Project__2_empty.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "toc_visible": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "MV4osAJhnrTv"
      },
      "source": [
        "## Definitions"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "6Gqh8kPlgiFJ",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "ff0a2596-78c3-43fb-bfac-5beddc8e00d1"
      },
      "source": [
        "%matplotlib inline\n",
        "import os\n",
        "import numpy as np\n",
        "import pickle\n",
        "import matplotlib.pyplot as plt\n",
        "\n",
        "from tensorflow.keras.utils import to_categorical\n",
        "from tensorflow.keras.callbacks import ReduceLROnPlateau, ModelCheckpoint, EarlyStopping\n",
        "from tensorflow.keras.optimizers import Adadelta, Adam, SGD\n",
        "from tensorflow.keras.layers import Input, Conv2D, Dense, MaxPooling2D, Dropout, Flatten, AveragePooling2D, Conv2DTranspose, UpSampling2D, Rescaling\n",
        "from tensorflow.keras.models import Sequential\n",
        "from tensorflow.keras.losses import categorical_crossentropy\n",
        "\n",
        "from google.colab import drive\n",
        "drive.mount(\"/content/gdrive\", force_remount=True)\n",
        "os.chdir(\"/content/gdrive/My Drive/Colab Notebooks\") # might have to change path to point to your Colab Notebooks folder"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Mounted at /content/gdrive\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "-eJJXcB1n2kn"
      },
      "source": [
        "## Load and format data"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "FgjqOlNDE3tJ",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 387
        },
        "outputId": "453e5ca7-5594-4d30-b5fa-5734f20e39b5"
      },
      "source": [
        "data = np.load('./MNIST_CorrNoise.npz')\n",
        "\n",
        "x_train = data['x_train']\n",
        "y_train = data['y_train']\n",
        "\n",
        "num_cls = len(np.unique(y_train))\n",
        "print('Number of classes: ' + str(num_cls))\n",
        "\n",
        "print('Example of handwritten digit with correlated noise: \\n')\n",
        "\n",
        "k = 3000\n",
        "plt.imshow(np.squeeze(x_train[k,:,:]))\n",
        "plt.show()\n",
        "print('Class: '+str(y_train[k])+'\\n')\n",
        "\n",
        "# RESHAPE and standarize\n",
        "x_train = np.expand_dims(x_train/255,axis=3)\n",
        "\n",
        "# convert class vectors to binary class matrices\n",
        "y_train = to_categorical(y_train, num_cls)\n",
        "\n",
        "print('Shape of x_train: '+str(x_train.shape))\n",
        "print('Shape of y_train: '+str(y_train.shape))"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Number of classes: 10\n",
            "Example of handwritten digit with correlated noise: \n",
            "\n"
          ]
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          }
        },
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Class: 9\n",
            "\n",
            "Shape of x_train: (60000, 28, 28, 1)\n",
            "Shape of y_train: (60000, 10)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "YFPhNWgloVIh"
      },
      "source": [
        "## Training"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "sAE8wRG_pypQ"
      },
      "source": [
        "model_name='CNN' # To compare models, you can give them different names\n",
        "\n",
        "pweight='./weights/weights_' + model_name  + '.hdf5'\n",
        "\n",
        "if not os.path.exists('./weights'):\n",
        "  os.mkdir('./weights')\n",
        "\n",
        "## EXPLORE VALUES AND FIND A GOOD SET\n",
        "b_size = 200 # batch size\n",
        "val_split = 0.75 # percentage of samples used for validation (e.g. 0.5)\n",
        "ep = 20 # number of epochs"
      ],
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "p2Ii3Irf3F0R",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 1000
        },
        "outputId": "8b08dbb0-61fe-4483-8bf9-545c3bede3d4"
      },
      "source": [
        "input_shape = x_train.shape[1:4] #(28,28,1)\n",
        "#  Sequential()\n",
        "# from tensorflow.keras.layers import Input, Conv2D, Dense, MaxPooling2D, Dropout, Flatten, AveragePooling2D, Conv2DTranspose, UpSampling2D, Rescaling\n",
        "import  tensorflow.keras.layers as layers\n",
        "# second model -- start\n",
        "model =Sequential([\n",
        "  layers.Normalization(input_shape=input_shape),\n",
        "  layers.Conv2D(16, 3, padding='same', activation='relu'),\n",
        "  layers.MaxPooling2D(pool_size=(2,2), strides=(2,2),padding='same',),\n",
        "  layers.Conv2D(128, 3, padding='same', activation='relu'),\n",
