Weights_Visualization_TensorBoard.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "Weights_Visualization_TensorBoard.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "authorship_tag": "ABX9TyPBI879GTuZWo/7sH4ong2g",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/sayakpaul/TF-2.0-Hacks/blob/master/Weights_Visualization_TensorBoard.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "qfrFw0GLjpeW",
        "colab_type": "text"
      },
      "source": [
        "This notebook accompanies with my article [The effects of weight initialization on neural nets](https://www.wandb.com/articles/the-effects-of-weight-initialization-on-neural-nets). "
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "TxmpsWyfQI5y",
        "colab_type": "code",
        "outputId": "de5ee3c0-adc0-4475-c04a-b3b5832581d6",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 51
        }
      },
      "source": [
        "# Selecting the right version of TensorFlow nad verifying it\n",
        "%tensorflow_version 2.x\n",
        "import tensorflow as tf\n",
        "print(tf.__version__)"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "TensorFlow 2.x selected.\n",
            "2.1.0\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "m39BAjMzQgww",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# Weights and Biases for experiment tracking\n",
        "!pip install wandb\n",
        "!wandb login"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LAjuau8YQxlf",
        "colab_type": "code",
        "outputId": "2709edf0-f0b4-49bd-ff3c-ef496d469644",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 153
        }
      },
      "source": [
        "# Gather data and preprocess it\n",
        "(X_train, y_train), (X_test, y_test) = tf.keras.datasets.fashion_mnist.load_data()\n",
        "X_train, X_test = X_train / 255.0, X_test / 255.0"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz\n",
            "32768/29515 [=================================] - 0s 0us/step\n",
            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz\n",
            "26427392/26421880 [==============================] - 0s 0us/step\n",
            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz\n",
            "8192/5148 [===============================================] - 0s 0us/step\n",
            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz\n",
            "4423680/4422102 [==============================] - 0s 0us/step\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "icxgnlO2RJMn",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "CLASS_NAMES = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',\n",
        "               'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "XWKYKwG7RMCz",
        "colab_type": "code",
        "outputId": "9b52a46b-19e9-44cc-9133-ae2c47d9f18c",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 589
        }
      },
      "source": [
        "import matplotlib.pyplot as plt\n",
        "\n",
        "plt.figure(figsize=(10,10))\n",
        "for i in range(25):\n",
        "    plt.subplot(5,5,i+1)\n",
        "    plt.xticks([])\n",
        "    plt.yticks([])\n",
        "    plt.grid(True)\n",
        "    plt.imshow(X_train[i], cmap=plt.cm.binary)\n",
        "    plt.xlabel(CLASS_NAMES[y_train[i]])\n",
        "plt.show()"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAj0AAAI8CAYAAAAazRqkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nQmCC\n",
            "text/plain": [
              "<Figure size 720x720 with 25 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Z42Fz6qzROwI",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "import wandb"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "dvYeX8d_RR2a",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def get_training_model(init_scheme='zeros'):\n",
        "    if isinstance(init_scheme, str):\n",
        "        wandb.init(project='weight-initialization-tb', sync_tensorboard=True,\n",
        "                   id=init_scheme)\n",
        "    else:\n",
        "        wandb.init(project='weight-initialization-tb', sync_tensorboard=True,\n",
        "                   id=str(init_scheme))\n",
        "    \n",
        "    model = tf.keras.models.Sequential([\n",
        "        tf.keras.layers.Flatten(input_shape=(28, 28)),\n",
        "        tf.keras.layers.Dense(256, activation='relu', kernel_initializer=init_scheme,\n",
        "            bias_initializer='zeros'),\n",
        "        tf.keras.layers.Dropout(0.2),\n",
        "        tf.keras.layers.Dense(128, activation='relu', kernel_initializer=init_scheme,\n",
        "            bias_initializer='zeros'),\n",
        "        tf.keras.layers.Dropout(0.2),\n",
