courses/udacity_deep_learning/4_convolutions.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "version": "0.3.2",
      "views": {},
      "default_view": {},
      "name": "4_convolutions.ipynb",
      "provenance": [],
      "toc_visible": true
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "4embtkV0pNxM",
        "colab_type": "text"
      },
      "source": [
        "Deep Learning\n",
        "=============\n",
        "\n",
        "Assignment 4\n",
        "------------\n",
        "\n",
        "Previously in `2_fullyconnected.ipynb` and `3_regularization.ipynb`, we trained fully connected networks to classify [notMNIST](http://yaroslavvb.blogspot.com/2011/09/notmnist-dataset.html) characters.\n",
        "\n",
        "The goal of this assignment is make the neural network convolutional."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "tm2CQN_Cpwj0",
        "colab_type": "code",
        "colab": {
          "autoexec": {
            "startup": false,
            "wait_interval": 0
          }
        },
        "cellView": "both"
      },
      "source": [
        "# These are all the modules we'll be using later. Make sure you can import them\n",
        "# before proceeding further.\n",
        "from __future__ import print_function\n",
        "import numpy as np\n",
        "import tensorflow as tf\n",
        "from six.moves import cPickle as pickle\n",
        "from six.moves import range"
      ],
      "outputs": [],
      "execution_count": 0
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "y3-cj1bpmuxc",
        "colab_type": "code",
        "colab": {
          "autoexec": {
            "startup": false,
            "wait_interval": 0
          },
          "output_extras": [
            {
              "item_id": 1
            }
          ]
        },
        "cellView": "both",
        "executionInfo": {
          "elapsed": 11948,
          "status": "ok",
          "timestamp": 1446658914837,
          "user": {
            "color": "",
            "displayName": "",
            "isAnonymous": false,
            "isMe": true,
            "permissionId": "",
            "photoUrl": "",
            "sessionId": "0",
            "userId": ""
          },
          "user_tz": 480
        },
        "outputId": "016b1a51-0290-4b08-efdb-8c95ffc3cd01"
      },
      "source": [
        "pickle_file = 'notMNIST.pickle'\n",
        "\n",
        "with open(pickle_file, 'rb') as f:\n",
        "  save = pickle.load(f)\n",
        "  train_dataset = save['train_dataset']\n",
        "  train_labels = save['train_labels']\n",
        "  valid_dataset = save['valid_dataset']\n",
        "  valid_labels = save['valid_labels']\n",
        "  test_dataset = save['test_dataset']\n",
        "  test_labels = save['test_labels']\n",
        "  del save  # hint to help gc free up memory\n",
        "  print('Training set', train_dataset.shape, train_labels.shape)\n",
        "  print('Validation set', valid_dataset.shape, valid_labels.shape)\n",
        "  print('Test set', test_dataset.shape, test_labels.shape)"
      ],
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Training set (200000, 28, 28) (200000,)\n",
            "Validation set (10000, 28, 28) (10000,)\n",
            "Test set (18724, 28, 28) (18724,)\n"
          ],
          "name": "stdout"
        }
      ],
      "execution_count": 0
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "L7aHrm6nGDMB",
        "colab_type": "text"
      },
      "source": [
        "Reformat into a TensorFlow-friendly shape:\n",
        "- convolutions need the image data formatted as a cube (width by height by #channels)\n",
        "- labels as float 1-hot encodings."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "IRSyYiIIGIzS",
        "colab_type": "code",
        "colab": {
          "autoexec": {
            "startup": false,
            "wait_interval": 0
          },
          "output_extras": [
