From 5e13ac39fbbb37e7a1dd0250890bca3467103195 Mon Sep 17 00:00:00 2001 From: keisuke umezawa Date: Tue, 27 Nov 2018 10:18:47 +0900 Subject: [PATCH] =?UTF-8?q?Colaboratory=20=E3=82=92=E4=BD=BF=E7=94=A8?= =?UTF-8?q?=E3=81=97=E3=81=A6=E4=BD=9C=E6=88=90=E3=81=97=E3=81=BE=E3=81=97?= =?UTF-8?q?=E3=81=9F=20(#74)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../01_Chainer_basic_tutorial.ipynb | 223 ++++++++++++------ 1 file changed, 145 insertions(+), 78 deletions(-) diff --git a/hands_on_ja/chainer/begginers_hands_on/01_Chainer_basic_tutorial.ipynb b/hands_on_ja/chainer/begginers_hands_on/01_Chainer_basic_tutorial.ipynb index 2e552ba..037e1fe 100644 --- a/hands_on_ja/chainer/begginers_hands_on/01_Chainer_basic_tutorial.ipynb +++ b/hands_on_ja/chainer/begginers_hands_on/01_Chainer_basic_tutorial.ipynb @@ -8,7 +8,8 @@ "provenance": [], "collapsed_sections": [ "7xCCSzxZlde1" - ] + ], + "toc_visible": true }, "kernelspec": { "name": "python3", @@ -68,24 +69,24 @@ "metadata": { "id": "QNxHfX_Blly7", "colab_type": "code", - "outputId": "767af462-7e3e-4a85-d9ee-2a085722c098", + "outputId": "aa1a5a65-8b78-48e9-ca13-5f4e8245b15b", "colab": { "base_uri": "https://localhost:8080/", - "height": 1193 + "height": 1173 } }, "cell_type": "code", "source": [ "!curl https://colab.chainer.org/install | sh -" ], - "execution_count": 1, + "execution_count": 0, "outputs": [ { "output_type": "stream", "text": [ " % Total % Received % Xferd Average Speed Time Time Time Current\n", " Dload Upload Total Spent Left Speed\n", - "\r 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0\r 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0\r100 1379 100 1379 0 0 6629 0 --:--:-- --:--:-- --:--:-- 6598\n", + "\r 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0\r100 1379 100 1379 0 0 10446 0 --:--:-- --:--:-- --:--:-- 10446\n", "+ apt -y -q install cuda-libraries-dev-9-2\n", "Reading package lists...\n", "Building dependency tree...\n", @@ -110,7 +111,7 @@ "Get:7 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1710/x86_64 cuda-nvgraph-dev-9-2 9.2.148-1 [30.1 MB]\n", "Get:8 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1710/x86_64 cuda-npp-dev-9-2 9.2.148-1 [52.0 MB]\n", "Get:9 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1710/x86_64 cuda-libraries-dev-9-2 9.2.148-1 [2,598 B]\n", - "Fetched 332 MB in 7s (47.5 MB/s)\n", + "Fetched 332 MB in 7s (48.7 MB/s)\n", "Selecting previously unselected package cuda-cublas-dev-9-2.\n", "(Reading database ... 22298 files and directories currently installed.)\n", "Preparing to unpack .../0-cuda-cublas-dev-9-2_9.2.148.1-1_amd64.deb ...\n", @@ -170,17 +171,17 @@ "metadata": { "id": "VePAE4PTmf-p", "colab_type": "code", - "outputId": "a7c2d627-764d-49d0-ff1e-d693af180137", + "outputId": "82010ac8-c095-4708-ebdc-ca3e043fc35b", "colab": { "base_uri": "https://localhost:8080/", - "height": 258 + "height": 238 } }, "cell_type": "code", "source": [ "!python -c 'import chainer; chainer.print_runtime_info()'" ], - "execution_count": 2, + "execution_count": 0, "outputs": [ { "output_type": "stream", @@ -210,17 +211,17 @@ }, "cell_type": "markdown", "source": [ - "うまくインストールできていますね。以下のチュートリアルでは、matplotlibを可視化に使いますので、これも同時にインストールしておきましょう。" + "うまくインストールできていますね。