<\/p>\n\n\n\n
The Directory Structure <\/h2>\n\n\n\n
It is not mandatory that you should use the same data as provided here. You can use your own data and by the end of this post, you will have a personal project ready. Just make sure that the folder structure of the data should be the same as provided. The following is the tree structure of the project directory.<\/p>\n\n\n\n
\u251c\u2500\u2500\u2500.ipynb_checkpoints\n\u251c\u2500\u2500\u2500data\n\u2502 \u251c\u2500\u2500\u2500golden_tiger\n\u2502 \u251c\u2500\u2500\u2500mountain_lion\n\u2502 \u2514\u2500\u2500\u2500white_tiger\n\u251c\u2500\u2500\u2500images\n\u2514\u2500\u2500\u2500logs\n \u251c\u2500\u2500\u2500wild-cats-1560494714\n \u2514\u2500\u2500\u2500wild-cats-1560495009<\/pre>\n\n\n\nWe have the
data<\/code> directory and three sub-directories inside it. The sub-directories (golden_tiger, mountain_lion, white_tiger<\/em>) contain the respective images of the wild cats. If you consider using your own data, then just replace the sub-directories with your own. The names of the sub-directories will act as labels. So, it will be better if you give intuitive names to those.<\/p>\n\n\n\nThe
images<\/code> directory contains the test images that we will use after the model has been trained.<\/p>\n\n\n\nThe
logs<\/code> directory contains the TensorBoard logs. We will be using TensorBoard to visualize the train and validation plots and also see the graph of the model. If you are going to use TensorBoard for the first time, then you may find this<\/a><\/strong> article useful. <\/p>\n\n\n\nMoreover, if you want to download new images for this tutorial, then please refer to my previous article – Create Your Own Deep Learning Image Dataset<\/strong><\/a>. Just one more thing. We will be using
tf.keras<\/code> module instead of the direct Keras API. If you find the model building codes too long, then feel free to use Keras directly. Okay, let’s start.<\/p>\n\n\n\nLoading Data and Preprocessing<\/h2>\n\n\n\n
The following block of code imports all the necessary packages.<\/p>\n\n\n\n
# import packages\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport numpy as np\nimport random\nimport pickle\nimport time\nimport cv2\nimport os\n\nfrom sklearn.preprocessing import LabelBinarizer\nfrom sklearn.model_selection import train_test_split\nfrom imutils import paths<\/pre>\n\n\n\nThe above code is mostly self-explanatory. <\/p>\n\n\n\n
time<\/code> module will help us to get a different timestamp <\/strong>for each of the TensorBoard logs so that the previous ones do not get replaced.<\/li>- Using
cv2<\/code> we will read the images, make necessary changes before feeding the image to the model and also visualize the predictions<\/strong>.<\/li>- We will use
LabelBinarizer<\/code> from Scikit-Learn to get the image labels<\/strong>.<\/li><\/ul>\n\n\n\nMoving on, let’s get the image paths:<\/p>\n\n\n\n
# path to the images\ndata_path = 'data'\n\nimages = []\nimage_labels = []\n\n# get the image paths inside the `data` directory\nimage_paths = sorted(list(paths.list_images(data_path)))\n# for reproducibility\nrandom.seed(42)\n# shuffle the images\nrandom.shuffle(image_paths)<\/pre>\n\n\n\nWe initialize two lists,
images<\/code> andimage_labels<\/code> which will hold our images and the labels of the images respectively. For this project, we have three labels, golden_tiger<\/em>, white_tiger<\/em> and mountain_lion<\/em>. Next, we userandom.seed()<\/code> for reproducibility, in case we need to run the code again. Finally, we shuffle all the images.<\/p>\n\n\n\nNow, let’s read the images and resize them using
cv2<\/code>. We will also append each image to theimages<\/code> list and the labels to theimage_labels<\/code> list.<\/p>\n\n\n\nfor image_path in image_paths:\n # load and resize the images\n image = cv2.imread(image_path)\n image = cv2.resize(image, (128, 128))\n images.append(image)\n \n # get the labels from the image paths and store in `image_labels`\n image_label = image_path.split(os.path.sep)[-2]\n image_labels.append(image_label)<\/pre>\n\n\n\nIt is a good idea to rescale the image pixels. Rescaling is a very important factor when considering the accuracy and performance of neural networks.<\/p>\n\n\n\n