        "        tf.keras.layers.Dense(10, activation='softmax', kernel_initializer=init_scheme,\n",
        "            bias_initializer='zeros')\n",
        "    ])\n",
        "\n",
        "    model.compile(optimizer='adam',\n",
        "                loss='sparse_categorical_crossentropy',\n",
        "                metrics=['accuracy'])\n",
        "    \n",
        "    return model, wandb.run.dir"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "BxFeU1L1RV6S",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "from wandb.keras import WandbCallback"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "1Q_PQE-6RYl1",
        "colab_type": "code",
        "outputId": "8f3f124b-42c7-4db9-e459-a27e3539ca34",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 819
        }
      },
      "source": [
        "model_w_zeros, run_dir = get_training_model()\n",
        "\n",
        "tb_callback = tf.keras.callbacks.TensorBoard(log_dir=run_dir, histogram_freq=4, write_images=True)\n",
        "\n",
        "model_w_zeros.fit(X_train, y_train, \n",
        "    validation_data=(X_test, y_test), \n",
        "    epochs=20, batch_size=128,\n",
        "    callbacks=[WandbCallback(data_type='image', labels=CLASS_NAMES, \n",
        "                             validation_data=(X_test, y_test)),\n",
        "               tb_callback])"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/html": [
              "\n",
              "                Logging results to <a href=\"https://wandb.com\" target=\"_blank\">Weights & Biases</a> <a href=\"https://docs.wandb.com/integrations/jupyter.html\" target=\"_blank\">(Documentation)</a>.<br/>\n",
              "                Project page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb</a><br/>\n",
              "                Run page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/zeros\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/zeros</a><br/>\n",
              "            "
            ],
            "text/plain": [
              "<IPython.core.display.HTML object>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "stream",
          "text": [
            "Train on 60000 samples, validate on 10000 samples\n",
            "Epoch 1/20\n",
            "  128/60000 [..............................] - ETA: 19:27 - loss: 2.3026 - accuracy: 0.0859WARNING:tensorflow:Method (on_train_batch_end) is slow compared to the batch update (0.153470). Check your callbacks.\n",
            "60000/60000 [==============================] - 5s 90us/sample - loss: 2.3027 - accuracy: 0.0989 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 2/20\n",
            "60000/60000 [==============================] - 3s 46us/sample - loss: 2.3027 - accuracy: 0.0961 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 3/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 2.3027 - accuracy: 0.0990 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 4/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 2.3027 - accuracy: 0.0992 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 5/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.3027 - accuracy: 0.0975 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 6/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 2.3027 - accuracy: 0.0977 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 7/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.3027 - accuracy: 0.0979 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 8/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 2.3027 - accuracy: 0.0997 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 9/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 2.3027 - accuracy: 0.0983 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 10/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 2.3027 - accuracy: 0.0965 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 11/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 2.3027 - accuracy: 0.0976 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 12/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.3027 - accuracy: 0.0987 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 13/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.3027 - accuracy: 0.0979 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 14/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 2.3027 - accuracy: 0.0983 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 15/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 2.3027 - accuracy: 0.0978 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 16/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 2.3027 - accuracy: 0.0984 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 17/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.3027 - accuracy: 0.0983 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 18/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 2.3027 - accuracy: 0.0984 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 19/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 2.3027 - accuracy: 0.1001 - val_loss: 2.3026 - val_accuracy: 0.1000\n",