以下のチュートリアルでは、pydotをネットワーク構造の可視化に使いますので、これも同時にインストールしておきましょう。" ] }, { "metadata": { "id": "_NGiH6r2c-KZ", "colab_type": "code", - "outputId": "87d55224-3e3c-4830-a864-9d60eba323bd", + "outputId": "07ba176c-d48b-4944-fcdb-06b390eb4e63", "colab": { "base_uri": "https://localhost:8080/", - "height": 122 + "height": 102 } }, "cell_type": "code", @@ -243,14 +244,24 @@ } ] }, + { + "metadata": { + "id": "LteyQnjgb_BB", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "最後に以前使用した関数をchutilというパッケージにまとめましたので、ここでインストールしましょう。" + ] + }, { "metadata": { "id": "H-jjJSw8DmtX", "colab_type": "code", - "outputId": "ce04beac-559a-4f9f-a404-2dd4313495a7", + "outputId": "9b02c817-7c7e-4fcc-c2a6-fc73e489fe27", "colab": { "base_uri": "https://localhost:8080/", - "height": 105 + "height": 85 } }, "cell_type": "code", @@ -500,7 +511,7 @@ "metadata": { "id": "vXXaYZUPrg4U", "colab_type": "code", - "outputId": "6164a462-b546-4ea0-9c9d-89ab74675b63", + "outputId": "a1a4c18a-333c-4f48-9a2d-bf5d5b94c92f", "colab": { "base_uri": "https://localhost:8080/", "height": 85 @@ -587,7 +598,9 @@ "* 各convolution層のチャネル数、ノード数を増やす\n", "* 各convolution層のフィルター数`ksize=3`, ストライド`stride=3`、パッディング`pad=1`に変更\n", "* convolution層を1層(`conv4`)追加\n", - "* 活性化関数を`ReLU`に変更\n" + "* 活性化関数を`ReLU`に変更\n", + "\n", + "ちなみに、`L.Convolution2D`の`in_channels`や`L.Linear`の第一引数を`None`にすることで明示的にサイズを指定していませんが、実はこれでもちゃんと動きます。なぜかというと、最初の`forward`呼び出し時に、入力されたサイズで動的にメモリを確保してくれるからです。\n" ] }, { @@ -618,7 +631,7 @@ " self.fc5 = L.Linear(None, 1000) # 84 -> 1000\n", " self.fc6 = L.Linear(None, 10)\n", "\n", - " def __call__(self, x):\n", + " def forward(self, x):\n", " h = F.relu(self.conv1(x.reshape((-1, 1, 28, 28))))\n", " h = F.max_pooling_2d(h, ksize=2, stride=2)\n", " h = F.relu(self.conv2(h)) # sigmoid -> relu\n", @@ -639,7 +652,7 @@ }, "cell_type": "markdown", "source": [ - "\n", + "では、上記変更を加えたモデル`MyConvNet`を訓練してみましょう。\n", "\n", "\n" ] @@ -656,8 +669,6 @@ }, "cell_type": "code", "source": [ - "enable_cupy = True\n", - "n_epoch = 1\n", "n_epoch = 20\n", "batchsize = 128\n", "\n", @@ -698,6 +709,68 @@ } ] }, + { + "metadata": { + "id": "qVbMQ0hyghQU", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "検証用データでも90%の精度を達成しています。これならテスト用データでも目標を達成してそうですね。\n", + "\n", + "以下の関数を実行して確かめてみましょう。" + ] + }, + { + "metadata": { + "id": "_5FxlxBv0n9p", + "colab_type": "code", + "outputId": "4da38f75-bd19-4374-9508-e3cef3eefa7f", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 34 + } + }, + "cell_type": "code", + "source": [ + "from chutil.visualize.show import show_test_performance\n", + "\n", + "show_test_performance(classifier_model, test)" + ], + "execution_count": 0, + "outputs": [ + { + "output_type": "stream", + "text": [ + "Test accuracy: 0.91982424\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "lPm8yXMHgxRJ", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "達成できていましたね！パチパチ！\n", + "\n", + "と終わりにしたいところですが、もう少し細かく結果を見てみましょう。" + ] + }, + { + "metadata": { + "id": "zIImirEIeoEr", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "まずは、`show_graph`を使ってネットワーク構造を表示してみましょう。\n", + "実際にどのようなネットワークができたか確認できますね。" + ] + }, { "metadata": { "id": "KiBp2v5ygD6n", @@ -734,6 +807,20 @@ } ] }, + { + "metadata": { + "id": "8z4qOyPafHRV", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "次に、`show_loss_and_accuracy`を使って、訓練用データと検証用データの損失と精度をplotしてみましょう。