Train_license_character_recognition.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "GxW3Qg-ScurV"
   },
   "source": [
    "#### Train a character recongition from input image using MobileNets Network."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 35
    },
    "colab_type": "code",
    "id": "XJmrQrB_curX",
    "outputId": "5a4bee5b-8c58-41bb-cd70-ff9ffa406ff9"
   },
   "outputs": [],
   "source": [
    "import os\n",
    "os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\n",
    "\n",
    "from keras.preprocessing.image import ImageDataGenerator\n",
    "from keras.applications import MobileNetV2\n",
    "from keras.layers import AveragePooling2D\n",
    "from keras.layers import Dropout\n",
    "from keras.layers import Flatten\n",
    "from keras.layers import Dense\n",
    "from keras.layers import Input\n",
    "from keras.models import Model\n",
    "from keras.optimizers import Adam\n",
    "from keras.preprocessing.image import img_to_array\n",
    "from keras.preprocessing.image import load_img\n",
    "from keras.utils import to_categorical\n",
    "from keras.callbacks import ModelCheckpoint, EarlyStopping\n",
    "from keras.models import model_from_json\n",
    "from sklearn.preprocessing import LabelEncoder\n",
    "from sklearn.model_selection import train_test_split\n",
    "import matplotlib.pyplot as plt\n",
    "import matplotlib.gridspec as gridspec\n",
    "import glob\n",
    "import numpy as np"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "z8ApWHt8curj"
   },
   "source": [
    "### Visualize dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 479
    },
    "colab_type": "code",
    "id": "-y1mnjyPcurk",
    "outputId": "de256a8f-5bb5-47c7-91c5-b29832539766",
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x576 with 12 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "dataset_paths = glob.glob(\"dataset_characters/**/*.jpg\")\n",
    "\n",
    "cols=4\n",
    "rows=3\n",
    "fig = plt.figure(figsize=(10,8))\n",
    "plt.rcParams.update({\"font.size\":14})\n",
    "grid = gridspec.GridSpec(ncols=cols,nrows=rows,figure=fig)\n",
    "\n",
    "# create a random list of images will be displayed\n",
    "np.random.seed(45)\n",
    "rand = np.random.randint(0,len(dataset_paths),size=(cols*rows))\n",
    "\n",
    "# Plot image\n",
    "for i in range(cols*rows):\n",
    "    fig.add_subplot(grid[i])\n",
    "    image = load_img(dataset_paths[rand[i]])\n",
    "    label = dataset_paths[rand[i]].split(os.path.sep)[-2]\n",
    "    plt.title('\"{:s}\"'.format(label))\n",
    "    plt.axis(False)\n",
    "    plt.imshow(image)\n",
    "\n",
    "plt.savefig(\"Visualize_dataset.jpg\",dpi=300)    \n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "LJzYBWt3bzAj"
   },
   "source": [
    "## Data processing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 35
    },
    "colab_type": "code",
    "id": "Yh443FGFRgFP",
    "outputId": "2fdc22f4-e393-4ae1-bb86-d5fb55ed1865"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[INFO] Find 37623 images with 36 classes\n"
     ]
    }
   ],
   "source": [
    "# Arange input data and corresponding labels\n",
    "X=[]\n",
    "labels=[]\n",
    "\n",
    "for image_path in dataset_paths:\n",
    "  label = image_path.split(os.path.sep)[-2]\n",
    "  image=load_img(image_path,target_size=(80,80))\n",
    "  image=img_to_array(image)\n",
    "\n",
    "  X.append(image)\n",
    "  labels.append(label)\n",
    "\n",
    "X = np.array(X,dtype=\"float16\")\n",
    "labels = np.array(labels)\n",
    "\n",
    "print(\"[INFO] Find {:d} images with {:d} classes\".format(len(X),len(set(labels))))\n",
    "\n",
    "\n",
    "# perform one-hot encoding on the labels\n",
    "lb = LabelEncoder()\n",
    "lb.fit(labels)\n",
    "labels = lb.transform(labels)\n",
    "y = to_categorical(labels)\n",
    "\n",
    "# save label file so we can use in another script\n",
    "np.save('license_character_classes.npy', lb.classes_)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "AOHM-oeWSsPd"
   },
   "outputs": [],
   "source": [
    "# split 10% of data as validation set\n",
    "(trainX, testX, trainY, testY) = train_test_split(X, y, test_size=0.10, stratify=y, random_state=42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "1DfUhlOTcurq"
   },
   "outputs": [],
   "source": [
    "# data augumentation\n",
    "image_gen = ImageDataGenerator(rotation_range=10,\n",
    "                              width_shift_range=0.1,\n",
    "                              height_shift_range=0.1,\n",
    "                              shear_range=0.1,\n",
    "                              zoom_range=0.1,\n",