            "Epoch 20/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.3027 - accuracy: 0.0989 - val_loss: 2.3026 - val_accuracy: 0.1000\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tensorflow.python.keras.callbacks.History at 0x7fb7d41ca358>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 9
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "AnhChzXgTjxC",
        "colab_type": "code",
        "outputId": "40c8cdb0-6ce7-44aa-a931-037232550fea",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 782
        }
      },
      "source": [
        "model_w_ones, run_dir = get_training_model(init_scheme='ones')\n",
        "\n",
        "tb_callback = tf.keras.callbacks.TensorBoard(log_dir=run_dir, histogram_freq=4, write_images=True)\n",
        "\n",
        "model_w_ones.fit(X_train, y_train, \n",
        "    validation_data=(X_test, y_test), \n",
        "    epochs=20, batch_size=128,\n",
        "    callbacks=[WandbCallback(data_type='image', labels=CLASS_NAMES, \n",
        "                             validation_data=(X_test, y_test)),\n",
        "               tb_callback])"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/html": [
              "\n",
              "                Logging results to <a href=\"https://wandb.com\" target=\"_blank\">Weights & Biases</a> <a href=\"https://docs.wandb.com/integrations/jupyter.html\" target=\"_blank\">(Documentation)</a>.<br/>\n",
              "                Project page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb</a><br/>\n",
              "                Run page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/ones\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/ones</a><br/>\n",
              "            "
            ],
            "text/plain": [
              "<IPython.core.display.HTML object>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "stream",
          "text": [
            "Train on 60000 samples, validate on 10000 samples\n",
            "Epoch 1/20\n",
            "60000/60000 [==============================] - 3s 51us/sample - loss: 2229.0527 - accuracy: 0.1007 - val_loss: 2100.3664 - val_accuracy: 0.1000\n",
            "Epoch 2/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 805.0972 - accuracy: 0.1002 - val_loss: 193.1670 - val_accuracy: 0.1000\n",
            "Epoch 3/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 262.6274 - accuracy: 0.0984 - val_loss: 182.9957 - val_accuracy: 0.1000\n",
            "Epoch 4/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 73.1662 - accuracy: 0.0994 - val_loss: 23.1904 - val_accuracy: 0.0971\n",
            "Epoch 5/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 10.1444 - accuracy: 0.1081 - val_loss: 3.3555 - val_accuracy: 0.1149\n",
            "Epoch 6/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 2.7196 - accuracy: 0.1383 - val_loss: 2.2665 - val_accuracy: 0.1499\n",
            "Epoch 7/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.4258 - accuracy: 0.1484 - val_loss: 2.2400 - val_accuracy: 0.1645\n",
            "Epoch 8/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 2.2450 - accuracy: 0.1588 - val_loss: 2.1060 - val_accuracy: 0.2324\n",
            "Epoch 9/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 2.1087 - accuracy: 0.2262 - val_loss: 2.0370 - val_accuracy: 0.2232\n",
            "Epoch 10/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 2.0512 - accuracy: 0.2228 - val_loss: 2.0122 - val_accuracy: 0.2448\n",
            "Epoch 11/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 1.9997 - accuracy: 0.2335 - val_loss: 1.9403 - val_accuracy: 0.2668\n",
            "Epoch 12/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 1.9307 - accuracy: 0.2739 - val_loss: 1.8894 - val_accuracy: 0.2892\n",
            "Epoch 13/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 1.8607 - accuracy: 0.3009 - val_loss: 1.8177 - val_accuracy: 0.3147\n",
            "Epoch 14/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 1.7852 - accuracy: 0.3181 - val_loss: 1.7507 - val_accuracy: 0.3321\n",
            "Epoch 15/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 1.7172 - accuracy: 0.3302 - val_loss: 1.6921 - val_accuracy: 0.3297\n",
            "Epoch 16/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 1.6624 - accuracy: 0.3453 - val_loss: 1.6418 - val_accuracy: 0.3398\n",
