Deep Learning

Model Types

Artificial Neural Networks

Requires Feature Scaling
- Scaling applied to all independent data points (including encoded ones)

Sample Code

import tensorflow as tf

# Initializing the ANN
dl_ann = tf.keras.models.Sequential()

# Adding the first hidden layer
# Here we are using Dense layers 
# units - neurons in hidden layer. its a hyperparameter and is determined by trial and error
# relu is the rectifier activation function
dl_ann.add(tf.keras.layers.Dense(units=6, activation='relu'))

# Adding the second hidden layer
dl_ann.add(tf.keras.layers.Dense(units=6, activation='relu'))

# Adding the output layer
# Since we are doing binary classification, the units in the output layer is set to 1
dl_ann.add(tf.keras.layers.Dense(units=1, activation='sigmoid'))

# Compile the ANN with an optimizer and loss function
# Adam optimizer performs Stochastic Gradient Descent
dl_ann.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

# Training the ANN on the Training set
# batch size and epochs are hyperparameters
dl_ann.fit(X_train, y_train, batch_size = 32, epochs = 100)

y_pred_ann = dl_ann.predict(sc.transform(X_test))
# Convert predicted probability to prediction
y_pred_ann = (y_pred_ann > 0.5)

Convolutional Neural Networks

Sample Code

import tensorflow as tf
from keras.preprocessing.image import ImageDataGenerator
import keras.utils as image

# Preprocess training set
# Image Augmentation - Transform data to avoid overfitting
train_datagen = ImageDataGenerator(rescale = 1./255,     # Feature scale all pixels (divides by 255)
                                shear_range = 0.2,
                                zoom_range = 0.2,
                                horizontal_flip = True)
training_set = train_datagen.flow_from_directory('data/CNNData1/training_set',
                                            target_size = (64, 64),   # Final size of images before feeding to CNN; smaller size results in faster training
                                            batch_size = 32,          # Number of images in each batch
                                            class_mode = 'binary')    # cat/dog in this example

# Preprocess test set
# Only apply feature scaling to the test set. No transformations applied.
test_datagen = ImageDataGenerator(rescale = 1./255)
test_set = test_datagen.flow_from_directory('data/CNNData1/test_set',
                                            target_size = (64, 64),
                                            batch_size = 32,
                                            class_mode = 'binary')

# Initializing the CNN
cnn = tf.keras.models.Sequential()

# Step1 - Add Convolution Layer
# input shae is needed only for the first layer
cnn.add(tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation='relu', input_shape=[64, 64, 3]))

# Step2 - Pooling
cnn.add(tf.keras.layers.MaxPool2D(pool_size=2, strides=2))

# Add a second convolution and pooling layer
cnn.add(tf.keras.layers.Conv2D(filters=32, kernel_size=3, activation='relu'))
cnn.add(tf.keras.layers.MaxPool2D(pool_size=2, strides=2))

# Step3 - Flattening
cnn.add(tf.keras.layers.Flatten())

# Full Connection
cnn.add(tf.keras.layers.Dense(units=128, activation='relu'))

# Output Layer
cnn.add(tf.keras.layers.Dense(units=1, activation='sigmoid'))

# Compile the CNN with an optimizer and loss function
cnn.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

# Make a single prediction
test_image = image.load_img('data/CNNData1/single_pred/cat_or_dog1.jpg', target_size = (64, 64)) # set size to 64 * 64 so that it matches the size used during training
test_image = image.img_to_array(test_image)    # convert to array
test_image = np.expand_dims(test_image, axis = 0)  # add a extra dummy dimension that will correspond to the batch size so that the format matches what the predict method expects 
result = cnn.predict(test_image)
training_set.class_indices  # allows us to know which index correspondes to which output. In this case, it will show that 1 is dog and 0 is cat
if result[0][0] == 1:
prediction = 'dog'
else:
prediction = 'cat'
print(prediction)