    "                              fill_mode=\"nearest\"\n",
    "                              )"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "vNMFc03Mb45i"
   },
   "source": [
    "## Initialize MobileNets architecture with pre-trained weight"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "AaemgepFcur0"
   },
   "outputs": [],
   "source": [
    "# Create our model with pre-trained MobileNetV2 architecture from imagenet\n",
    "def create_model(lr=1e-4,decay=1e-4/25, training=False,output_shape=y.shape[1]):\n",
    "    baseModel = MobileNetV2(weights=\"imagenet\", \n",
    "                            include_top=False,\n",
    "                            input_tensor=Input(shape=(80, 80, 3)))\n",
    "\n",
    "    headModel = baseModel.output\n",
    "    headModel = AveragePooling2D(pool_size=(3, 3))(headModel)\n",
    "    headModel = Flatten(name=\"flatten\")(headModel)\n",
    "    headModel = Dense(128, activation=\"relu\")(headModel)\n",
    "    headModel = Dropout(0.5)(headModel)\n",
    "    headModel = Dense(output_shape, activation=\"softmax\")(headModel)\n",
    "    \n",
    "    model = Model(inputs=baseModel.input, outputs=headModel)\n",
    "    \n",
    "    if training:\n",
    "        # define trainable lalyer\n",
    "        for layer in baseModel.layers:\n",
    "            layer.trainable = True\n",
    "        # compile model\n",
    "        optimizer = Adam(lr=lr, decay = decay)\n",
    "        model.compile(loss=\"categorical_crossentropy\", optimizer=optimizer,metrics=[\"accuracy\"])    \n",
    "        \n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 108
    },
    "colab_type": "code",
    "id": "-B5O6zG_cur5",
    "outputId": "83bfaffb-85bd-4219-da8e-ed68d7c656ad"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:`input_shape` is undefined or non-square, or `rows` is not in [96, 128, 160, 192, 224]. Weights for input shape (224, 224) will be loaded as the default.\n",
      "Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/mobilenet_v2/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_224_no_top.h5\n",
      "9412608/9406464 [==============================] - 3s 0us/step\n"
     ]
    }
   ],
   "source": [
    "# initilaize initial hyperparameter\n",
    "INIT_LR = 1e-4\n",
    "EPOCHS = 30\n",
    "\n",
    "model = create_model(lr=INIT_LR, decay=INIT_LR/EPOCHS,training=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "JR0L3nWgb-gK"
   },
   "source": [
    "## Train model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 1000
    },
    "colab_type": "code",
    "id": "4d4NQLwDcusA",
    "outputId": "29c5fad5-3f23-427c-f04c-01af2a140e40"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/30\n",
      "529/529 [==============================] - 1069s 2s/step - loss: 2.4566 - accuracy: 0.3716 - val_loss: 1.7755 - val_accuracy: 0.5876\n",
      "\n",
      "Epoch 00001: saving model to License_character_recognition.h5\n",
      "Epoch 2/30\n",
      "529/529 [==============================] - 982s 2s/step - loss: 0.4508 - accuracy: 0.8779 - val_loss: 0.4975 - val_accuracy: 0.8794\n",
      "\n",
      "Epoch 00002: saving model to License_character_recognition.h5\n",
      "Epoch 3/30\n",
      "529/529 [==============================] - 992s 2s/step - loss: 0.2955 - accuracy: 0.9184 - val_loss: 0.2489 - val_accuracy: 0.9336\n",
      "\n",
      "Epoch 00003: saving model to License_character_recognition.h5\n",
      "Epoch 4/30\n",
      "529/529 [==============================] - 977s 2s/step - loss: 0.2352 - accuracy: 0.9323 - val_loss: 0.1782 - val_accuracy: 0.9490\n",
      "\n",
      "Epoch 00004: saving model to License_character_recognition.h5\n",
      "Epoch 5/30\n",
      "529/529 [==============================] - 993s 2s/step - loss: 0.1939 - accuracy: 0.9453 - val_loss: 0.1607 - val_accuracy: 0.9562\n",
      "\n",
      "Epoch 00005: saving model to License_character_recognition.h5\n",
      "Epoch 6/30\n",
      "529/529 [==============================] - 948s 2s/step - loss: 0.1698 - accuracy: 0.9509 - val_loss: 0.1403 - val_accuracy: 0.9596\n",
      "\n",
      "Epoch 00006: saving model to License_character_recognition.h5\n",
      "Epoch 7/30\n",
      "529/529 [==============================] - 899s 2s/step - loss: 0.1534 - accuracy: 0.9544 - val_loss: 0.1306 - val_accuracy: 0.9660\n",
      "\n",
      "Epoch 00007: saving model to License_character_recognition.h5\n",
      "Epoch 8/30\n",
      "529/529 [==============================] - 895s 2s/step - loss: 0.1399 - accuracy: 0.9595 - val_loss: 0.1279 - val_accuracy: 0.9625\n",
      "\n",
      "Epoch 00008: saving model to License_character_recognition.h5\n",