            "Epoch 17/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 1.6161 - accuracy: 0.3647 - val_loss: 1.6059 - val_accuracy: 0.3707\n",
            "Epoch 18/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 1.5729 - accuracy: 0.4020 - val_loss: 1.5562 - val_accuracy: 0.4360\n",
            "Epoch 19/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 1.4061 - accuracy: 0.4997 - val_loss: 1.1115 - val_accuracy: 0.5946\n",
            "Epoch 20/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 1.0171 - accuracy: 0.6227 - val_loss: 0.9393 - val_accuracy: 0.6414\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tensorflow.python.keras.callbacks.History at 0x7fb7d4309860>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 10
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "BeH4SsViTqPI",
        "colab_type": "code",
        "outputId": "66abba04-b27e-472d-e879-0d9b66169df9",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 782
        }
      },
      "source": [
        "model_w_uniform, run_dir = get_training_model(init_scheme=tf.keras.initializers.RandomUniform(minval=0, maxval=1, seed=666))\n",
        "\n",
        "tb_callback = tf.keras.callbacks.TensorBoard(log_dir=run_dir, histogram_freq=4, write_images=True)\n",
        "\n",
        "model_w_uniform.fit(X_train, y_train, \n",
        "    validation_data=(X_test, y_test), \n",
        "    epochs=20, batch_size=128,\n",
        "    callbacks=[WandbCallback(data_type='image', labels=CLASS_NAMES, \n",
        "                             validation_data=(X_test, y_test)),\n",
        "               tb_callback])"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/html": [
              "\n",
              "                Logging results to <a href=\"https://wandb.com\" target=\"_blank\">Weights & Biases</a> <a href=\"https://docs.wandb.com/integrations/jupyter.html\" target=\"_blank\">(Documentation)</a>.<br/>\n",
              "                Project page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb</a><br/>\n",
              "                Run page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/%3Ctensorflow.python.ops.init_ops_v2.RandomUniform+object+at+0x7fb7c018a3c8%3E\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/%3Ctensorflow.python.ops.init_ops_v2.RandomUniform+object+at+0x7fb7c018a3c8%3E</a><br/>\n",
              "            "
            ],
            "text/plain": [
              "<IPython.core.display.HTML object>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "stream",
          "text": [
            "Train on 60000 samples, validate on 10000 samples\n",
            "Epoch 1/20\n",
            "60000/60000 [==============================] - 3s 48us/sample - loss: 21325.0095 - accuracy: 0.1001 - val_loss: 665.8834 - val_accuracy: 0.2236\n",
            "Epoch 2/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 1303.5034 - accuracy: 0.2453 - val_loss: 3.5182 - val_accuracy: 0.4493\n",
            "Epoch 3/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 3.8314 - accuracy: 0.3570 - val_loss: 1.9515 - val_accuracy: 0.4465\n",
            "Epoch 4/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 2.3605 - accuracy: 0.3988 - val_loss: 1.6066 - val_accuracy: 0.4830\n",
            "Epoch 5/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 1.9021 - accuracy: 0.4436 - val_loss: 1.4260 - val_accuracy: 0.5482\n",
            "Epoch 6/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 1.6767 - accuracy: 0.4931 - val_loss: 1.3005 - val_accuracy: 0.5831\n",
            "Epoch 7/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 1.5313 - accuracy: 0.5255 - val_loss: 1.1737 - val_accuracy: 0.6185\n",
            "Epoch 8/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 1.4014 - accuracy: 0.5566 - val_loss: 1.0859 - val_accuracy: 0.6422\n",
            "Epoch 9/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 1.3203 - accuracy: 0.5743 - val_loss: 1.0299 - val_accuracy: 0.6611\n",
            "Epoch 10/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 1.2391 - accuracy: 0.5918 - val_loss: 0.9859 - val_accuracy: 0.6664\n",
            "Epoch 11/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 1.1887 - accuracy: 0.6037 - val_loss: 0.9357 - val_accuracy: 0.6805\n",
            "Epoch 12/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 1.1346 - accuracy: 0.6151 - val_loss: 0.9069 - val_accuracy: 0.6825\n",
            "Epoch 13/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 1.0940 - accuracy: 0.6225 - val_loss: 0.8799 - val_accuracy: 0.6854\n",
            "Epoch 14/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 1.0529 - accuracy: 0.6340 - val_loss: 0.8582 - val_accuracy: 0.7022\n",