            {
              "item_id": 1
            }
          ]
        },
        "cellView": "both",
        "executionInfo": {
          "elapsed": 11952,
          "status": "ok",
          "timestamp": 1446658914857,
          "user": {
            "color": "",
            "displayName": "",
            "isAnonymous": false,
            "isMe": true,
            "permissionId": "",
            "photoUrl": "",
            "sessionId": "0",
            "userId": ""
          },
          "user_tz": 480
        },
        "outputId": "650a208c-8359-4852-f4f5-8bf10e80ef6c"
      },
      "source": [
        "image_size = 28\n",
        "num_labels = 10\n",
        "num_channels = 1 # grayscale\n",
        "\n",
        "import numpy as np\n",
        "\n",
        "def reformat(dataset, labels):\n",
        "  dataset = dataset.reshape(\n",
        "    (-1, image_size, image_size, num_channels)).astype(np.float32)\n",
        "  labels = (np.arange(num_labels) == labels[:,None]).astype(np.float32)\n",
        "  return dataset, labels\n",
        "train_dataset, train_labels = reformat(train_dataset, train_labels)\n",
        "valid_dataset, valid_labels = reformat(valid_dataset, valid_labels)\n",
        "test_dataset, test_labels = reformat(test_dataset, test_labels)\n",
        "print('Training set', train_dataset.shape, train_labels.shape)\n",
        "print('Validation set', valid_dataset.shape, valid_labels.shape)\n",
        "print('Test set', test_dataset.shape, test_labels.shape)"
      ],
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Training set (200000, 28, 28, 1) (200000, 10)\n",
            "Validation set (10000, 28, 28, 1) (10000, 10)\n",
            "Test set (18724, 28, 28, 1) (18724, 10)\n"
          ],
          "name": "stdout"
        }
      ],
      "execution_count": 0
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "AgQDIREv02p1",
        "colab_type": "code",
        "colab": {
          "autoexec": {
            "startup": false,
            "wait_interval": 0
          }
        },
        "cellView": "both"
      },
      "source": [
        "def accuracy(predictions, labels):\n",
        "  return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))\n",
        "          / predictions.shape[0])"
      ],
      "outputs": [],
      "execution_count": 0
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5rhgjmROXu2O",
        "colab_type": "text"
      },
      "source": [
        "Let's build a small network with two convolutional layers, followed by one fully connected layer. Convolutional networks are more expensive computationally, so we'll limit its depth and number of fully connected nodes."
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "IZYv70SvvOan",
        "colab_type": "code",
        "colab": {
          "autoexec": {
            "startup": false,
            "wait_interval": 0
          }
        },
        "cellView": "both"
      },
      "source": [
        "batch_size = 16\n",
        "patch_size = 5\n",
        "depth = 16\n",
        "num_hidden = 64\n",
        "\n",
        "graph = tf.Graph()\n",
        "\n",
        "with graph.as_default():\n",
        "\n",
        "  # Input data.\n",
        "  tf_train_dataset = tf.placeholder(\n",
        "    tf.float32, shape=(batch_size, image_size, image_size, num_channels))\n",
        "  tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))\n",
        "  tf_valid_dataset = tf.constant(valid_dataset)\n",
        "  tf_test_dataset = tf.constant(test_dataset)\n",
        "  \n",
        "  # Variables.\n",
        "  layer1_weights = tf.Variable(tf.truncated_normal(\n",
        "      [patch_size, patch_size, num_channels, depth], stddev=0.1))\n",
        "  layer1_biases = tf.Variable(tf.zeros([depth]))\n",
        "  layer2_weights = tf.Variable(tf.truncated_normal(\n",
        "      [patch_size, patch_size, depth, depth], stddev=0.1))\n",
