expnp_tsubame.ipynb

{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "5OcSaXhD8Nh0"
   },
   "source": [
    "# 手書き文字認識を行うNeural Networkのレシピ"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "113c4jyo8V3c"
   },
   "source": [
    "## このスクリプトについて\n",
    "ニューラルネットワークを使った画像認識のレシピです。今回はチュートリアルなどでよく使われている手書き文字データセットMNISTを学習してモデルを作成し、テスト画像を入力してその認識精度を確認します。"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "eMZptlN9J73d"
   },
   "source": [
    "## 実験条件の設定"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "ItESsJMIKEz8"
   },
   "source": [
    "実験で用いるデータセットのサイズを指定します。MNIST(FashionMNIST)は60000枚の画像を含むデータセットですが、ここで指定した割合のデータを学習・評価に用います。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "GjORKgwiKnOO"
   },
   "outputs": [],
   "source": [
    "# 20% -> 0.2\n",
    "util_rate = 0.2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 100"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "UpTablEP8mlR"
   },
   "source": [
    "## 環境のセットアップ\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "rhLnQJ0w9WgI"
   },
   "source": [
    "### GPUの種類を確認する\n",
    "スピード　：　Tesla K80 < Tesla T4 < Tesla P100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "RrSPtdQWiL1U"
   },
   "outputs": [],
   "source": [
    "!nvidia-smi\n",
    "!module purge\n",
    "!module load python-extension/3.4\n",
    "!pip install matplotlib numpy\n",
    "!python3 -m pip install --user seaborn torchvision ipywidgets"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "Zr9I4qM9-sib"
   },
   "source": [
    "### 必要なライブラリのインポート"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "ZXNuuII9icFj"
   },
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "from torchvision import datasets, transforms\n",
    "from torch.utils.data import DataLoader\n",
    "import time"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "XHcKzY0G-0Dl"
   },
   "source": [
    "### GPUが有効になっているか確認\n",
    "CUDAはGPUを供給するメーカーのnVIDIA社が提供するドライバである。CPUと表示された場合はランタイムのタイプをGPUに変更してライブラリをもう一度インポートする。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "CXEit4LSifJx"
   },
   "outputs": [],
   "source": [
    "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
    "print(device)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "UfAjrWfT_2vb"
   },
   "source": [
    "## ニューラルネットワークのパラメータと学習の設定"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "NILRwCI6AonM"
   },
   "source": [
    "データセットのダウンロードには6分程度かかる場合があります。  \n",
    "ダウンロードエラーになる場合は8行目の先頭に#を追加して、9行目の先頭の#を削除してください。  \n",
    "(コメントアウト)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "e45DUQ1ZiiGt"
   },
   "outputs": [],
   "source": [
    "# 前処理\n",
    "transform = transforms.Compose([\n",
    "    # 画像をTensorに変換してくれる\n",
    "    # チャネルラストをチャネルファーストに\n",
    "    # 0〜255の整数値を0.0〜1.0の浮動小数点に変換してくれる\n",
    "    transforms.ToTensor()                              \n",
    "])\n",
    "dataset = datasets.MNIST(root=\"./data\", train=True, download=True, transform=transform)\n",
    "#dataset = datasets.FashionMNIST(root=\"./data\", train=True, download=True, transform=transform)\n",
    "\n",
    "# datasetのうち1/6 (10000件)をテスト用に分ける\n",
    "# 残りの5/6の中で、実験条件で設定した割合のデータをdatasetから取り出す\n",
    "n_imgs = len(dataset)\n",
    "n_val = int(n_imgs / 6)\n",
    "n_train = int((n_imgs - n_val) * util_rate)\n",
    "n_rem = n_imgs - n_val - n_train\n",
    "train_dataset, val_dataset, _ = torch.utils.data.random_split(\n",
    "    dataset, [n_train, n_val, n_rem])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "0_Y0CMJliqhJ"
   },
   "outputs": [],
   "source": [
    "print('Batch Size : ', batch_size)\n",
    "train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, drop_last=True)\n",
    "val_dataloader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)\n",
