Keras Crack Classifier

CrackClassifier.py

import tensorflow as tf
from IPython.display import display
from PIL import Image
import keras
%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
import math
print(tf.__version__)
print(keras.__version__)

# View Available GPUs
from keras import backend as K
K.tensorflow_backend._get_available_gpus()

"""------------------------
Data Loaders
------------------------"""
from keras.preprocessing.image import ImageDataGenerator
from keras.applications.vgg16 import preprocess_input

img_height, img_width = 224, 224
batch_size = 10
train_data_dir = "./dataset"
train_datagen = ImageDataGenerator(shear_range=0.2,
                                   zoom_range=0.2,
                                   horizontal_flip=True,
                                    rotation_range=45,
                                   validation_split=0.2,
                                   preprocessing_function=preprocess_input) # set validation split

train_generator = train_datagen.flow_from_directory(train_data_dir,
                                                   target_size=(img_height, img_width),
                                                   batch_size=batch_size,
                                                   class_mode='binary',
                                                   subset='training') # set as training data
validation_generator = train_datagen.flow_from_directory(train_data_dir,
                                                   target_size=(img_height, img_width),
                                                   batch_size=batch_size,
                                                   class_mode='binary',
                                                   subset='validation') # set as validation data
                                                   

"""------------------------
Create And Traing Model
------------------------"""
from keras import applications
from keras.models import Sequential, load_model
from keras.layers import Dropout, Flatten, Dense, Reshape, MaxPooling2D, Conv2D
from keras.regularizers import l2
from keras.layers.normalization import BatchNormalization
from keras import optimizers

# Function to create the model, required for KerasClassifier
def create_model(learning_rate, num_dense_layers,
                 num_dense_nodes, activation, dropout_rate,
                 momentum): 
    """
    Hyper-parameters:
    learning_rate:     Learning-rate for the optimizer.
    num_dense_layers:  Number of dense layers.
    num_dense_nodes:   Number of nodes in each dense layer.
    activation:        Activation function for all layers.
    """

    # build the VGG16 network from base imagenet model
    base_model = applications.VGG16(weights='imagenet', include_top=True, input_shape=(img_height, img_width, 3))

    # Build out network for transfer learning
    model = Sequential()
    # take off last two layers
    for layer in base_model.layers[:-1]:
        model.add(layer)
        
    for i in range(num_dense_layers):
        name = 'layer_dense_{0}'.format(i+1)
        model.add(Dense(num_dense_nodes, activation=activation, name=name))
        model.add(Dropout(dropout_rate))
        model.add(BatchNormalization())
    model.add(Dense(1, activation='sigmoid'))

    # set the first 11 layers (up to the last conv block)
    for layer in model.layers[:11]:
        layer.trainable = False
    # setup optimization strategy
    model.compile(loss='binary_crossentropy',
                 optimizer=optimizers.SGD(lr=learning_rate, momentum=momentum),
                 metrics=['accuracy'])
    return model
   
from keras.callbacks import ModelCheckpoint, EarlyStopping
from keras.backend import manual_variable_initialization

def train_model(model, epochs=40, epochs_to_wait=10):
    """
    Hyper-parameters:
    learning_rate:     Learning-rate for the optimizer.
    num_dense_layers:  Number of dense layers.
    num_dense_nodes:   Number of nodes in each dense layer.
    activation:        Activation function for all layers.
    """        
    checkpoint_callback = ModelCheckpoint('crack_classifier_chkpt.h5', monitor='val_loss', verbose=1, save_best_only=True, mode='min')
    early_stopping_callback = EarlyStopping(monitor='val_loss', patience=epochs_to_wait)
    
    # Use Keras to train the model.
    history = model.fit_generator(
        train_generator,
        steps_per_epoch=5249//batch_size,
        epochs=epochs,
        validation_data=validation_generator,
        validation_steps=1312//batch_size, verbose=True,
        callbacks=[early_stopping_callback, checkpoint_callback])
    return history
    
# Create the neural network with these hyper-parameters.
model = create_model(1e-5, 1, 16, 'relu', .5,.8,)
# Train
history = train_model(model)
# Output Below
"""
Epoch 1/40
524/524 [==============================] - 1329s 3s/step - loss: 0.8319 - acc: 0.5246 - val_loss: 0.6275 - val_acc: 0.7221

Epoch 00001: val_loss improved from inf to 0.62746, saving model to crack_classifier_chkpt.h5
Epoch 2/40
524/524 [==============================] - 1312s 3s/step - loss: 0.7429 - acc: 0.5464 - val_loss: 0.6715 - val_acc: 0.7107