        "  layers.MaxPooling2D(pool_size=(2,2), strides=(2,2),padding='same',),\n",
        "  # layers.Conv2D(128, 3, padding='same', activation='relu'),\n",
        "  # layers.MaxPooling2D(pool_size=(2,2), strides=(2,2),padding='same',),\n",
        "  layers.Flatten(),\n",
        "  layers.Dense(1024, activation='relu'),\n",
        "  layers.Dropout(0.5),\n",
        "  layers.Dense(num_cls, activation='softmax')\n",
        "])\n",
        "# -- sencond model -- end\n",
        "\n",
        "## EXPLORE LAYERS, # CHANNELS... Conv2D, Dense, MaxPooling2D, Dropout, Flatten, AveragePooling2D, Conv2DTranspose, UpSampling2D\n",
        "## SEE KERAS MANUAL https://keras.io/layers/about-keras-layers/\n",
        "\n",
        "# -- first model -- start\n",
        "# model.add(Conv2D(1, kernel_size=(3, 3), activation='relu', input_shape=input_shape)) # explore num channels and adding more layers\n",
        "\n",
        "# model.add(MaxPooling2D(\n",
        "#     pool_size=(2,2),  # The  down-sampling lattice is 2*2\n",
        "#     strides=(2,2),  # Step length: down to the right\n",
        "#     padding='same', # padding mode is 'same'\n",
        "# ))\n",
        "\n",
        "# model.add(Conv2D(64, (5, 5), padding='same',activation='relu'))\n",
        "# # model.add(Dropout(0.5))\n",
        "# model.add(MaxPooling2D(strides=(2, 2), padding='same'))\n",
        "\n",
        "# # Make multidimensional inputs one-dimensional\n",
        "# model.add(Flatten()) # transforms matrix feature map to vector for dense layer (fully connected)\n",
        "# ##\n",
        "# model.add(Dense(1024,activation='relu'))  # Full connection layer 1024 points\n",
        "# model.add(Dropout(0.5))\n",
        "# model.add(Dense(num_cls, activation='softmax')) # This is a necessary output layer, however, you can add more dense layers before with different activation functions\n",
        "# -- first model -- end\n",
        "lr = 0.001\n",
        "model.compile(loss=categorical_crossentropy,\n",
        "              optimizer=Adam(learning_rate=lr), # explore other optimizers:Adadelta, Adam, SGD\n",
        "              metrics=['accuracy'])\n",
        "model.summary()\n",
        "\n",
        "checkpointer = ModelCheckpoint(filepath=pweight, verbose=1, save_best_only=True)\n",
        "callbacks_list = [checkpointer] # explore adding other callbacks such as ReduceLROnPlateau, EarlyStopping\n",
        "\n",
        "history=model.fit(x_train, y_train,\n",
        "                        epochs=ep,\n",
        "                        batch_size=b_size,\n",
        "                        verbose=1,\n",
        "                        shuffle=True,\n",
        "                        validation_split = val_split,\n",
        "                        callbacks=callbacks_list)\n",
        "\n",
        "print('CNN weights saved in ' + pweight)\n",
        "\n",
        "# Plot loss vs epochs\n",
        "plt.plot(history.history['loss'])\n",
        "plt.plot(history.history['val_loss'])\n",
        "plt.title('model loss')\n",
        "plt.ylabel('loss')\n",
        "plt.xlabel('epoch')\n",
        "plt.legend(['train', 'val'], loc='upper right')\n",
        "plt.show()\n",
        "\n",
        "# for i in history.history:\n",
        "#   print(i)\n",
        "\n",
        "# Plot accuracy vs epochs\n",
        "plt.plot(history.history['accuracy'])\n",
        "plt.plot(history.history['val_accuracy'])\n",
        "plt.title('model accuracy')\n",
        "plt.ylabel('accuracy')\n",
        "plt.xlabel('epoch')\n",
        "plt.legend(['train', 'val'], loc='upper left')\n",
        "plt.show()"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Model: \"sequential_18\"\n",
            "_________________________________________________________________\n",
            " Layer (type)                Output Shape              Param #   \n",
            "=================================================================\n",
            " normalization_5 (Normalizat  (None, 28, 28, 1)        3         \n",
            " ion)                                                            \n",
            "                                                                 \n",
            " conv2d_40 (Conv2D)          (None, 28, 28, 16)        160       \n",
            "                                                                 \n",
            " max_pooling2d_40 (MaxPoolin  (None, 14, 14, 16)       0         \n",
            " g2D)                                                            \n",
            "                                                                 \n",
            " conv2d_41 (Conv2D)          (None, 14, 14, 128)       18560     \n",