            "Epoch 15/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 1.0173 - accuracy: 0.6422 - val_loss: 0.8372 - val_accuracy: 0.6966\n",
            "Epoch 16/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 0.9815 - accuracy: 0.6488 - val_loss: 0.8120 - val_accuracy: 0.7085\n",
            "Epoch 17/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.9524 - accuracy: 0.6563 - val_loss: 0.7886 - val_accuracy: 0.7082\n",
            "Epoch 18/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 0.9291 - accuracy: 0.6630 - val_loss: 0.7767 - val_accuracy: 0.7112\n",
            "Epoch 19/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 0.9006 - accuracy: 0.6690 - val_loss: 0.7590 - val_accuracy: 0.7200\n",
            "Epoch 20/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 0.8729 - accuracy: 0.6757 - val_loss: 0.7514 - val_accuracy: 0.7239\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tensorflow.python.keras.callbacks.History at 0x7fb7c0209048>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 11
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "kW70RwlvUWfz",
        "colab_type": "code",
        "outputId": "2f85d88a-7a8e-4e05-fac1-997c17c105f6",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 782
        }
      },
      "source": [
        "model_w_centered, run_dir = get_training_model(init_scheme=tf.keras.initializers.RandomUniform(minval=-0.5, maxval=0.5, seed=666))\n",
        "\n",
        "tb_callback = tf.keras.callbacks.TensorBoard(log_dir=run_dir, histogram_freq=4, write_images=True)\n",
        "\n",
        "model_w_centered.fit(X_train, y_train, \n",
        "    validation_data=(X_test, y_test), \n",
        "    epochs=20, batch_size=128,\n",
        "    callbacks=[WandbCallback(data_type='image', labels=CLASS_NAMES, \n",
        "                             validation_data=(X_test, y_test)),\n",
        "               tb_callback])"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/html": [
              "\n",
              "                Logging results to <a href=\"https://wandb.com\" target=\"_blank\">Weights & Biases</a> <a href=\"https://docs.wandb.com/integrations/jupyter.html\" target=\"_blank\">(Documentation)</a>.<br/>\n",
              "                Project page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb</a><br/>\n",
              "                Run page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/%3Ctensorflow.python.ops.init_ops_v2.RandomUniform+object+at+0x7fb7c0089198%3E\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/%3Ctensorflow.python.ops.init_ops_v2.RandomUniform+object+at+0x7fb7c0089198%3E</a><br/>\n",
              "            "
            ],
            "text/plain": [
              "<IPython.core.display.HTML object>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "stream",
          "text": [
            "Train on 60000 samples, validate on 10000 samples\n",
            "Epoch 1/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 2.7490 - accuracy: 0.6333 - val_loss: 0.6562 - val_accuracy: 0.7626\n",
            "Epoch 2/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 0.7808 - accuracy: 0.7416 - val_loss: 0.5450 - val_accuracy: 0.8023\n",
            "Epoch 3/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 0.6312 - accuracy: 0.7782 - val_loss: 0.4916 - val_accuracy: 0.8186\n",
            "Epoch 4/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 0.5515 - accuracy: 0.8013 - val_loss: 0.4644 - val_accuracy: 0.8301\n",
            "Epoch 5/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 0.5014 - accuracy: 0.8180 - val_loss: 0.4341 - val_accuracy: 0.8386\n",
            "Epoch 6/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.4678 - accuracy: 0.8292 - val_loss: 0.4218 - val_accuracy: 0.8476\n",
            "Epoch 7/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 0.4453 - accuracy: 0.8396 - val_loss: 0.4069 - val_accuracy: 0.8492\n",
            "Epoch 8/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 0.4224 - accuracy: 0.8456 - val_loss: 0.3955 - val_accuracy: 0.8569\n",
            "Epoch 9/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 0.4045 - accuracy: 0.8516 - val_loss: 0.3915 - val_accuracy: 0.8512\n",
            "Epoch 10/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 0.3883 - accuracy: 0.8565 - val_loss: 0.3815 - val_accuracy: 0.8578\n",
            "Epoch 11/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 0.3759 - accuracy: 0.8599 - val_loss: 0.3740 - val_accuracy: 0.8658\n",
            "Epoch 12/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 0.3599 - accuracy: 0.8669 - val_loss: 0.3691 - val_accuracy: 0.8665\n",
            "Epoch 13/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 0.3498 - accuracy: 0.8713 - val_loss: 0.3715 - val_accuracy: 0.8651\n",