Epoch 00002: val_loss did not improve from 0.62746
Epoch 3/40
524/524 [==============================] - 1156s 2s/step - loss: 0.7017 - acc: 0.5786 - val_loss: 0.6216 - val_acc: 0.8298

Epoch 00003: val_loss improved from 0.62746 to 0.62158, saving model to crack_classifier_chkpt.h5
Epoch 4/40
524/524 [==============================] - 1280s 2s/step - loss: 0.6711 - acc: 0.6103 - val_loss: 0.6136 - val_acc: 0.8649

Epoch 00004: val_loss improved from 0.62158 to 0.61362, saving model to crack_classifier_chkpt.h5
Epoch 5/40
524/524 [==============================] - 1460s 3s/step - loss: 0.6485 - acc: 0.6380 - val_loss: 0.5730 - val_acc: 0.9176

Epoch 00005: val_loss improved from 0.61362 to 0.57303, saving model to crack_classifier_chkpt.h5
Epoch 6/40
524/524 [==============================] - 1369s 3s/step - loss: 0.6248 - acc: 0.6727 - val_loss: 0.5979 - val_acc: 0.8962

Epoch 00006: val_loss did not improve from 0.57303
Epoch 7/40
524/524 [==============================] - 1079s 2s/step - loss: 0.6032 - acc: 0.6977 - val_loss: 0.6165 - val_acc: 0.8458

Epoch 00007: val_loss did not improve from 0.57303
Epoch 8/40
524/524 [==============================] - 997s 2s/step - loss: 0.5784 - acc: 0.7260 - val_loss: 0.5539 - val_acc: 0.8733

Epoch 00008: val_loss improved from 0.57303 to 0.55389, saving model to crack_classifier_chkpt.h5
Epoch 9/40
524/524 [==============================] - 976s 2s/step - loss: 0.5486 - acc: 0.7969 - val_loss: 0.5246 - val_acc: 0.9344

Epoch 00009: val_loss improved from 0.55389 to 0.52456, saving model to crack_classifier_chkpt.h5
Epoch 10/40
524/524 [==============================] - 1004s 2s/step - loss: 0.5252 - acc: 0.8483 - val_loss: 0.5030 - val_acc: 0.9420

Epoch 00010: val_loss improved from 0.52456 to 0.50295, saving model to crack_classifier_chkpt.h5
Epoch 11/40
524/524 [==============================] - 964s 2s/step - loss: 0.5062 - acc: 0.8912 - val_loss: 0.4630 - val_acc: 0.9893

Epoch 00011: val_loss improved from 0.50295 to 0.46302, saving model to crack_classifier_chkpt.h5
Epoch 12/40
524/524 [==============================] - 1045s 2s/step - loss: 0.4782 - acc: 0.9924 - val_loss: 0.4979 - val_acc: 0.9916

Epoch 00012: val_loss did not improve from 0.46302
Epoch 13/40
524/524 [==============================] - 1113s 2s/step - loss: 0.4631 - acc: 0.9968 - val_loss: 0.4762 - val_acc: 0.9931

Epoch 00013: val_loss did not improve from 0.46302
Epoch 14/40
524/524 [==============================] - 1002s 2s/step - loss: 0.4499 - acc: 0.9954 - val_loss: 0.4477 - val_acc: 0.9962

Epoch 00014: val_loss improved from 0.46302 to 0.44765, saving model to crack_classifier_chkpt.h5
Epoch 15/40
524/524 [==============================] - 954s 2s/step - loss: 0.4368 - acc: 0.9973 - val_loss: 0.4460 - val_acc: 0.9954

Epoch 00015: val_loss improved from 0.44765 to 0.44603, saving model to crack_classifier_chkpt.h5
Epoch 16/40
524/524 [==============================] - 1197s 2s/step - loss: 0.4242 - acc: 0.9985 - val_loss: 0.4209 - val_acc: 0.9954

Epoch 00016: val_loss improved from 0.44603 to 0.42094, saving model to crack_classifier_chkpt.h5
Epoch 17/40
524/524 [==============================] - 1043s 2s/step - loss: 0.4121 - acc: 0.9983 - val_loss: 0.4155 - val_acc: 0.9977

Epoch 00017: val_loss improved from 0.42094 to 0.41545, saving model to crack_classifier_chkpt.h5
Epoch 18/40
524/524 [==============================] - 987s 2s/step - loss: 0.4008 - acc: 0.9983 - val_loss: 0.4102 - val_acc: 0.9947