            "                                                                 \n",
            " max_pooling2d_41 (MaxPoolin  (None, 7, 7, 128)        0         \n",
            " g2D)                                                            \n",
            "                                                                 \n",
            " flatten_18 (Flatten)        (None, 6272)              0         \n",
            "                                                                 \n",
            " dense_36 (Dense)            (None, 1024)              6423552   \n",
            "                                                                 \n",
            " dropout_10 (Dropout)        (None, 1024)              0         \n",
            "                                                                 \n",
            " dense_37 (Dense)            (None, 10)                10250     \n",
            "                                                                 \n",
            "=================================================================\n",
            "Total params: 6,452,525\n",
            "Trainable params: 6,452,522\n",
            "Non-trainable params: 3\n",
            "_________________________________________________________________\n",
            "Epoch 1/20\n",
            "74/75 [============================>.] - ETA: 0s - loss: 2.1804 - accuracy: 0.2272\n",
            "Epoch 00001: val_loss improved from inf to 1.62957, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 61s 72ms/step - loss: 2.1726 - accuracy: 0.2303 - val_loss: 1.6296 - val_accuracy: 0.4587\n",
            "Epoch 2/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 1.3910 - accuracy: 0.5188\n",
            "Epoch 00002: val_loss improved from 1.62957 to 1.07248, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 4s 58ms/step - loss: 1.3861 - accuracy: 0.5209 - val_loss: 1.0725 - val_accuracy: 0.6538\n",
            "Epoch 3/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.9380 - accuracy: 0.6879\n",
            "Epoch 00003: val_loss improved from 1.07248 to 0.79087, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 4s 59ms/step - loss: 0.9348 - accuracy: 0.6890 - val_loss: 0.7909 - val_accuracy: 0.7372\n",
            "Epoch 4/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.7198 - accuracy: 0.7605\n",
            "Epoch 00004: val_loss improved from 0.79087 to 0.66278, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 3s 44ms/step - loss: 0.7186 - accuracy: 0.7611 - val_loss: 0.6628 - val_accuracy: 0.7820\n",
            "Epoch 5/20\n",
            "75/75 [==============================] - ETA: 0s - loss: 0.5991 - accuracy: 0.7997\n",
            "Epoch 00005: val_loss improved from 0.66278 to 0.61667, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 4s 59ms/step - loss: 0.5991 - accuracy: 0.7997 - val_loss: 0.6167 - val_accuracy: 0.7938\n",
            "Epoch 6/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.5181 - accuracy: 0.8321\n",
            "Epoch 00006: val_loss improved from 0.61667 to 0.56673, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 3s 44ms/step - loss: 0.5173 - accuracy: 0.8322 - val_loss: 0.5667 - val_accuracy: 0.8109\n",
            "Epoch 7/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.4545 - accuracy: 0.8499\n",
            "Epoch 00007: val_loss improved from 0.56673 to 0.55205, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 3s 43ms/step - loss: 0.4542 - accuracy: 0.8501 - val_loss: 0.5520 - val_accuracy: 0.8145\n",
            "Epoch 8/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.3951 - accuracy: 0.8675\n",
            "Epoch 00008: val_loss did not improve from 0.55205\n",
            "75/75 [==============================] - 3s 39ms/step - loss: 0.3929 - accuracy: 0.8681 - val_loss: 0.5635 - val_accuracy: 0.8139\n",
            "Epoch 9/20\n",
            "75/75 [==============================] - ETA: 0s - loss: 0.3466 - accuracy: 0.8853\n",
            "Epoch 00009: val_loss did not improve from 0.55205\n",
            "75/75 [==============================] - 4s 55ms/step - loss: 0.3466 - accuracy: 0.8853 - val_loss: 0.5544 - val_accuracy: 0.8184\n",
            "Epoch 10/20\n",
            "75/75 [==============================] - ETA: 0s - loss: 0.2899 - accuracy: 0.9070\n",
            "Epoch 00010: val_loss improved from 0.55205 to 0.52969, saving model to ./weights/weights_CNN.hdf5\n",
            "75/75 [==============================] - 5s 68ms/step - loss: 0.2899 - accuracy: 0.9070 - val_loss: 0.5297 - val_accuracy: 0.8252\n",
            "Epoch 11/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.2387 - accuracy: 0.9249\n",
