Upload autoencoder.ipynb

autoencoder.ipynb

@@ -0,0 +1,293 @@
+{
+  "nbformat": 4,
+  "nbformat_minor": 0,
+  "metadata": {
+    "colab": {
+      "provenance": []
+    },
+    "kernelspec": {
+      "name": "python3",
+      "display_name": "Python 3"
+    },
+    "language_info": {
+      "name": "python"
+    }
+  },
+  "cells": [
+    {
+      "cell_type": "code",
+      "execution_count": 13,
+      "metadata": {
+        "id": "y9Z3qVyH_sZT"
+      },
+      "outputs": [],
+      "source": [
+        "import numpy as np\n",
+        "import matplotlib.pyplot as plt\n",
+        "from tensorflow import keras\n",
+        "from keras.datasets import fashion_mnist\n",
+        "import keras.backend as K\n",
+        "from keras.layers import Input, Flatten, Dense, Reshape, Lambda, Dropout, BatchNormalization"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()\n",
+        "x_train = x_train / 255\n",
+        "y_train = y_train / 255"
+      ],
+      "metadata": {
+        "id": "epAnpmmLAzCY",
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "outputId": "539c3e52-9946-4a63-8b69-87843e1d6413"
+      },
+      "execution_count": 15,
+      "outputs": [
+        {
+          "output_type": "stream",
+          "name": "stdout",
+          "text": [
+            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz\n",
+            "29515/29515 [==============================] - 0s 0us/step\n",
+            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz\n",
+            "26421880/26421880 [==============================] - 0s 0us/step\n",
+            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz\n",
+            "5148/5148 [==============================] - 0s 0us/step\n",
+            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz\n",
+            "4422102/4422102 [==============================] - 0s 0us/step\n"
+          ]
+        }
+      ]
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "x_train.shape"
+      ],
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "VEJpcZ-wBKkg",
+        "outputId": "ead4dc8c-575b-4210-cb9f-ffe57a5981eb"
+      },
+      "execution_count": 16,
+      "outputs": [
+        {
+          "output_type": "execute_result",
+          "data": {
+            "text/plain": [
+              "(60000, 28, 28)"
+            ]
+          },
+          "metadata": {},
+          "execution_count": 16
+        }
+      ]
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "hidden_dim = 10\n",
+        "batch_size = 32"
+      ],
+      "metadata": {
+        "id": "gAO_-1kyBOCw"
+      },
+      "execution_count": 17,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "input_img = Input((28, 28))\n",
+        "x = Flatten()(input_img)\n",
+        "x = Dense(256, activation = 'relu')(x)\n",
+        "x = Dense(128, activation = 'relu')(x)\n",
+        "\n",
+        "z_mean = Dense(hidden_dim)(x)\n",
+        "z_log_var = Dense(hidden_dim)(x)"
+      ],
+      "metadata": {
+        "id": "1upRqDErBdtS"
+      },
+      "execution_count": 18,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "loss_z_mean, loss_z_log_var = [(), ()]\n",
+        "def foo(args):\n",
+        "  global loss_z_mean, loss_z_log_var\n",
+        "  loss_z_mean, loss_z_log_var = args\n",
+        "\n",
+        "  z_mean, z_log_var = args\n",
+        "  N = K.random_normal(shape = (batch_size, hidden_dim))\n",
+        "  delta = K.exp(z_log_var / 2) * N\n",
+        "  return z_mean + delta"
+      ],
+      "metadata": {
+        "id": "M0Nx30v1DBtq"
+      },
+      "execution_count": 19,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "h = Lambda(foo, output_shape = (hidden_dim,))([z_mean, z_log_var])"
+      ],
+      "metadata": {
+        "id": "fUqgmlf1B-Cz"
+      },
+      "execution_count": 20,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "input_dec = Input(shape = (hidden_dim, ))\n",