Epoch 00018: val_loss improved from 0.41545 to 0.41017, saving model to crack_classifier_chkpt.h5
Epoch 19/40
524/524 [==============================] - 1065s 2s/step - loss: 0.3900 - acc: 0.9981 - val_loss: 0.4013 - val_acc: 0.9939

Epoch 00019: val_loss improved from 0.41017 to 0.40132, saving model to crack_classifier_chkpt.h5
Epoch 20/40
524/524 [==============================] - 990s 2s/step - loss: 0.3793 - acc: 0.9983 - val_loss: 0.3932 - val_acc: 0.9947

Epoch 00020: val_loss improved from 0.40132 to 0.39316, saving model to crack_classifier_chkpt.h5
Epoch 21/40
524/524 [==============================] - 1082s 2s/step - loss: 0.3692 - acc: 0.9989 - val_loss: 0.3665 - val_acc: 0.9985

Epoch 00021: val_loss improved from 0.39316 to 0.36652, saving model to crack_classifier_chkpt.h5
Epoch 22/40
524/524 [==============================] - 1116s 2s/step - loss: 0.3594 - acc: 0.9985 - val_loss: 0.3581 - val_acc: 0.9962

Epoch 00022: val_loss improved from 0.36652 to 0.35806, saving model to crack_classifier_chkpt.h5
Epoch 23/40
524/524 [==============================] - 1180s 2s/step - loss: 0.3502 - acc: 0.9989 - val_loss: 0.3446 - val_acc: 0.9969

Epoch 00023: val_loss improved from 0.35806 to 0.34459, saving model to crack_classifier_chkpt.h5
Epoch 24/40
524/524 [==============================] - 1100s 2s/step - loss: 0.3410 - acc: 0.9994 - val_loss: 0.3483 - val_acc: 0.9985

Epoch 00024: val_loss did not improve from 0.34459
Epoch 25/40
524/524 [==============================] - 1057s 2s/step - loss: 0.3323 - acc: 0.9998 - val_loss: 0.3295 - val_acc: 0.9992

Epoch 00025: val_loss improved from 0.34459 to 0.32946, saving model to crack_classifier_chkpt.h5
Epoch 26/40
524/524 [==============================] - 1295s 2s/step - loss: 0.3241 - acc: 0.9998 - val_loss: 0.3204 - val_acc: 0.9992

Epoch 00026: val_loss improved from 0.32946 to 0.32040, saving model to crack_classifier_chkpt.h5
Epoch 27/40
524/524 [==============================] - 1089s 2s/step - loss: 0.3161 - acc: 0.9996 - val_loss: 0.3248 - val_acc: 0.9977

Epoch 00027: val_loss did not improve from 0.32040
Epoch 28/40
524/524 [==============================] - 994s 2s/step - loss: 0.3084 - acc: 0.9996 - val_loss: 0.3103 - val_acc: 0.9985

Epoch 00028: val_loss improved from 0.32040 to 0.31030, saving model to crack_classifier_chkpt.h5
Epoch 29/40
524/524 [==============================] - 994s 2s/step - loss: 0.3009 - acc: 1.0000 - val_loss: 0.3058 - val_acc: 0.9985

Epoch 00029: val_loss improved from 0.31030 to 0.30578, saving model to crack_classifier_chkpt.h5
Epoch 30/40
524/524 [==============================] - 967s 2s/step - loss: 0.2938 - acc: 0.9996 - val_loss: 0.2994 - val_acc: 0.9992

Epoch 00030: val_loss improved from 0.30578 to 0.29936, saving model to crack_classifier_chkpt.h5
Epoch 31/40
524/524 [==============================] - 986s 2s/step - loss: 0.2868 - acc: 0.9998 - val_loss: 0.2961 - val_acc: 0.9985

Epoch 00031: val_loss improved from 0.29936 to 0.29615, saving model to crack_classifier_chkpt.h5
Epoch 32/40
524/524 [==============================] - 986s 2s/step - loss: 0.2801 - acc: 1.0000 - val_loss: 0.2820 - val_acc: 0.9985

Epoch 00032: val_loss improved from 0.29615 to 0.28197, saving model to crack_classifier_chkpt.h5
Epoch 33/40
524/524 [==============================] - 1058s 2s/step - loss: 0.2736 - acc: 1.0000 - val_loss: 0.2736 - val_acc: 0.9992

Epoch 00033: val_loss improved from 0.28197 to 0.27365, saving model to crack_classifier_chkpt.h5
Epoch 34/40
524/524 [==============================] - 971s 2s/step - loss: 0.2674 - acc: 0.9998 - val_loss: 0.2641 - val_acc: 1.0000

