pll.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Partial Label Learningのノートブック"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from matplotlib import pyplot as plt\n",
    "import polars as pl\n",
    "import seaborn as sns\n",
    "import japanize_matplotlib\n",
    "\n",
    "from common import *"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# ハイパーパラメータ\n",
    "N_COMPONENTS = 50\n",
    "TEST_SIZE = 0.2\n",
    "VAL_SIZE = 0.2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Files already downloaded\n",
      "Files already downloaded\n"
     ]
    }
   ],
   "source": [
    "# データセットの読み込み\n",
    "X_train, X_val, X_test, y_train, y_val, y_test = load_mnist_pca_train_test_val(\n",
    "    n_components=N_COMPONENTS, test_size=TEST_SIZE, val_size=VAL_SIZE\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "646\n",
      "71\n",
      "5\n",
      "2\n",
      "0\n"
     ]
    }
   ],
   "source": [
    "# Partial Label Learningの設定にラベルを変更\n",
    "X_train, y_train = rewrite_label_with_pll_setting(X_train, y_train, positive_rate=0.2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "sns.histplot(pl.DataFrame(y_train.sum(-1), schema=[\"num_instances\"]), x=\"num_instances\")\n",
    "plt.title(\"各インスタンスに与えられたラベルの数\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "\n",
    "def train_lgbm_with_custom_loss(X_train, X_val, y_train, y_val):\n",
    "    \"\"\"Train LightGBM with custom loss function.\"\"\"\n",
    "    train_data = lgb.Dataset(X_train, label=np.zeros(len(y_train)), free_raw_data=False)\n",
    "    train_data._partial_labels = y_train\n",
    "    train_data._pweight = y_train / y_train.sum(axis=-1, keepdims=True)\n",
    "    valid_data = lgb.Dataset(X_val, label=y_val, free_raw_data=False)\n",
    "    params = {\n",
    "        \"objective\": \"custom\",\n",
    "        \"metric\": \"custom\",\n",
    "        \"verbose\": -1,\n",
    "        \"learning_rate\": 0.1,\n",
    "        \"num_boost_round\": 200,\n",
    "        \"num_class\": 10\n",
    "    }\n",
    "    valid_accuracies = []\n",
    "\n",
    "    def record_accuracies(p: lgb.Booster, train_data: lgb.Dataset, valid_data: lgb.Dataset):\n",
    "        valid_pred = p.predict(valid_data.data).argmax(-1)\n",
    "        valid_acc = accuracy_score(valid_data.label, valid_pred)\n",
    "        valid_accuracies.append(valid_acc)\n",
    "        print(valid_acc)\n",
    "    \n",
    "    gbm = lgb.train(\n",
    "        params,\n",
    "        train_data,\n",
    "        valid_sets=[valid_data],\n",
    "        fobj=pll_loss_objective,\n",
    "        feval=multiclass_metric,\n",
    "        callbacks=[lambda p: record_accuracies(p.model, train_data, valid_data)]\n",
    "    )\n",
    "    return gbm, valid_accuracies"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/ryoma.kobayashi/gits/techblog_git/wsl_w_lgbm/.venv/lib/python3.10/site-packages/lightgbm/engine.py:177: UserWarning: Found `num_boost_round` in params. Will use it instead of argument\n",
      "  _log_warning(f\"Found `{alias}` in params. Will use it instead of argument\")\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.7949166666666667\n",
      "0.82475\n",
      "0.8288333333333333\n",
      "0.8333333333333334\n",
      "0.8394166666666667\n",
      "0.842\n",
      "0.8456666666666667\n",
      "0.8480833333333333\n",
      "0.8495\n",
      "0.8504166666666667\n",
      "0.8530833333333333\n",
      "0.8574166666666667\n",
      "0.85975\n",
      "0.8635\n",
      "0.8654166666666666\n",
      "0.8689166666666667\n",
      "0.8715833333333334\n",
      "0.874\n",
      "0.8765\n",
      "0.8785\n",
      "0.8811666666666667\n",
      "0.884\n",
      "0.88525\n",
      "0.88875\n",
      "0.8914166666666666\n",
      "0.89275\n",
      "0.8945833333333333\n",
      "0.89675\n",
      "0.8983333333333333\n",
      "0.90125\n",
      "0.9026666666666666\n",
      "0.9040833333333333\n",
      "0.9069166666666667\n",
      "0.909\n",
      "0.9094166666666667\n",
      "0.911\n",
      "0.9139166666666667\n",
      "0.9148333333333334\n",
      "0.9169166666666667\n",
      "0.918\n",
      "0.9193333333333333\n",
      "0.9206666666666666\n",
      "0.922\n",
      "0.9235833333333333\n",
      "0.925\n",
      "0.9265833333333333\n",
      "0.9266666666666666\n",
      "0.92775\n",
      "0.9285833333333333\n",
      "0.9295\n",