            "Epoch 14/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 0.3393 - accuracy: 0.8752 - val_loss: 0.3653 - val_accuracy: 0.8640\n",
            "Epoch 15/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 0.3331 - accuracy: 0.8756 - val_loss: 0.3570 - val_accuracy: 0.8701\n",
            "Epoch 16/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.3246 - accuracy: 0.8788 - val_loss: 0.3598 - val_accuracy: 0.8717\n",
            "Epoch 17/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.3155 - accuracy: 0.8818 - val_loss: 0.3492 - val_accuracy: 0.8733\n",
            "Epoch 18/20\n",
            "60000/60000 [==============================] - 2s 40us/sample - loss: 0.3096 - accuracy: 0.8838 - val_loss: 0.3441 - val_accuracy: 0.8763\n",
            "Epoch 19/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 0.3021 - accuracy: 0.8859 - val_loss: 0.3435 - val_accuracy: 0.8768\n",
            "Epoch 20/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 0.2930 - accuracy: 0.8898 - val_loss: 0.3431 - val_accuracy: 0.8751\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tensorflow.python.keras.callbacks.History at 0x7fb7a83b4a90>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 12
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "_Mf4Rf87U7wg",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "def get_training_model_w_rule(init_scheme):\n",
        "    if isinstance(init_scheme, str):\n",
        "        wandb.init(project='weight-initialization-tb', sync_tensorboard=True,\n",
        "                   id=init_scheme)\n",
        "    else:\n",
        "        wandb.init(project='weight-initialization-tb', sync_tensorboard=True,\n",
        "                   id=str(init_scheme))\n",
        "    \n",
        "    model = tf.keras.models.Sequential([\n",
        "        tf.keras.layers.Flatten(input_shape=(28, 28)),\n",
        "        tf.keras.layers.Dense(256, activation='relu'),\n",
        "        tf.keras.layers.Dropout(0.2),\n",
        "        tf.keras.layers.Dense(128, activation='relu'),\n",
        "        tf.keras.layers.Dropout(0.2),\n",
        "        tf.keras.layers.Dense(10, activation='softmax')\n",
        "    ])\n",
        "\n",
        "    model.compile(optimizer='adam',\n",
        "                loss='sparse_categorical_crossentropy',\n",
        "                metrics=['accuracy'])\n",
        "    \n",
        "    return model, wandb.run.dir"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "W6Nv-zA2Viul",
        "colab_type": "code",
        "outputId": "b7a05738-a950-45df-a086-c822c8f6f2c9",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 782
        }
      },
      "source": [
        "import numpy as np\n",
        "\n",
        "model_w_uniform_rule, run_dir = get_training_model_w_rule('uniform-rule')\n",
        "\n",
        "tb_callback = tf.keras.callbacks.TensorBoard(log_dir=run_dir, histogram_freq=4, write_images=True)\n",
        "\n",
        "for layer in model_w_uniform_rule.layers:\n",
        "    if isinstance(layer, tf.keras.layers.Dense):\n",
        "        shape = (layer.weights[0].shape[0], layer.weights[0].shape[1])\n",
        "        y = 1.0/np.sqrt(shape[0])\n",
        "        rule_weights = np.random.uniform(-y, y, shape)\n",
        "        layer.weights[0] = rule_weights\n",
        "        layer.weights[1] = 0\n",
        "\n",
        "model_w_uniform_rule.compile(optimizer='adam',\n",
        "                loss='sparse_categorical_crossentropy',\n",
        "                metrics=['accuracy'])\n",
        "\n",
        "model_w_uniform_rule.fit(X_train, y_train, \n",
        "    validation_data=(X_test, y_test), \n",
        "    epochs=20, batch_size=128,\n",
        "    callbacks=[WandbCallback(data_type='image', labels=CLASS_NAMES, \n",
        "                             validation_data=(X_test, y_test)),\n",
        "               tb_callback])"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/html": [
              "\n",
              "                Logging results to <a href=\"https://wandb.com\" target=\"_blank\">Weights & Biases</a> <a href=\"https://docs.wandb.com/integrations/jupyter.html\" target=\"_blank\">(Documentation)</a>.<br/>\n",
              "                Project page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb</a><br/>\n",
              "                Run page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/uniform-rule\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/uniform-rule</a><br/>\n",
              "            "
            ],
            "text/plain": [
              "<IPython.core.display.HTML object>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "stream",