            "Epoch 00011: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 3s 39ms/step - loss: 0.2402 - accuracy: 0.9242 - val_loss: 0.5478 - val_accuracy: 0.8241\n",
            "Epoch 12/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.1914 - accuracy: 0.9392\n",
            "Epoch 00012: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 4s 54ms/step - loss: 0.1913 - accuracy: 0.9392 - val_loss: 0.5708 - val_accuracy: 0.8216\n",
            "Epoch 13/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.1618 - accuracy: 0.9497\n",
            "Epoch 00013: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 4s 54ms/step - loss: 0.1621 - accuracy: 0.9496 - val_loss: 0.5731 - val_accuracy: 0.8224\n",
            "Epoch 14/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.1284 - accuracy: 0.9595\n",
            "Epoch 00014: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 4s 54ms/step - loss: 0.1276 - accuracy: 0.9599 - val_loss: 0.6288 - val_accuracy: 0.8185\n",
            "Epoch 15/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.0973 - accuracy: 0.9727\n",
            "Epoch 00015: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 3s 39ms/step - loss: 0.0973 - accuracy: 0.9725 - val_loss: 0.6451 - val_accuracy: 0.8234\n",
            "Epoch 16/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.0784 - accuracy: 0.9775\n",
            "Epoch 00016: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 4s 54ms/step - loss: 0.0785 - accuracy: 0.9775 - val_loss: 0.6556 - val_accuracy: 0.8198\n",
            "Epoch 17/20\n",
            "75/75 [==============================] - ETA: 0s - loss: 0.0687 - accuracy: 0.9815\n",
            "Epoch 00017: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 3s 39ms/step - loss: 0.0687 - accuracy: 0.9815 - val_loss: 0.6783 - val_accuracy: 0.8201\n",
            "Epoch 18/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.0555 - accuracy: 0.9858\n",
            "Epoch 00018: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 4s 54ms/step - loss: 0.0555 - accuracy: 0.9860 - val_loss: 0.6967 - val_accuracy: 0.8185\n",
            "Epoch 19/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.0476 - accuracy: 0.9885\n",
            "Epoch 00019: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 4s 54ms/step - loss: 0.0475 - accuracy: 0.9886 - val_loss: 0.7327 - val_accuracy: 0.8178\n",
            "Epoch 20/20\n",
            "73/75 [============================>.] - ETA: 0s - loss: 0.0439 - accuracy: 0.9889\n",
            "Epoch 00020: val_loss did not improve from 0.52969\n",
            "75/75 [==============================] - 3s 39ms/step - loss: 0.0436 - accuracy: 0.9891 - val_loss: 0.7633 - val_accuracy: 0.8143\n",
            "CNN weights saved in ./weights/weights_CNN.hdf5\n"
          ]
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          }
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "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\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "XK3dbqKloipP"
      },
      "source": [
        "## Make predictions in test set"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "OPLSp4PppXE1",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "7a22210f-6ee7-4b34-8bf2-34a2d4279b2a"
      },
      "source": [
        "from tensorflow.keras.models import load_model\n",
        "\n",
        "## LOAD DATA\n",
        "data = np.load('./MNIST_CorrNoise.npz')\n",
        "\n",
        "x_test = data['x_test']\n",
        "y_test = data['y_test']\n",
        "\n",
        "num_cls = len(np.unique(y_test))\n",
        "print('Number of classes: ' + str(num_cls))\n",
        "\n",
        "# RESHAPE and standarize\n",
        "x_test = np.expand_dims(x_test/255,axis=3)\n",
        "\n",
        "print('Shape of x_train: '+str(x_test.shape)+'\\n')\n",
        "## Define model parameters\n",
        "model_name='CNN' # To compare models, you can give them different names\n",
        "pweight='./weights/weights_' + model_name  + '.hdf5'\n",
        "\n",
        "model = load_model(pweight)\n",
        "\n",
        "proba = model.predict(x_test, use_multiprocessing=True)\n",
        "if proba.shape[-1] > 1:\n",
        "    y_pred = proba.argmax(axis=-1)\n",
        "else:\n",
        "    y_pred = (proba > 0.5).astype('int32')\n",
        "\n",
        "Acc_pred = sum(y_test == y_pred)/len(y_test)\n",
        "print('Accuracy in test set is: '+str(Acc_pred))\n"
      ],
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Number of classes: 10\n",
            "Shape of x_train: (10000, 28, 28, 1)\n",
            "\n",
            "Accuracy in test set is: 0.8182\n"
          ]
        }
      ]
    }
  ]
}