        "  layer2_biases = tf.Variable(tf.constant(1.0, shape=[depth]))\n",
        "  layer3_weights = tf.Variable(tf.truncated_normal(\n",
        "      [image_size // 4 * image_size // 4 * depth, num_hidden], stddev=0.1))\n",
        "  layer3_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))\n",
        "  layer4_weights = tf.Variable(tf.truncated_normal(\n",
        "      [num_hidden, num_labels], stddev=0.1))\n",
        "  layer4_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))\n",
        "  \n",
        "  # Model.\n",
        "  def model(data):\n",
        "    conv = tf.nn.conv2d(data, layer1_weights, [1, 2, 2, 1], padding='SAME')\n",
        "    hidden = tf.nn.relu(conv + layer1_biases)\n",
        "    conv = tf.nn.conv2d(hidden, layer2_weights, [1, 2, 2, 1], padding='SAME')\n",
        "    hidden = tf.nn.relu(conv + layer2_biases)\n",
        "    shape = hidden.get_shape().as_list()\n",
        "    reshape = tf.reshape(hidden, [shape[0], shape[1] * shape[2] * shape[3]])\n",
        "    hidden = tf.nn.relu(tf.matmul(reshape, layer3_weights) + layer3_biases)\n",
        "    return tf.matmul(hidden, layer4_weights) + layer4_biases\n",
        "  \n",
        "  # Training computation.\n",
        "  logits = model(tf_train_dataset)\n",
        "  loss = tf.reduce_mean(\n",
        "    tf.nn.softmax_cross_entropy_with_logits(labels=tf_train_labels, logits=logits))\n",
        "    \n",
        "  # Optimizer.\n",
        "  optimizer = tf.train.GradientDescentOptimizer(0.05).minimize(loss)\n",
        "  \n",
        "  # Predictions for the training, validation, and test data.\n",
        "  train_prediction = tf.nn.softmax(logits)\n",
        "  valid_prediction = tf.nn.softmax(model(tf_valid_dataset))\n",
        "  test_prediction = tf.nn.softmax(model(tf_test_dataset))"
      ],
      "outputs": [],
      "execution_count": 0
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "noKFb2UovVFR",
        "colab_type": "code",
        "colab": {
          "autoexec": {
            "startup": false,
            "wait_interval": 0
          },
          "output_extras": [
            {
              "item_id": 37
            }
          ]
        },
        "cellView": "both",
        "executionInfo": {
          "elapsed": 63292,
          "status": "ok",
          "timestamp": 1446658966251,
          "user": {
            "color": "",
            "displayName": "",
            "isAnonymous": false,
            "isMe": true,
            "permissionId": "",
            "photoUrl": "",
            "sessionId": "0",
            "userId": ""
          },
          "user_tz": 480
        },
        "outputId": "28941338-2ef9-4088-8bd1-44295661e628"
      },
      "source": [
        "num_steps = 1001\n",
        "\n",
        "with tf.Session(graph=graph) as session:\n",
        "  tf.global_variables_initializer().run()\n",
        "  print('Initialized')\n",
        "  for step in range(num_steps):\n",
        "    offset = (step * batch_size) % (train_labels.shape[0] - batch_size)\n",
        "    batch_data = train_dataset[offset:(offset + batch_size), :, :, :]\n",
        "    batch_labels = train_labels[offset:(offset + batch_size), :]\n",
        "    feed_dict = {tf_train_dataset : batch_data, tf_train_labels : batch_labels}\n",
        "    _, l, predictions = session.run(\n",
        "      [optimizer, loss, train_prediction], feed_dict=feed_dict)\n",
        "    if (step % 50 == 0):\n",
        "      print('Minibatch loss at step %d: %f' % (step, l))\n",
        "      print('Minibatch accuracy: %.1f%%' % accuracy(predictions, batch_labels))\n",
        "      print('Validation accuracy: %.1f%%' % accuracy(\n",
        "        valid_prediction.eval(), valid_labels))\n",
        "  print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))"
      ],
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Initialized\n",
            "Minibatch loss at step 0 : 3.51275\n",