    "train_iter = iter(train_dataloader)\n",
    "# batch_size個だけミニバッチからデータをロードする\n",
    "imgs, labels = next(iter(train_iter))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "LsDw4pIKJ-Aa"
   },
   "source": [
    "## trainingに用いる画像の確認"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "cEll--hXiykC"
   },
   "outputs": [],
   "source": [
    "print(labels)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "-Jhnc48ui696"
   },
   "outputs": [],
   "source": [
    "img = imgs[0]\n",
    "# 画像データを表示するために、チャネルファーストのデータをチャネルラストに変換する\n",
    "img_permute = img.permute(1, 2, 0)\n",
    "# tensorから2次元のarrayに変換する\n",
    "sns.heatmap(img_permute.numpy()[:, :, 0])"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "CwO_2FsFAPgO"
   },
   "source": [
    "## ニューラルネットワークの定義"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "UVPMtjWNi-S2"
   },
   "outputs": [],
   "source": [
    "class MLP(nn.Module):\n",
    "    def __init__(self):\n",
    "        super().__init__()\n",
    "        self.fc1 = nn.Linear(28 * 28, 100)\n",
    "        self.relu1 = nn.ReLU(inplace=True)\n",
    "        self.fc2 = nn.Linear(100, 100)\n",
    "        self.relu2 = nn.ReLU(inplace=True)\n",
    "        self.fc3 = nn.Linear(100, 10)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.fc1(x)\n",
    "        x = self.relu1(x)\n",
    "        x = self.fc2(x)\n",
    "        x = self.relu2(x)\n",
    "        x = self.fc3(x)\n",
    "\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "4aWaB7_5jBV8"
   },
   "outputs": [],
   "source": [
    "model = MLP()\n",
    "\n",
    "print(model.to(device))"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "iID6N8p7jJMU"
   },
   "source": [
    "## ニューラルネットワークのトレーニング"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "D9YuNPEpjrRy"
   },
   "outputs": [],
   "source": [
    "criterion = nn.CrossEntropyLoss()\n",
    "optimizer = optim.Adam(model.parameters(), lr=0.001)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "iSE8OJRWjN62"
   },
   "outputs": [],
   "source": [
    "num_epochs = 10\n",
    "train_losses = []\n",
    "train_accs = []\n",
    "val_losses = []\n",
    "val_accs = []\n",
    "\n",
    "start_time = time.process_time()\n",
    "\n",
    "for epoch in range(num_epochs):\n",
    "    #train\n",
    "    running_train_loss = 0.0\n",
    "    running_train_acc = 0.0\n",
    "    for imgs, labels in train_dataloader:\n",
    "        imgs = imgs.view(-1, 28*28)\n",
    "        imgs = imgs.to(device)\n",
    "        labels = labels.to(device)\n",
    "        optimizer.zero_grad()\n",
    "        output = model(imgs)\n",
    "        loss = criterion(output, labels)\n",
    "        running_train_loss += loss.item()\n",
    "        # dim=1 => 0-9の分類方向のMax値を返す\n",
    "        pred = torch.argmax(output, dim=1)\n",
    "        running_train_acc += torch.mean(pred.eq(labels).float())\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "    # 画像数で割る\n",
    "    running_train_loss /= len(train_dataloader)\n",
    "    running_train_acc /= len(train_dataloader)\n",
    "    train_losses.append(running_train_loss)\n",
    "    train_accs.append(running_train_acc.to('cpu'))\n",
    "\n",
    "    #test\n",
    "    running_val_loss = 0.0\n",
    "    running_val_acc = 0.0\n",
    "    with torch.no_grad():\n",
    "      for imgs, labels in val_dataloader:\n",
    "        imgs = imgs.view(-1, 28*28)\n",
    "        imgs = imgs.to(device)\n",
    "        labels = labels.to(device)\n",
    "        output = model(imgs)\n",
    "        loss = criterion(output, labels)\n",
    "        running_val_loss += loss.item()\n",
    "        # dim=1 => 0-9の分類方向のMax値を返す\n",
    "        pred = torch.argmax(output, dim=1)\n",
    "        running_val_acc += torch.mean(pred.eq(labels).float())\n",