          "text": [
            "Train on 60000 samples, validate on 10000 samples\n",
            "Epoch 1/20\n",
            "60000/60000 [==============================] - 3s 54us/sample - loss: 0.5833 - accuracy: 0.7910 - val_loss: 0.4314 - val_accuracy: 0.8425\n",
            "Epoch 2/20\n",
            "60000/60000 [==============================] - 3s 45us/sample - loss: 0.4082 - accuracy: 0.8526 - val_loss: 0.3905 - val_accuracy: 0.8590\n",
            "Epoch 3/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 0.3686 - accuracy: 0.8658 - val_loss: 0.3806 - val_accuracy: 0.8617\n",
            "Epoch 4/20\n",
            "60000/60000 [==============================] - 3s 46us/sample - loss: 0.3459 - accuracy: 0.8739 - val_loss: 0.3625 - val_accuracy: 0.8664\n",
            "Epoch 5/20\n",
            "60000/60000 [==============================] - 3s 46us/sample - loss: 0.3274 - accuracy: 0.8795 - val_loss: 0.3483 - val_accuracy: 0.8735\n",
            "Epoch 6/20\n",
            "60000/60000 [==============================] - 3s 45us/sample - loss: 0.3159 - accuracy: 0.8837 - val_loss: 0.3474 - val_accuracy: 0.8763\n",
            "Epoch 7/20\n",
            "60000/60000 [==============================] - 3s 46us/sample - loss: 0.3044 - accuracy: 0.8894 - val_loss: 0.3304 - val_accuracy: 0.8809\n",
            "Epoch 8/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2938 - accuracy: 0.8898 - val_loss: 0.3374 - val_accuracy: 0.8782\n",
            "Epoch 9/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2869 - accuracy: 0.8942 - val_loss: 0.3306 - val_accuracy: 0.8802\n",
            "Epoch 10/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2786 - accuracy: 0.8960 - val_loss: 0.3251 - val_accuracy: 0.8827\n",
            "Epoch 11/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.2688 - accuracy: 0.8999 - val_loss: 0.3241 - val_accuracy: 0.8828\n",
            "Epoch 12/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 0.2642 - accuracy: 0.9006 - val_loss: 0.3298 - val_accuracy: 0.8814\n",
            "Epoch 13/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2585 - accuracy: 0.9026 - val_loss: 0.3176 - val_accuracy: 0.8853\n",
            "Epoch 14/20\n",
            "60000/60000 [==============================] - 2s 42us/sample - loss: 0.2532 - accuracy: 0.9047 - val_loss: 0.3348 - val_accuracy: 0.8768\n",
            "Epoch 15/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.2504 - accuracy: 0.9057 - val_loss: 0.3248 - val_accuracy: 0.8863\n",
            "Epoch 16/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.2446 - accuracy: 0.9082 - val_loss: 0.3064 - val_accuracy: 0.8915\n",
            "Epoch 17/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2393 - accuracy: 0.9096 - val_loss: 0.3118 - val_accuracy: 0.8909\n",
            "Epoch 18/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2346 - accuracy: 0.9103 - val_loss: 0.3248 - val_accuracy: 0.8885\n",
            "Epoch 19/20\n",
            "60000/60000 [==============================] - 2s 38us/sample - loss: 0.2307 - accuracy: 0.9130 - val_loss: 0.3244 - val_accuracy: 0.8871\n",
            "Epoch 20/20\n",
            "60000/60000 [==============================] - 2s 39us/sample - loss: 0.2237 - accuracy: 0.9154 - val_loss: 0.3163 - val_accuracy: 0.8913\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tensorflow.python.keras.callbacks.History at 0x7fb7a822f5c0>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 14
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "ERcQ8vBBWYDb",
        "colab_type": "code",
        "outputId": "13aa9b2d-6a47-4048-aab4-08217c9a4660",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 782
        }
      },
      "source": [
        "import numpy as np\n",
        "\n",
        "model_w_normal_rule, run_dir = get_training_model_w_rule('normal-rule')\n",
        "\n",
        "tb_callback = tf.keras.callbacks.TensorBoard(log_dir=run_dir, histogram_freq=4, write_images=True)\n",
        "\n",
        "for layer in model_w_normal_rule.layers:\n",
        "    if isinstance(layer, tf.keras.layers.Dense):\n",
        "        shape = (layer.weights[0].shape[0], layer.weights[0].shape[1])\n",
        "        y = 1.0/np.sqrt(shape[0])\n",
        "        rule_weights = np.random.uniform(-y, y, shape)\n",
        "        layer.weights[0] = rule_weights\n",
        "        layer.weights[1] = 0\n",
        "\n",
        "model_w_normal_rule.compile(optimizer='adam',\n",
        "                loss='sparse_categorical_crossentropy',\n",
        "                metrics=['accuracy'])\n",
        "\n",
        "model_w_normal_rule.fit(X_train, y_train, \n",
        "    validation_data=(X_test, y_test), \n",