# rescale the image pixels \nimages = np.array(images, dtype='float') \/ 255.0\n# make `image_labels` as array\nimage_labels = np.array(image_labels)<\/pre>\n\n\n\nMoving on, we will now split the data into train and validation test. We will be using 80% of the data for training and 20% for validation.<\/p>\n\n\n\n
# divide the image data into train set and test set\n(train_X, test_X, train_y, test_y) = train_test_split(images, image_labels, \n test_size=0.2, \n random_state=42)<\/pre>\n\n\n\nThe
image_labels<\/code> is a list of strings. To use it properly during training and inference (prediction) we need to convert those into one-hot labels. <\/p>\n\n\n\n# one-hot encode the labels\nlb = LabelBinarizer()\ntrain_y = lb.fit_transform(train_y)\ntest_y = lb.fit_transform(test_y)<\/pre>\n\n\n\nSo, now for golden_tiger<\/em>, we will have [1, 0, 0]. This means that the image is true for golden_tiger<\/em> only (hence, the 1). Similarly, it will be [0, 1, 0], [0, 0, 1] for mountain_lion<\/em> and white_tiger<\/em> respectively.<\/p>\n\n\n\n
Next, we will insert the code image augmentation generator. This will provide our model with a lot of different cases to learn while training. Hopefully, this step will help in achieving better accuracy.<\/p>\n\n\n\n
# generator for image augmentation\nimage_aug = tf.keras.preprocessing.image.ImageDataGenerator(rotation_range=30, shear_range=0.2, \n zoom_range=0.2, height_shift_range=0.2, width_shift_range=0.2, \n horizontal_flip=True, fill_mode='nearest')<\/pre>\n\n\n\nFor data augmentation, we are using the
ImageDataGenerator<\/code> fromtf.keras<\/code> module.<\/p>\n\n\n\nBuilding the Model<\/h2>\n\n\n\n
Now, its time to build our model. Frankly, the model does have a lot of layers. Let’s insert the code fist, then I will move on to explain the layers.<\/p>\n\n\n\n
# build the model\nmodel = tf.keras.models.Sequential()\ninput_shape = (128, 128, 3)\n\nmodel.add(tf.keras.layers.Conv2D(32, (3, 3), padding='same', \n activation='relu', input_shape=input_shape))\nmodel.add(tf.keras.layers.BatchNormalization(axis=-1)) \nmodel.add(tf.keras.layers.MaxPooling2D(pool_size=(2, 2)))\nmodel.add(tf.keras.layers.Dropout(0.2))\n \nmodel.add(tf.keras.layers.Conv2D(64, (3, 3), padding='same', \n activation='relu'))\nmodel.add(tf.keras.layers.BatchNormalization(axis=-1))\nmodel.add(tf.keras.layers.Conv2D(64, (3, 3), padding='same', \n activation='relu'))\nmodel.add(tf.keras.layers.BatchNormalization(axis=-1))\nmodel.add(tf.keras.layers.MaxPooling2D(pool_size=(2, 2)))\nmodel.add(tf.keras.layers.Dropout(0.2))\n\nmodel.add(tf.keras.layers.Conv2D(128, (3, 3), padding='same', \n activation='relu'))\nmodel.add(tf.keras.layers.BatchNormalization(axis=-1))\nmodel.add(tf.keras.layers.Conv2D(128, (3, 3), padding='same', \n activation='relu'))\nmodel.add(tf.keras.layers.BatchNormalization(axis=-1))\nmodel.add(tf.keras.layers.Conv2D(128, (3, 3), padding='same', \n activation='relu'))\nmodel.add(tf.keras.layers.BatchNormalization(axis=-1))\nmodel.add(tf.keras.layers.MaxPooling2D(pool_size=(2, 2)))\nmodel.add(tf.keras.layers.Dropout(0.2))\n\nmodel.add(tf.keras.layers.Flatten())\nmodel.add(tf.keras.layers.Dense(512, activation='relu'))\nmodel.add(tf.keras.layers.BatchNormalization())\nmodel.add(tf.keras.layers.Dropout(0.2))\n\nmodel.add(tf.keras.layers.Dense(len(lb.classes_), activation='softmax'))<\/pre>\n\n\n\nFirst of all, we are using the
Sequntial()<\/code> API to build the model. Secondly, the we have given the input shape as(128, 128, 3)<\/code> for the height, width and channels. So, it is a channels last input shape.<\/p>\n\n\n\nNow, moving on to the actual layers, we are using
Conv2D<\/code> withrelu<\/code> activation function andpadding=same<\/code>. For the first convolutional layer we have given the output dimensionality as 32. Then we gradually increase it to 64 and 128. Each of theConv2D<\/code> layer is having its ownBatchNormalization<\/code> layer as well. Theaxis<\/code> is -1 as we are using channels last. Then we have theMaxPooling<\/code> layer withpool_size=(2, 2)<\/code> andDropout<\/code> layer with a rate of 0.2.<\/p>\n\n\n\nIn the model, the only exceptions are the last hidden layer and the output layer. For the last hidden layer we are using
Dense<\/code> with 512 dimensionality. The output layer is usingsoftmax<\/code> activation function and the output dimensionality islen(lb.classes)<\/code>. That would amount to 3 as we have 3 classes in total. <\/p>\n\n\n\n
![\"Graph](\"https:\/\/debuggercafe.com\/wp-content\/uploads\/2019\/06\/plot-e1560743463116.png\")
![\"Image](\"https:\/\/debuggercafe.com\/wp-content\/uploads\/2019\/06\/Tensorboard-plots-1.png\")
![\"Image](\"https:\/\/debuggercafe.com\/wp-content\/uploads\/2019\/06\/Tensorboard-graph.png\")
![\"Image](\"https:\/\/debuggercafe.com\/wp-content\/uploads\/2019\/06\/tiger-e1560756008126.png\")