      "0.9298333333333333\n",
      "0.931\n",
      "0.9321666666666667\n",
      "0.9326666666666666\n",
      "0.9334166666666667\n",
      "0.9344166666666667\n",
      "0.9351666666666667\n",
      "0.9360833333333334\n",
      "0.937\n",
      "0.93775\n",
      "0.93825\n",
      "0.93875\n",
      "0.93925\n",
      "0.9404166666666667\n",
      "0.9404166666666667\n",
      "0.9408333333333333\n",
      "0.9413333333333334\n",
      "0.9414166666666667\n",
      "0.94225\n",
      "0.94225\n",
      "0.9429166666666666\n",
      "0.9431666666666667\n",
      "0.9434166666666667\n",
      "0.944\n",
      "0.944\n",
      "0.94425\n",
      "0.945\n",
      "0.94575\n",
      "0.946\n",
      "0.9460833333333334\n",
      "0.9464166666666667\n",
      "0.94675\n",
      "0.9474166666666667\n",
      "0.94775\n",
      "0.948\n",
      "0.9489166666666666\n",
      "0.9486666666666667\n",
      "0.94925\n",
      "0.9495\n",
      "0.9493333333333334\n",
      "0.9498333333333333\n",
      "0.95\n",
      "0.95075\n",
      "0.951\n",
      "0.9510833333333333\n",
      "0.9511666666666667\n",
      "0.9516666666666667\n",
      "0.952\n",
      "0.9520833333333333\n",
      "0.9528333333333333\n",
      "0.9529166666666666\n",
      "0.953\n",
      "0.9531666666666667\n",
      "0.9530833333333333\n",
      "0.9533333333333334\n",
      "0.954\n",
      "0.9539166666666666\n",
      "0.9540833333333333\n",
      "0.9539166666666666\n",
      "0.9539166666666666\n",
      "0.9539166666666666\n",
      "0.9540833333333333\n",
      "0.9540833333333333\n",
      "0.9543333333333334\n",
      "0.9544166666666667\n",
      "0.9543333333333334\n",
      "0.95425\n",
      "0.9541666666666667\n",
      "0.9546666666666667\n",
      "0.9545833333333333\n",
      "0.9549166666666666\n",
      "0.955\n",
      "0.955\n",
      "0.9555\n",
      "0.9556666666666667\n",
      "0.95575\n",
      "0.9555833333333333\n",
      "0.95575\n",
      "0.95575\n",
      "0.9558333333333333\n",
      "0.9560833333333333\n",
      "0.95625\n",
      "0.9564166666666667\n",
      "0.956\n",
      "0.95625\n",
      "0.9565833333333333\n",
      "0.9565\n",
      "0.9565\n",
      "0.9565\n",
      "0.95625\n",
      "0.9566666666666667\n",
      "0.9570833333333333\n",
      "0.9568333333333333\n",
      "0.9571666666666667\n",
      "0.9574166666666667\n",
      "0.9573333333333334\n",
      "0.9569166666666666\n",
      "0.95725\n",
      "0.9574166666666667\n",
      "0.957\n",
      "0.9570833333333333\n",
      "0.9575\n",
      "0.9574166666666667\n",
      "0.9574166666666667\n",
      "0.9574166666666667\n",
      "0.95775\n",
      "0.95775\n",
      "0.9575833333333333\n",
      "0.9575833333333333\n",
      "0.9579166666666666\n",
      "0.9579166666666666\n",
      "0.958\n",
      "0.9579166666666666\n",
      "0.95775\n",
      "0.9578333333333333\n",
      "0.9578333333333333\n",
      "0.958\n",
      "0.9579166666666666\n",
      "0.958\n",
      "0.9581666666666667\n",
      "0.9581666666666667\n",
      "0.95825\n",
      "0.9583333333333334\n",
      "0.9585\n",
      "0.9585833333333333\n",
      "0.9588333333333333\n",
      "0.9588333333333333\n",
      "0.9589166666666666\n",
      "0.9590833333333333\n",
      "0.9590833333333333\n",
      "0.959\n",
      "0.9589166666666666\n",
      "0.9588333333333333\n",
      "0.9588333333333333\n",
      "0.9589166666666666\n",
      "0.9585833333333333\n",
      "0.95875\n",
      "0.9589166666666666\n",
      "0.9589166666666666\n",
      "0.9589166666666666\n",
      "0.9591666666666666\n",
      "0.95925\n",
      "0.95925\n",
      "0.95925\n",
      "0.9593333333333334\n",
      "0.9595\n",
      "0.9595\n",
      "0.9595833333333333\n",
      "0.9594166666666667\n",
      "0.9596666666666667\n",
      "Accuracy: 0.9596666666666667\n"
     ]
    }
   ],
   "source": [
    "gbm, valid_accuracies =train_lgbm_with_custom_loss(X_train, X_val, y_train, y_val)\n",
    "y_pred = gbm.predict(X_val)\n",
    "print(\"Accuracy: {}\".format(accuracy_score(y_val, y_pred.argmax(-1))))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.save(\"pll_valid_accuracies.npy\", valid_accuracies)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": ".venv",
   "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": "3.10.4"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}