{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"In this notebook, referring to [Getting Started](https://www.kaggle.com/code/robikscube/nfl-player-contact-detection-getting-started), I created a training model using xgboost with some features other than distance, such as speed and acceleration.","metadata":{}},{"cell_type":"code","source":"import os\nimport torch\n\nclass Config:\n    AUTHOR = \"colum2131\"\n\n    NAME = \"NFLC-\" + \"Exp001-simple-xgb-baseline\"\n\n    COMPETITION = \"nfl-player-contact-detection\"\n\n    seed = 42\n    num_fold = 5\n    \n    xgb_params = {\n        'objective': 'binary:logistic',\n        'eval_metric': 'auc',\n        'learning_rate':0.03,\n        'tree_method':'hist' if not torch.cuda.is_available() else 'gpu_hist'\n    }","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2022-12-12T12:44:18.775920Z","iopub.execute_input":"2022-12-12T12:44:18.776863Z","iopub.status.idle":"2022-12-12T12:44:20.635053Z","shell.execute_reply.started":"2022-12-12T12:44:18.776763Z","shell.execute_reply":"2022-12-12T12:44:20.633948Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import os\nimport gc\nimport subprocess\n\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport matplotlib.patches as patches\nfrom IPython.display import Video, display\n\nfrom scipy.optimize import minimize\nimport cv2\nfrom glob import glob\nfrom tqdm import tqdm\n\nfrom sklearn.model_selection import GroupKFold\nfrom sklearn.metrics import (\n    roc_auc_score,\n    matthews_corrcoef,\n)\n\nimport xgboost as xgb\n\nimport torch\n\nif torch.cuda.is_available():\n    import cupy \n    import cudf\n    from cuml import ForestInference","metadata":{"execution":{"iopub.status.busy":"2022-12-12T12:44:20.637146Z","iopub.execute_input":"2022-12-12T12:44:20.637941Z","iopub.status.idle":"2022-12-12T12:44:25.099549Z","shell.execute_reply.started":"2022-12-12T12:44:20.637903Z","shell.execute_reply":"2022-12-12T12:44:25.098556Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def setup(cfg):\n    cfg.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n    \n    # set dirs\n    cfg.INPUT = f'../input/{cfg.COMPETITION}'\n    cfg.EXP = cfg.NAME\n    cfg.OUTPUT_EXP = cfg.NAME\n    cfg.SUBMISSION = './'\n    cfg.DATASET = '../input/'\n\n    cfg.EXP_MODEL = os.path.join(cfg.EXP, 'model')\n    cfg.EXP_FIG = os.path.join(cfg.EXP, 'fig')\n    cfg.EXP_PREDS = os.path.join(cfg.EXP, 'preds')\n\n    # make dirs\n    for d in [cfg.EXP_MODEL, cfg.EXP_FIG, cfg.EXP_PREDS]:\n        os.makedirs(d, exist_ok=True)\n        \n    return cfg","metadata":{"execution":{"iopub.status.busy":"2022-12-12T12:44:25.100904Z","iopub.execute_input":"2022-12-12T12:44:25.101251Z","iopub.status.idle":"2022-12-12T12:44:25.110841Z","shell.execute_reply.started":"2022-12-12T12:44:25.101217Z","shell.execute_reply":"2022-12-12T12:44:25.109583Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# ==============================\n# function\n# ==============================\n# ref: https://www.kaggle.com/code/robikscube/nfl-player-contact-detection-getting-started\ndef add_contact_id(df):\n    # Create contact ids\n    df[\"contact_id\"] = (\n        df[\"game_play\"]\n        + \"_\"\n        + df[\"step\"].astype(\"str\")\n        + \"_\"\n        + df[\"nfl_player_id_1\"].astype(\"str\")\n        + \"_\"\n        + df[\"nfl_player_id_2\"].astype(\"str\")\n    )\n    return df\n\ndef expand_contact_id(df):\n    \"\"\"\n    Splits out contact_id into seperate columns.\n    \"\"\"\n    df[\"game_play\"] = df[\"contact_id\"].str[:12]\n    df[\"step\"] = df[\"contact_id\"].str.split(\"_\").str[-3].astype(\"int\")\n    df[\"nfl_player_id_1\"] = df[\"contact_id\"].str.split(\"_\").str[-2]\n    df[\"nfl_player_id_2\"] = df[\"contact_id\"].str.split(\"_\").str[-1]\n    return df\n\n# cross validation\ndef get_groupkfold(train, target_col, group_col, n_splits):\n    kf = GroupKFold(n_splits=n_splits)\n    generator = kf.split(train, train[target_col], train[group_col])\n    fold_series = []\n    for fold, (idx_train, idx_valid) in enumerate(generator):\n        fold_series.append(pd.Series(fold, index=idx_valid))\n    fold_series = pd.concat(fold_series).sort_index()\n    return fold_series\n\n# xgboost code\ndef fit_xgboost(cfg, X, y, params, add_suffix=''):\n    \"\"\"\n    xgb_params = {\n        'objective': 'binary:logistic',\n        'eval_metric': 'auc',\n        'learning_rate':0.01,\n        'tree_method':'gpu_hist'\n    }\n    \"\"\"\n    oof_pred = np.zeros(len(y), dtype=np.float32)\n    for fold in sorted(cfg.folds.unique()):\n        if fold == -1: continue\n        idx_train = (cfg.folds!