        "    epochs=20, batch_size=128,\n",
        "    callbacks=[WandbCallback(data_type='image', labels=CLASS_NAMES, \n",
        "                             validation_data=(X_test, y_test)),\n",
        "               tb_callback])"
      ],
      "execution_count": 0,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/html": [
              "\n",
              "                Logging results to <a href=\"https://wandb.com\" target=\"_blank\">Weights & Biases</a> <a href=\"https://docs.wandb.com/integrations/jupyter.html\" target=\"_blank\">(Documentation)</a>.<br/>\n",
              "                Project page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb</a><br/>\n",
              "                Run page: <a href=\"https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/normal-rule\" target=\"_blank\">https://app.wandb.ai/sayakpaul/weight-initialization-tb/runs/normal-rule</a><br/>\n",
              "            "
            ],
            "text/plain": [
              "<IPython.core.display.HTML object>"
            ]
          },
          "metadata": {
            "tags": []
          }
        },
        {
          "output_type": "stream",
          "text": [
            "Train on 60000 samples, validate on 10000 samples\n",
            "Epoch 1/20\n",
            "60000/60000 [==============================] - 3s 51us/sample - loss: 0.5845 - accuracy: 0.7929 - val_loss: 0.4272 - val_accuracy: 0.8436\n",
            "Epoch 2/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 0.4073 - accuracy: 0.8529 - val_loss: 0.3816 - val_accuracy: 0.8631\n",
            "Epoch 3/20\n",
            "60000/60000 [==============================] - 3s 45us/sample - loss: 0.3710 - accuracy: 0.8653 - val_loss: 0.3747 - val_accuracy: 0.8670\n",
            "Epoch 4/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 0.3469 - accuracy: 0.8742 - val_loss: 0.3540 - val_accuracy: 0.8732\n",
            "Epoch 5/20\n",
            "60000/60000 [==============================] - 3s 45us/sample - loss: 0.3291 - accuracy: 0.8800 - val_loss: 0.3509 - val_accuracy: 0.8740\n",
            "Epoch 6/20\n",
            "60000/60000 [==============================] - 3s 45us/sample - loss: 0.3148 - accuracy: 0.8840 - val_loss: 0.3487 - val_accuracy: 0.8711\n",
            "Epoch 7/20\n",
            "60000/60000 [==============================] - 3s 45us/sample - loss: 0.3063 - accuracy: 0.8857 - val_loss: 0.3348 - val_accuracy: 0.8781\n",
            "Epoch 8/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 0.2935 - accuracy: 0.8915 - val_loss: 0.3298 - val_accuracy: 0.8791\n",
            "Epoch 9/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 0.2856 - accuracy: 0.8944 - val_loss: 0.3210 - val_accuracy: 0.8851\n",
            "Epoch 10/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.2794 - accuracy: 0.8957 - val_loss: 0.3268 - val_accuracy: 0.8809\n",
            "Epoch 11/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 0.2707 - accuracy: 0.8991 - val_loss: 0.3241 - val_accuracy: 0.8846\n",
            "Epoch 12/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 0.2650 - accuracy: 0.9008 - val_loss: 0.3289 - val_accuracy: 0.8811\n",
            "Epoch 13/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2593 - accuracy: 0.9024 - val_loss: 0.3194 - val_accuracy: 0.8863\n",
            "Epoch 14/20\n",
            "60000/60000 [==============================] - 3s 43us/sample - loss: 0.2512 - accuracy: 0.9056 - val_loss: 0.3106 - val_accuracy: 0.8842\n",
            "Epoch 15/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 0.2477 - accuracy: 0.9060 - val_loss: 0.3161 - val_accuracy: 0.8907\n",
            "Epoch 16/20\n",
            "60000/60000 [==============================] - 3s 42us/sample - loss: 0.2420 - accuracy: 0.9081 - val_loss: 0.3106 - val_accuracy: 0.8899\n",
            "Epoch 17/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.2400 - accuracy: 0.9086 - val_loss: 0.3164 - val_accuracy: 0.8882\n",
            "Epoch 18/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.2346 - accuracy: 0.9111 - val_loss: 0.3138 - val_accuracy: 0.8906\n",
            "Epoch 19/20\n",
            "60000/60000 [==============================] - 2s 41us/sample - loss: 0.2323 - accuracy: 0.9120 - val_loss: 0.3123 - val_accuracy: 0.8899\n",
            "Epoch 20/20\n",
            "60000/60000 [==============================] - 3s 44us/sample - loss: 0.2257 - accuracy: 0.9132 - val_loss: 0.3060 - val_accuracy: 0.8907\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tensorflow.python.keras.callbacks.History at 0x7fb7a808ce48>"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 15
        }
      ]
    }
  ]
}