    "        \n",
    "    # 画像数で割る\n",
    "    running_val_loss /= len(val_dataloader)\n",
    "    running_val_acc /= len(val_dataloader)\n",
    "    val_losses.append(running_val_loss)\n",
    "    val_accs.append(running_val_acc.to('cpu'))\n",
    "\n",
    "    print(\"epoch: {}, train_loss: {}, train_acc: {}, test_loss: {}, test_acc: {}\".format(epoch, running_train_loss, running_train_acc, running_val_loss, running_val_acc))\n",
    "\n",
    "end_time = time.process_time()\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "S6jtwOJxCaXH"
   },
   "source": [
    "## 結果の確認\n",
    "\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "TM0zWh-mCtEU"
   },
   "source": [
    "Loss：正しい答えとニューラルネットワーク出力の誤差  \n",
    "Accuracy：ニューラルネットワーク推論結果の正解率"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "LCJ1dJBgA0hJ"
   },
   "source": [
    "### 学習データ"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "TtkjC_6wBajD"
   },
   "source": [
    "Loss:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "84C118MABFiS"
   },
   "outputs": [],
   "source": [
    "plt.plot(train_losses)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "wprzcVRwBgZ-"
   },
   "source": [
    "Accuracy:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "bMmh3MGYBAkA"
   },
   "outputs": [],
   "source": [
    "plt.plot(train_accs)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "baNhdi0rA9c_"
   },
   "source": [
    "### テストデータ"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "YX3NSzteBiKl"
   },
   "source": [
    "Loss:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "kxbaeyOzj39C"
   },
   "outputs": [],
   "source": [
    "plt.plot(val_losses)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "R8BXvItXBjMi"
   },
   "source": [
    "Accuracy:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "RqmtPYz9j7KT"
   },
   "outputs": [],
   "source": [
    "plt.plot(val_accs)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "fsv2Mi3BIThJ"
   },
   "source": [
    "## Summary"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "fJVy1JxpItwq"
   },
   "outputs": [],
   "source": [
    "!nvidia-smi\n",
    "print('Accuracy:{:.04f}'.format(running_val_acc))\n",
    "print('Time cost:{:.04f}'.format(end_time-start_time), 's')"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {
    "id": "vGLWPrTLOq9I"
   },
   "source": [
    "## Appendix\n",
    "ニューラルネットワークの推論が間違えていた画像を確認する。  \n",
    "最大10枚まで表示"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "wusTs8i35QJM"
   },
   "outputs": [],
   "source": [
    "demo_dataloader = DataLoader(val_dataset, batch_size=1, shuffle=False)\n",
    "max_images = 10\n",
    "ind = 0\n",
    "fig = plt.figure(figsize=(6,4*((max_images+1)//2)))\n",
    "with torch.no_grad():\n",
    "  for imgs, labels in demo_dataloader:\n",
    "    imgs_reshape = imgs.view(1, -1)\n",
    "    imgs_reshape = imgs_reshape.to(device)\n",
    "    labels = labels.to(device)\n",
    "    output = model(imgs_reshape)\n",
    "    # dim=1 => 0-9の分類方向のMax値を返す\n",
    "    pred = torch.argmax(output, dim=1)\n",
    "    if not labels==pred:\n",
    "      ind += 1\n",
    "      ax = fig.add_subplot((max_images+1)//2, 2, ind)\n",
    "      ax.imshow(imgs_reshape.cpu().reshape(28,28), cmap = \"gray\")\n",
    "      ax.set_title(\"correct: \" + str(int(labels)) + \"\\npredict: \" + str(int(pred)))\n",
    "    if max_images <= ind:\n",
    "      break\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "LQLgn7UWfBe5"
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "colab": {
   "collapsed_sections": [],
   "name": "expnp_tsubame.ipynb",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.9.18"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}