      "Epoch 9/30\n",
      "529/529 [==============================] - 897s 2s/step - loss: 0.1245 - accuracy: 0.9633 - val_loss: 0.1116 - val_accuracy: 0.9665\n",
      "\n",
      "Epoch 00009: saving model to License_character_recognition.h5\n",
      "Epoch 10/30\n",
      "529/529 [==============================] - 894s 2s/step - loss: 0.1118 - accuracy: 0.9659 - val_loss: 0.1062 - val_accuracy: 0.9681\n",
      "\n",
      "Epoch 00010: saving model to License_character_recognition.h5\n",
      "Epoch 11/30\n",
      "529/529 [==============================] - 911s 2s/step - loss: 0.1192 - accuracy: 0.9635 - val_loss: 0.1010 - val_accuracy: 0.9689\n",
      "\n",
      "Epoch 00011: saving model to License_character_recognition.h5\n",
      "Epoch 12/30\n",
      "529/529 [==============================] - 925s 2s/step - loss: 0.1014 - accuracy: 0.9671 - val_loss: 0.0833 - val_accuracy: 0.9737\n",
      "\n",
      "Epoch 00012: saving model to License_character_recognition.h5\n",
      "Epoch 13/30\n",
      "529/529 [==============================] - 892s 2s/step - loss: 0.0972 - accuracy: 0.9711 - val_loss: 0.1058 - val_accuracy: 0.9705\n",
      "\n",
      "Epoch 00013: saving model to License_character_recognition.h5\n",
      "Epoch 14/30\n",
      "529/529 [==============================] - 895s 2s/step - loss: 0.0945 - accuracy: 0.9715 - val_loss: 0.1005 - val_accuracy: 0.9729\n",
      "\n",
      "Epoch 00014: saving model to License_character_recognition.h5\n",
      "Epoch 15/30\n",
      "529/529 [==============================] - 900s 2s/step - loss: 0.0893 - accuracy: 0.9725 - val_loss: 0.0986 - val_accuracy: 0.9702\n",
      "\n",
      "Epoch 00015: saving model to License_character_recognition.h5\n",
      "Epoch 16/30\n",
      "529/529 [==============================] - 893s 2s/step - loss: 0.0880 - accuracy: 0.9736 - val_loss: 0.0913 - val_accuracy: 0.9729\n",
      "\n",
      "Epoch 00016: saving model to License_character_recognition.h5\n",
      "Epoch 17/30\n",
      "529/529 [==============================] - 890s 2s/step - loss: 0.0821 - accuracy: 0.9729 - val_loss: 0.0924 - val_accuracy: 0.9713\n",
      "\n",
      "Epoch 00017: saving model to License_character_recognition.h5\n"
     ]
    }
   ],
   "source": [
    "BATCH_SIZE = 64\n",
    "\n",
    "my_checkpointer = [\n",
    "                EarlyStopping(monitor='val_loss', patience=5, verbose=0),\n",
    "                ModelCheckpoint(filepath=\"License_character_recognition.h5\", verbose=1, save_weights_only=True)\n",
    "                ]\n",
    "\n",
    "result = model.fit(image_gen.flow(trainX, trainY, batch_size=BATCH_SIZE), \n",
    "                   steps_per_epoch=len(trainX) // BATCH_SIZE, \n",
    "                   validation_data=(testX, testY), \n",
    "                   validation_steps=len(testX) // BATCH_SIZE, \n",
    "                   epochs=EPOCHS, callbacks=my_checkpointer)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "va-WvqTpcCJw"
   },
   "source": [
    "## Visualize training result"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 362
    },
    "colab_type": "code",
    "id": "qUYOQrVzcusF",
    "outputId": "66bf8f56-e839-40bb-f1b7-3d7307c7fe26"
   },
   "outputs": [],
   "source": [
    "\n",
    "fig = plt.figure(figsize=(14,5))\n",
    "grid=gridspec.GridSpec(ncols=2,nrows=1,figure=fig)\n",
    "fig.add_subplot(grid[0])\n",
    "plt.plot(result.history['accuracy'], label='training accuracy')\n",
    "plt.plot(result.history['val_accuracy'], label='val accuracy')\n",
    "plt.title('Accuracy')\n",
    "plt.xlabel('epochs')\n",
    "plt.ylabel('accuracy')\n",
    "plt.legend()\n",
    "\n",
    "fig.add_subplot(grid[1])\n",
    "plt.plot(result.history['loss'], label='training loss')\n",
    "plt.plot(result.history['val_loss'], label='val loss')\n",
    "plt.title('Loss')\n",
    "plt.xlabel('epochs')\n",
    "plt.ylabel('loss')\n",
    "plt.legend()\n",
    "plt.savefig(\"Training_result.jpg\",dpi=300)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "w8JOq60Xw1UY"
   },
   "outputs": [],
   "source": [
    "# save model architectur as json file\n",
    "model_json = model.to_json()\n",
    "with open(\"MobileNets_character_recognition.json\", \"w\") as json_file:\n",
    "  json_file.write(model_json)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "t92Je4kBbs0g"
   },
   "source": []
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "collapsed_sections": [],
   "name": "Train_license_character_recognition.ipynb",
   "provenance": [],
   "toc_visible": true
  },
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}