA_PATH_QUERY, QUERY_PREFIX, preprocess, NUM_SEGMENTS)\n",
    "dataloader = DataLoader(query_dataset, num_workers=4, batch_size=BATCH_SIZE)\n",
    "query_embeddings, query_ids = extract_feature(dataloader, model)\n",
    "query_embeddings = normalize(query_embeddings)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open(QUERY_MAP_PATH, 'r') as file:\n",
    "    query_map = json.load(file)\n",
    "save_features(query_embeddings, query_ids, f'{OUTPUT_DIR_QUERY}/descriptors.npz', query_map)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#extract feature for reference in DVSC23\n",
    "ref_dataset = VideoClsDataset(DATA_PATH_DATABASE, DATABASE_PREFIX, preprocess, NUM_SEGMENTS)\n",
    "dataloader = DataLoader(ref_dataset, num_workers=4, batch_size=BATCH_SIZE)\n",
    "ref_embeddings, ref_ids = extract_feature(dataloader, model)\n",
    "ref_embeddings = normalize(ref_embeddings)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open(DATABASE_MAP_PATH, 'r') as file:\n",
    "    database_map = json.load(file)\n",
    "save_features(ref_embeddings, ref_ids, f'{OUTPUT_DIR_DATABASE}/descriptors.npz', database_map)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%cd DVSC23\n",
    "!python descriptor_eval.py  \\\n",
    "    --query_features '{OUTPUT_DIR_QUERY}/descriptors.npz' \\\n",
    "    --ref_features '{OUTPUT_DIR_DATABASE}/descriptors.npz' \\\n",
    "    --ground_truth ../annotations/DVSC23/test_ground_truth.csv\n",
    "%cd .."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# FOR FIVR-5K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "DATA_PATH_QUERY='annotations/FIVR-5K/queries/test.csv'\n",
    "DATA_PATH_DATABASE='annotations/FIVR-5K/database/test.csv'\n",
    "\n",
    "QUERY_PREFIX=f'/path/to/prefix_path_of_query'\n",
    "DATABASE_PREFIX=f'/path/to/prefix_path_of_database'\n",
    "\n",
    "\n",
    "QUERY_MAP_PATH = \"annotations/FIVR-5K/queries/class_map.json\"\n",
    "DATABASE_MAP_PATH = \"annotations/FIVR-5K/database/class_map.json\"\n",
    "\n",
    "\n",
    "OUTPUT_DIR='FIVR-5K'\n",
    "os.makedirs(OUTPUT_DIR, exist_ok=True)\n",
    "\n",
    "NUM_SEGMENTS = 1\n",
    "BATCH_SIZE = 4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def calculate_similarity(test_feature, train_feature):\n",
    "    test_feature /= np.linalg.norm(test_feature, axis=-1, keepdims=True) + 1e-6\n",
    "    train_feature /= np.linalg.norm(train_feature, axis=-1, keepdims=True) + 1e-6\n",
    "    return test_feature @ (train_feature.T)\n",
    "\n",
    "\n",
    "def save_predictions(logits, save_path, query_map, database_map):\n",
    "    prediction = dict()\n",
    "    for query_id, query_prob in enumerate(logits):\n",
    "        prob_per_query = dict()\n",
    "        for ref_id, prob in enumerate(query_prob):\n",
    "            if prob.item() > 0 :\n",
    "                prob_per_query[str(database_map[str(ref_id)]).split('.')[0]] = prob.item()\n",
    "        prediction[query_map[str(query_id)].split('.')[0]] = prob_per_query\n",
    "    with open(save_path, 'w') as f:\n",
    "        json.dump(prediction, f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#extract feature for query in FIVR-5K\n",
    "query_dataset = VideoClsDataset(DATA_PATH_QUERY, QUERY_PREFIX, preprocess, NUM_SEGMENTS)\n",
    "dataloader = DataLoader(query_dataset, num_workers=4, batch_size=BATCH_SIZE)\n",
    "query_embeddings, query_ids = extract_feature(dataloader, model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#extract feature for reference in FIVR-5K\n",
    "ref_dataset = VideoClsDataset(DATA_PATH_DATABASE, DATABASE_PREFIX, preprocess, NUM_SEGMENTS)\n",
    "dataloader = DataLoader(ref_dataset, num_workers=4, batch_size=BATCH_SIZE)\n",
    "ref_embeddings, ref_ids = extract_feature(dataloader, model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open(QUERY_MAP_PATH, 'r') as file:\n",
    "    query_map = json.load(file)\n",
    "\n",
    "with open(DATABASE_MAP_PATH, 'r') as file:\n",
    "    database_map = json.load(file)\n",
    "\n",
    "similarity = calculate_similarity(query_embeddings.mean(1), ref_embeddings.mean(1))\n",
    "save_predictions(similarity, f\"{OUTPUT_DIR}/prediction.json\", query_map, database_map)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%cd FIVR-200K\n",
    "!python evaluation.py --result_file \"../{OUTPUT_DIR}/prediction.json\" \\\n",
    "                      --relevant_labels ND,DS \\\n",
    "                      --dataset_ids ../annotations/FIVR-5K/used_videos.txt \\\n",
    "                      --annotations_file ../annotations/FIVR-5K/annotation.json\n",
    "%cd .."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%cd FIVR-200K\n",
    "!python evaluation.py --result_file \"../{OUTPUT_DIR}/prediction.json\" \\\n",
    "                      --relevant_labels ND,DS,CS \\\n",
    "                      --dataset_ids ../annotations/FIVR-5K/used_videos.txt \\\n",
    "                      --annotations_file ../annotations/FIVR-5K/annotation.json\n",
    "%cd .."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%cd FIVR-200K\n",
    "!python evaluation.py --result_file \"../{OUTPUT_DIR}/prediction.json\" \\\n",
    "                      --relevant_labels ND,DS,CS,IS \\\n",
    "                      --dataset_ids ../annotations/FIVR-5K/used_videos.txt \\\n",
    "                      --annotations_file ../annotations/FIVR-5K/annotation.json\n",
    "%cd .."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "base",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.19"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}