\n", + "おそらく以下のことに気づくと思います。\n", + "\n", + "* 訓練用データでは、`epoch`数に伴い、損失は減少・精度は増加し、共に改善している\n", + "* しかし、検証用データでは、`epoch=5`程度で精度は頭打ちになっており、損失も最小値になったあと悪化している" + ] + }, { "metadata": { "id": "HbKYXpnk0l6q", @@ -780,29 +867,42 @@ }, { "metadata": { - "id": "_5FxlxBv0n9p", - "colab_type": "code", - "outputId": "4da38f75-bd19-4374-9508-e3cef3eefa7f", - "colab": { - "base_uri": "https://localhost:8080/", - "height": 34 - } + "id": "FSn4eo7HhlvN", + "colab_type": "text" }, - "cell_type": "code", + "cell_type": "markdown", "source": [ - "from chutil.visualize.show import show_test_performance\n", - "\n", - "show_test_performance(classifier_model, test)" - ], - "execution_count": 0, - "outputs": [ - { - "output_type": "stream", - "text": [ - "Test accuracy: 0.91982424\n" - ], - "name": "stdout" - } + "訓練用データでは精度が順調に伸びるのに、検証用データでは伸びない・・・。このような現象をどこかで聞いたことがあるような気がしますよね。" + ] + }, + { + "metadata": { + "id": "Zc-kZPtWhQed", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## 過学習と汎化性能" + ] + }, + { + "metadata": { + "id": "qMvkkQwLigT3", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### 過学習とは" + ] + }, + { + "metadata": { + "id": "xwlTnVjahQqw", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Dropout" ] }, { @@ -847,16 +947,11 @@ "metadata": { "id": "nO84XgsNDiN9", "colab_type": "code", - "outputId": "b6526c08-2ac0-4aee-f1f2-a24550c60604", - "colab": { - "base_uri": "https://localhost:8080/", - "height": 374 - } + "colab": {} }, "cell_type": "code", "source": [ - "enable_cupy = True\n", - "n_epoch = 20\n", + "n_epoch = 30\n", "batchsize = 128\n", "\n", "model = MyConvNet()\n", @@ -866,35 +961,7 @@ " classifier_model, optimizer, train, validation, n_epoch, batchsize)" ], "execution_count": 0, - "outputs": [ - { - "output_type": "stream", - "text": [ - "epoch main/loss main/accuracy val/main/loss val/main/accuracy elapsed_time\n", - "\u001b[J1 0.553697 0.794977 0.387361 0.86254 10.9894 \n", - "\u001b[J2 0.352021 0.870724 0.343159 0.873418 21.6753 \n", - "\u001b[J3 0.299985 0.88756 0.281276 0.900218 32.2488 \n", - "\u001b[J4 0.270547 0.901894 0.265802 0.902591 42.9115 \n", - "\u001b[J5 0.256918 0.90643 0.245724 0.911887 53.5615 \n", - "\u001b[J6 0.240187 0.910557 0.24804 0.910008 64.0522 \n", - "\u001b[J7 0.225157 0.91652 0.223051 0.918315 74.6426 \n", - "\u001b[J8 0.216988 0.91867 0.227083 0.914953 85.1441 \n", - "\u001b[J9 0.207234 0.923813 0.216907 0.920293 95.6334 \n", - "\u001b[J10 0.199658 0.925492 0.217243 0.917227 106.077 \n", - "\u001b[J11 0.196208 0.926382 0.230436 0.913667 117.514 \n", - "\u001b[J12 0.19097 0.930047 0.218538 0.919007 129.225 \n", - "\u001b[J13 0.186419 0.930687 0.199121 0.925732 141.079 \n", - "\u001b[J14 0.176123 0.933854 0.202717 0.92415 152.826 \n", - "\u001b[J15 0.171971 0.935322 0.204852 0.921578 164.535 \n", - "\u001b[J16 0.169379 0.936138 0.204297 0.924644 176.419 \n", - "\u001b[J17 0.167468 0.9377 0.215237 0.921282 188.069 \n", - "\u001b[J18 0.163362 0.939618 0.202339 0.92682 199.796 \n", - "\u001b[J19 0.157467 0.940365 0.206196 0.925831 211.327 \n", - "\u001b[J20 0.154054 0.941216 0.20949 0.922567 222.881 \n" - ], - "name": "stdout" - } - ] + "outputs": [] }, { "metadata": { @@ -1246,4 +1313,4 @@ ] } ] -} +} \ No newline at end of file