Epoch 00034: val_loss improved from 0.27365 to 0.26415, saving model to crack_classifier_chkpt.h5
Epoch 35/40
524/524 [==============================] - 953s 2s/step - loss: 0.2615 - acc: 1.0000 - val_loss: 0.2630 - val_acc: 0.9992

Epoch 00035: val_loss improved from 0.26415 to 0.26295, saving model to crack_classifier_chkpt.h5
Epoch 36/40
524/524 [==============================] - 1176s 2s/step - loss: 0.2556 - acc: 1.0000 - val_loss: 0.2660 - val_acc: 0.9977

Epoch 00036: val_loss did not improve from 0.26295
Epoch 37/40
524/524 [==============================] - 1040s 2s/step - loss: 0.2502 - acc: 0.9996 - val_loss: 0.2531 - val_acc: 0.9992

Epoch 00037: val_loss improved from 0.26295 to 0.25314, saving model to crack_classifier_chkpt.h5

Epoch 38/40
524/524 [==============================] - 1019s 2s/step - loss: 0.2446 - acc: 1.0000 - val_loss: 0.2431 - val_acc: 1.0000

Epoch 00038: val_loss improved from 0.25314 to 0.24308, saving model to crack_classifier_chkpt.h5
Epoch 39/40
524/524 [==============================] - 1087s 2s/step - loss: 0.2393 - acc: 1.0000 - val_loss: 0.2380 - val_acc: 1.0000

Epoch 00039: val_loss improved from 0.24308 to 0.23800, saving model to crack_classifier_chkpt.h5
Epoch 40/40
524/524 [==============================] - 988s 2s/step - loss: 0.2342 - acc: 1.0000 - val_loss: 0.2344 - val_acc: 1.0000

Epoch 00040: val_loss improved from 0.23800 to 0.23444, saving model to crack_classifier_chkpt.h5

"""

"""------------------------
Test Model
------------------------"""
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array

from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array
def test_crack_model(model):
    crack_test_images = ["crack.jpg", "crack1.jpg", "crack2.jpg", "crack3.jpg",
                    "crack4.jpg", "crack5.jpg", "crack6.jpg", "crack7.jpg",
                    "crack8.jpg", "crack9.jpg", "crack10.jpg", "crack11.jpg",
                    "crack12.jpg", "crack13.jpg"]
    nocrack_test_images = ["nocrack1.jpg", "nocrack2.jpg", "nocrack3.jpg",
                          "nocrack4.jpg", "nocrack5.jpg", "nocrack6.jpg",
                          "nocrack7.jpg", "nocrack8.jpg"]
    print("============== Cracks ===============")

    for i, crack_test_image in enumerate(crack_test_images):
        # load an image from file
        image = load_img('Images/crack/'+crack_test_image, target_size=(224, 224))
        # convert the image pixels to a numpy array
        image = img_to_array(image)

        # reshape data for the model
        image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
        # preprocess input
        image = preprocess_input(image)
                # predict the probability across all output classes
        yhat = model.predict(image)
        print("{}: {}".format(i, yhat))


    print("============== No Cracks ===============")
    for i, nocrack_test_image in enumerate(nocrack_test_images):
        #break
        # load an image from file
        image = load_img('Images/nocrack/'+nocrack_test_image, target_size=(224, 224))
        # convert the image pixels to a numpy array
        image = img_to_array(image)
        # reshape data for the model
        image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
        # preprocess input
        image = preprocess_input(image)
        
        # predict the probability across all output classes
        yhat = model.predict(image)
        print("{}: {}".format(i, yhat))
        
   test_crack_model(model)
  
  # Output Below (Model outputs results like it was not trained
  """
  ============== Cracks ===============
============== Cracks ===============
0: [[0.20922232]]
1: [[0.21883659]]
2: [[0.20403491]]
3: [[0.20747341]]
4: [[0.21976307]]
5: [[0.2170394]]
6: [[0.19359364]]
7: [[0.2190856]]
8: [[0.2077934]]
9: [[0.20873123]]
10: [[0.21976307]]
11: [[0.20891656]]
12: [[0.02694948]]
13: [[0.2004625]]
============== No Cracks ===============
0: [[0.21425322]]
1: [[0.9999826]]
2: [[0.2190724]]
3: [[0.21664117]]
4: [[0.21976307]]
5: [[0.21329497]]
6: [[0.20993665]]
7: [[0.20981254]]
  """