+        "d = Dense(128, activation = 'relu')(input_dec)\n",
+        "d = Dense(256, activation = 'relu')(d)\n",
+        "d = Dense(28*28, activation = 'sigmoid')(d)\n",
+        "decoded = Reshape((28, 28))(d)"
+      ],
+      "metadata": {
+        "id": "M8kD92EJCMoW"
+      },
+      "execution_count": 21,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "def loss(y, z):\n",
+        "  y = K.reshape(y, shape = (batch_size, 28*28))\n",
+        "  z = K.reshape(z, shape = (batch_size, 28*28))\n",
+        "  mse = K.sum(K.square(y - z), axis = 1)\n",
+        "  kl = -.5 * K.sum(1 + loss_z_log_var - K.square(loss_z_mean) - K.exp(loss_z_log_var), axis = 1)\n",
+        "\n",
+        "  return mse# + kl"
+      ],
+      "metadata": {
+        "id": "SEzuODKrDtUO"
+      },
+      "execution_count": 22,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "encoder = keras.Model(input_img, h, name = 'encoder')\n",
+        "decoder = keras.Model(input_dec, decoded, name = 'decoder')\n",
+        "vae = keras.Model(input_img, decoder(encoder(input_img)), name = 'vae')"
+      ],
+      "metadata": {
+        "id": "vA_k2uOiCgyV"
+      },
+      "execution_count": 23,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "vae.compile(optimizer = 'adam', loss = loss)"
+      ],
+      "metadata": {
+        "id": "z9foG6KEC7h9"
+      },
+      "execution_count": 24,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "vae.fit(x_train, x_train, epochs = 5, batch_size = batch_size, shuffle = True)"
+      ],
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "NhpBXD8wEctG",
+        "outputId": "fe9d7518-58ce-476f-8792-92a9310e26d7"
+      },
+      "execution_count": 25,
+      "outputs": [
+        {
+          "output_type": "stream",
+          "name": "stdout",
+          "text": [
+            "Epoch 1/5\n",
+            "1875/1875 [==============================] - 24s 12ms/step - loss: 18.1302\n",
+            "Epoch 2/5\n",
+            "1875/1875 [==============================] - 21s 11ms/step - loss: 12.2267\n",
+            "Epoch 3/5\n",
+            "1875/1875 [==============================] - 21s 11ms/step - loss: 11.2004\n",
+            "Epoch 4/5\n",
+            "1875/1875 [==============================] - 21s 11ms/step - loss: 10.6651\n",
+            "Epoch 5/5\n",
+            "1875/1875 [==============================] - 22s 12ms/step - loss: 10.3514\n"
+          ]
+        },
+        {
+          "output_type": "execute_result",
+          "data": {
+            "text/plain": [
+              "<keras.callbacks.History at 0x7f0ca6480fd0>"
+            ]
+          },
+          "metadata": {},
+          "execution_count": 25
+        }
+      ]
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "plt.imshow(vae.predict([x_test[2].reshape(-1, 28, 28, 1)])[0], cmap = 'gray_r')"
+      ],
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/",
+          "height": 465
+        },
+        "id": "ROTQ8mkoE9i6",
+        "outputId": "e2d34505-e8cf-40f5-bba6-b6496bb2e084"
+      },
+      "execution_count": 29,
+      "outputs": [
+        {
+          "output_type": "stream",
+          "name": "stdout",
+          "text": [
+            "1/1 [==============================] - 0s 62ms/step\n"
+          ]
+        },
+        {
+          "output_type": "execute_result",
+          "data": {
+            "text/plain": [
+              "<matplotlib.image.AxesImage at 0x7f0c7cf4ceb0>"
+            ]
+          },
+          "metadata": {},
+          "execution_count": 29
+        },
+        {
+          "output_type": "display_data",
+          "data": {
+            "text/plain": [
+              "<Figure size 640x480 with 1 Axes>"
+            ],
+            "image/png": "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\n"
+          },
+          "metadata": {}
+        }
+      ]
+    }
+  ]
+}