=fold)\n        idx_valid = (cfg.folds==fold)\n        x_train, y_train = X[idx_train], y[idx_train]\n        x_valid, y_valid = X[idx_valid], y[idx_valid]\n        display(pd.Series(y_valid).value_counts())\n\n        xgb_train = xgb.DMatrix(x_train, label=y_train)\n        xgb_valid = xgb.DMatrix(x_valid, label=y_valid)\n        evals = [(xgb_train,'train'),(xgb_valid,'eval')]\n\n        model = xgb.train(\n            params,\n            xgb_train,\n            num_boost_round=10_000,\n            early_stopping_rounds=100,\n            evals=evals,\n            verbose_eval=100,\n        )\n\n        model_path = os.path.join(cfg.EXP_MODEL, f'xgb_fold{fold}{add_suffix}.model')\n        model.save_model(model_path)\n        if not torch.cuda.is_available():\n            model = xgb.Booster().load_model(model_path)\n        else:\n            model = ForestInference.load(model_path, output_class=True, model_type='xgboost')\n        pred_i = model.predict_proba(x_valid)[:, 1]\n        oof_pred[x_valid.index] = pred_i\n        score = round(roc_auc_score(y_valid, pred_i), 5)\n        print(f'Performance of the prediction: {score}\\n')\n        del model; gc.collect()\n\n    np.save(os.path.join(cfg.EXP_PREDS, f'oof_pred{add_suffix}'), oof_pred)\n    score = round(roc_auc_score(y, oof_pred), 5)\n    print(f'All Performance of the prediction: {score}')\n    return oof_pred\n\ndef pred_xgboost(X, data_dir, add_suffix=''):\n    models = glob(os.path.join(data_dir, f'xgb_fold*{add_suffix}.model'))\n    if not torch.cuda.is_available():\n         models = [xgb.Booster().load_model(model_path) for model in models]\n    else:\n        models = [ForestInference.load(model, output_class=True, model_type='xgboost') for model in models]\n    preds = np.array([model.predict_proba(X)[:, 1] for model in models])\n    preds = np.mean(preds, axis=0)\n    return preds","metadata":{"execution":{"iopub.status.busy":"2022-12-12T12:44:25.114287Z","iopub.execute_input":"2022-12-12T12:44:25.114703Z","iopub.status.idle":"2022-12-12T12:44:25.134987Z","shell.execute_reply.started":"2022-12-12T12:44:25.114599Z","shell.execute_reply":"2022-12-12T12:44:25.134156Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# ==============================\n# read data\n# ==============================\ncfg = setup(Config)\n\nif not torch.cuda.is_available():\n    tr_tracking = pd.read_csv(os.path.join(cfg.INPUT, 'train_player_tracking.csv'), parse_dates=[\"datetime\"])\n    te_tracking = pd.read_csv(os.path.join(cfg.INPUT, 'test_player_tracking.csv'), parse_dates=[\"datetime\"])\n    # tr_helmets = pd.read_csv(os.path.join(cfg.INPUT, 'train_baseline_helmets.csv'))\n    # te_helmets = pd.read_csv(os.path.join(cfg.INPUT, 'test_baseline_helmets.csv'))\n    # tr_video_metadata = pd.read_csv(os.path.join(cfg.INPUT, 'train_video_metadata.csv'))\n    # te_video_metadata = pd.read_csv(os.path.join(cfg.INPUT, 'test_video_metadata.csv'))\n    sub = pd.read_csv(os.path.join(cfg.INPUT, 'sample_submission.csv'))\n\n    train = pd.read_csv(os.path.join(cfg.INPUT, 'train_labels.csv'), parse_dates=[\"datetime\"])\n    test = expand_contact_id(sub)\n    \nelse:\n    tr_tracking = cudf.read_csv(os.path.join(cfg.INPUT, 'train_player_tracking.csv'), parse_dates=[\"datetime\"])\n    te_tracking = cudf.read_csv(os.path.join(cfg.INPUT, 'test_player_tracking.csv'), parse_dates=[\"datetime\"])\n    # tr_helmets = cudf.read_csv(os.path.join(cfg.INPUT, 'train_baseline_helmets.csv'))\n    # te_helmets = cudf.read_csv(os.path.join(cfg.INPUT, 'test_baseline_helmets.csv'))\n    # tr_video_metadata = cudf.read_csv(os.path.join(cfg.INPUT, 'train_video_metadata.csv'))\n    # te_video_metadata = cudf.read_csv(os.path.join(cfg.INPUT, 'test_video_metadata.csv'))\n    sub = pd.read_csv(os.path.join(cfg.INPUT, 'sample_submission.csv'))\n\n    train = cudf.read_csv(os.path.join(cfg.INPUT, 'train_labels.csv'), parse_dates=[\"datetime\"])\n    test = cudf.DataFrame(expand_contact_id(sub))","metadata":{"execution":{"iopub.status.busy":"2022-12-12T12:44:25.137653Z","iopub.execute_input":"2022-12-12T12:44:25.137947Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"The following code is used to create the features.  \nBasically, the numerical features contained in player_tracking.csv are merged into player_id_1 and player_id_2 respectively.","metadata":{}},{"cell_type":"code","source":"# ==============================\n# feature engineering\n# ==============================\ndef create_features(df, tr_tracking, merge_col=\"step\", use_cols=[\"x_position\", \"y_position\"]):\n    output_cols = []\n    df_combo = (\n        df.astype({\"nfl_player_id_1\": \"str\"})\n        .merge(\n            tr_tracking.astype({\"nfl_player_id\": \"str\"})[\n                [\"game_play\", merge_col, \"nfl_player_id\",] + use_cols\n            ],\n            left_on=[\"game_play\", merge_col, \"nfl_player_id_1\"],\n            right_on=[\"game_play\", merge_col, \"nfl_player_id\"],\n            how=\"left\",\n        )\n        .rename(columns={c: c+\"_1\" for c in use_cols})\n        .drop(\"nfl_player_id\", axis=1)\n        .merge(\n            tr_tracking.astype({\"nfl_player_id\": \"str\"})[\n                [\"game_play\", merge_col, \"nfl_player_id\"] + use_cols\n            ],\n            left_on=[\"game_play\", merge_col, \"nfl_player_id_2\"],\n            right_on=[\"game_play\", merge_col, \"nfl_player_id\"],\n            how=\"left\",\n        )\n        .drop(\"nfl_player_id\", axis=1)\n        .rename(columns={c: c+\"_2\" for c in use_cols})\n        .sort_values([\"game_play\", merge_col, \"nfl_player_id_1\", \"nfl_player_id_2\"])\n        .reset_index(drop=True)\n    )\n    output_cols += [c+\"_1\" for c in use_cols]\n    output_cols += [c+\"_2\" for c in use_cols]\n    \n    if (\"x_position\" in use_cols) & (\"y_position\" in use_cols):\n        index = df_combo['x_position_2'].notnull()\n        if torch.cuda.is_available():\n            index = index.to_array()\n        distance_arr = np.full(len(index), np.nan)\n        tmp_distance_arr = np.sqrt(\n            np.square(df_combo.loc[index, \"x_position_1\"] - df_combo.loc[index, \"x_position_2\"])\n            + np.square(df_combo.loc[index, \"y_position_1\"]- df_combo.loc[index, \"y_position_2\"])\n        )\n        if torch.cuda.is_available():\n            tmp_distance_arr = tmp_distance_arr.to_array()\n        distance_arr[index] = tmp_distance_arr\n        df_combo['distance'] = distance_arr\n        output_cols += [\"distance\"]\n        \n    df_combo['G_flug'] = (df_combo['nfl_player_id_2']==\"G\")\n    output_cols += [\"G_flug\"]\n    return df_combo, output_cols\n\n\nuse_cols = [\n    'x_position', 'y_position', 'speed', 'distance',\n    'direction', 'orientation', 'acceleration', 'sa'\n]\ntrain, feature_cols = create_features(train, tr_tracking, use_cols=use_cols)\ntest, feature_cols = create_features(test, te_tracking, use_cols=use_cols)\nif torch.cuda.is_available():\n    train = train.to_pandas()\n    test = test.to_pandas()\n\ndisplay(train)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# ==============================\n# training & inference\n# ==============================\ntrain_X = train[feature_cols]\ntest_X = test[feature_cols]\ntrain_y = train['contact']\ncfg.folds = get_groupkfold(train, 'contact', 'game_play', cfg.num_fold)\ncfg.folds.to_csv(os.path.join(cfg.EXP_PREDS, 'folds.csv'), index=False)\n\noof_pred = fit_xgboost(cfg, train_X, train_y, cfg.xgb_params, add_suffix=\"_xgb_1st\")\nsub_pred = pred_xgboost(test_X, cfg.EXP_MODEL, add_suffix=\"_xgb_1st\")","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# ==============================\n# optimize\n# ==============================\ndef func(x_list):\n    score = matthews_corrcoef(train['contact'], oof_pred>x_list[0])\n    return -score\n\nx0 = [0.5]\nresult = minimize(func, x0,  method=\"nelder-mead\")\ncfg.threshold = result.x[0]\nprint(\"score:\", round(matthews_corrcoef(train['contact'], oof_pred>cfg.threshold), 5))\nprint(\"threshold\", round(cfg.threshold, 5))\n\ntest = add_contact_id(test)\ntest['contact'] = (sub_pred > cfg.threshold).astype(int)\ntest[['contact_id', 'contact']].to_csv('submission.csv', index=False)\ndisplay(test[['contact_id', 'contact']].head())","metadata":{"trusted":true},"execution_count":null,"outputs":[]}]}