            "Minibatch accuracy: 6.2%\n",
            "Validation accuracy: 12.8%\n",
            "Minibatch loss at step 50 : 1.48703\n",
            "Minibatch accuracy: 43.8%\n",
            "Validation accuracy: 50.4%\n",
            "Minibatch loss at step 100 : 1.04377\n",
            "Minibatch accuracy: 68.8%\n",
            "Validation accuracy: 67.4%\n",
            "Minibatch loss at step 150 : 0.601682\n",
            "Minibatch accuracy: 68.8%\n",
            "Validation accuracy: 73.0%\n",
            "Minibatch loss at step 200 : 0.898649\n",
            "Minibatch accuracy: 75.0%\n",
            "Validation accuracy: 77.8%\n",
            "Minibatch loss at step 250 : 1.3637\n",
            "Minibatch accuracy: 56.2%\n",
            "Validation accuracy: 75.4%\n",
            "Minibatch loss at step 300 : 1.41968\n",
            "Minibatch accuracy: 62.5%\n",
            "Validation accuracy: 76.0%\n",
            "Minibatch loss at step 350 : 0.300648\n",
            "Minibatch accuracy: 81.2%\n",
            "Validation accuracy: 80.2%\n",
            "Minibatch loss at step 400 : 1.32092\n",
            "Minibatch accuracy: 56.2%\n",
            "Validation accuracy: 80.4%\n",
            "Minibatch loss at step 450 : 0.556701\n",
            "Minibatch accuracy: 81.2%\n",
            "Validation accuracy: 79.4%\n",
            "Minibatch loss at step 500 : 1.65595\n",
            "Minibatch accuracy: 43.8%\n",
            "Validation accuracy: 79.6%\n",
            "Minibatch loss at step 550 : 1.06995\n",
            "Minibatch accuracy: 75.0%\n",
            "Validation accuracy: 81.2%\n",
            "Minibatch loss at step 600 : 0.223684\n",
            "Minibatch accuracy: 100.0%\n",
            "Validation accuracy: 82.3%\n",
            "Minibatch loss at step 650 : 0.619602\n",
            "Minibatch accuracy: 87.5%\n",
            "Validation accuracy: 81.8%\n",
            "Minibatch loss at step 700 : 0.812091\n",
            "Minibatch accuracy: 75.0%\n",
            "Validation accuracy: 82.4%\n",
            "Minibatch loss at step 750 : 0.276302\n",
            "Minibatch accuracy: 87.5%\n",
            "Validation accuracy: 82.3%\n",
            "Minibatch loss at step 800 : 0.450241\n",
            "Minibatch accuracy: 81.2%\n",
            "Validation accuracy: 82.3%\n",
            "Minibatch loss at step 850 : 0.137139\n",
            "Minibatch accuracy: 93.8%\n",
            "Validation accuracy: 82.3%\n",
            "Minibatch loss at step 900 : 0.52664\n",
            "Minibatch accuracy: 75.0%\n",
            "Validation accuracy: 82.2%\n",
            "Minibatch loss at step 950 : 0.623835\n",
            "Minibatch accuracy: 87.5%\n",
            "Validation accuracy: 82.1%\n",
            "Minibatch loss at step 1000 : 0.243114\n",
            "Minibatch accuracy: 93.8%\n",
            "Validation accuracy: 82.9%\n",
            "Test accuracy: 90.0%\n"
          ],
          "name": "stdout"
        }
      ],
      "execution_count": 0
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KedKkn4EutIK",
        "colab_type": "text"
      },
      "source": [
        "---\n",
        "Problem 1\n",
        "---------\n",
        "\n",
        "The convolutional model above uses convolutions with stride 2 to reduce the dimensionality. Replace the strides by a max pooling operation (`nn.max_pool()`) of stride 2 and kernel size 2.\n",
        "\n",
        "---"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "klf21gpbAgb-",
        "colab_type": "text"
      },
      "source": [
        "---\n",
        "Problem 2\n",
        "---------\n",
        "\n",
        "Try to get the best performance you can using a convolutional net. Look for example at the classic [LeNet5](http://yann.lecun.com/exdb/lenet/) architecture, adding Dropout, and/or adding learning rate decay.\n",
        "\n",
        "---"
      ]
    }
  ]
}