def main():
""" Load images.
Extract HOG feature descriptors.
"""
# Load stored data
X_train = np.load('../data/augment/ImageAugment_input.npy')
print('=== TRAIN DATA ===')
print(X_train.shape)
X_test = np.load('../data/test/ImageTest_input.npy')
print('=== TEST DATA ===')
print(X_test.shape)
print("Extracting HOG features...")
RF_train = np.zeros([len(X_train), 32400])
for i in range(len(X_train)):
RF_train[i] = hog_feature(X_train[i])
print("FEATURE DESCRIPTORS")
print(RF_train.shape)
RF_test = np.zeros([len(X_test), 32400])
for i in range(len(X_test)):
RF_test[i] = hog_feature(X_test[i])
print(RF_test.shape)
# Save data
make_folder('../data/processed')
np.save('../data/processed/ImageTestHOG_input.npy', RF_test)
np.save('../data/processed/ImageTrainHOG_input.npy', RF_train)
def main():
""" Load the data.
Train SVM model using linear kernel.
Print accuracy on test data.
"""
# Load stored data
X_train = np.load('../data/processed/ImageTrainHOG_input.npy')
y_train = np.load('../data/augment/DiseaseAugment_input.npy')
print("=== TRAIN DATA ===")
print(X_train.shape)
print(y_train.shape)
X_test = np.load('../data/processed/ImageTestHOG_input.npy')
y_test = np.load('../data/test/DiseaseTest_input.npy')
print("=== TEST DATA ===")
print(X_test.shape)
print(y_test.shape)
# Classifier
svm_model = LinearSVC(C=0.01)
svm_model.fit(X_train, y_train)
print(svm_model.score(X_test, y_test))
make_folder('../results/models')
filename = '../results/models/SVM_model.sav'
joblib.dump(svm_model, filename)
def main():
""" Load data.
Train random forest model.
Print accuracy on test data.
"""
# Load stored data
X_train = np.load('../data/processed/ImageTrainHOG_input.npy')
y_train = np.load('../data/augment/DiseaseAugment_input.npy')
print("=== TRAIN DATA ===")
print(X_train.shape)
print(y_train.shape)
X_test = np.load('../data/processed/ImageTestHOG_input.npy')
y_test = np.load('../data/test/DiseaseTest_input.npy')
print("=== TEST DATA ===")
print(X_test.shape)
print(y_test.shape)
# Classifier
Random_classifier = RandomForestClassifier(n_estimators=500,
max_depth=35,
n_jobs=-1,
warm_start=True,
oob_score=True,
max_features='sqrt')
Random_classifier.fit(X_train, y_train)
print(Random_classifier.score(X_test, y_test))
make_folder('../results/models')
filename = '../results/models/Random_model.sav'
joblib.dump(Random_classifier, filename)
def con_matrix(model, x, y):
""" Plot confusion matrix for the given model.
Save the png in the output folder.
Args:
model (str): model for which visualization needs to be plot.
x (numpy array): test images.
y (numpy array): test labels.
"""
corr = []
if model == "random_forest":
loaded_model = joblib.load(os.path.join(ROOT_DIR, 'Random_model.sav'))
classifier_prediction = loaded_model.predict(x)
corr = confusion_matrix(y, classifier_prediction)
elif model == "svm":
loaded_model = joblib.load(os.path.join(ROOT_DIR, 'SVM_model.sav'))
classifier_prediction = loaded_model.predict(x)
corr = confusion_matrix(y, classifier_prediction)
elif model == "majority_voting":
classifier_prediction = np.load(os.path.join(ROOT_DIR, 'Ensemble.npy'))
corr = confusion_matrix(y, classifier_prediction)
elif model == "stacked_prediction":
labeler = LabelEncoder()
labeler.fit(y)
loaded_model = load_model(os.path.join(ROOT_DIR, 'custom_ensemble.h5'))
y_prediction = loaded_model.predict(
np.load('data/test/X_test_ensemble.npy'))
prediction = np.argmax(y_prediction, axis=-1)
prediction = labeler.inverse_transform(prediction)
corr = confusion_matrix(y, prediction)
make_confusion_matrix(
corr,
categories=['blackrot', 'ecsa', 'healthy', 'leafblight', 'pmildew'],
count=True,
percent=True,
color_bar=False,
xy_ticks=True,
xy_plot_labels=True,
sum_stats=True,
fig_size=(8, 6),
c_map='OrRd',
title='Confusion matrix')
# error correction - cropped heat map
b, t = plt.ylim() # discover the values for bottom and top
b += 0.5 # Add 0.5 to the bottom
t -= 0.5 # Subtract 0.5 from the top
plt.ylim(b, t) # update the ylim(bottom, top) values
make_folder('results/visualization')
plt.savefig('results/visualization/confusion_matrix_{}.png'.format(model),
bbox_inches='tight')
def tree():
""" Plot the tree for random forest model.
Save the dot file in output folder.
Convert dot file to png by using the command:
'dot -Tpng tree.dot -o tree.png'
"""
model = joblib.load(os.path.join(ROOT_DIR, 'Random_model.sav'))
tree_num = model.estimators_
make_folder('results/visualization')
for tree_in_forest in tree_num:
export_graphviz(tree_in_forest,
out_file='results/visualization/tree.dot',
filled=True,
rounded=True,
precision=2)
def plot(image, label, index, model, x_hog):
""" Display image, true label and predicted label.
Save the result in the output folder
Args:
image (numpy array): image to predict.
label (numpy array): true labels of corresponding images.
index (int): index of the test image entered by the user.
model (str): model to use. (Entered by user)
x_hog (numpy array): feature descriptors of the image.
"""
plt.figure(figsize=(8, 6))
plt.imshow(image[index])
plt.axis('off')
plt.title('True label: {}'.format(label[index]),
fontdict={
'fontweight': 'bold',
'fontsize': 'x-large'
})
predictions, percent = model_predict(model, image, x_hog, label)
if model == "majority_voting":
if predictions[index] == label[index]:
plt.suptitle('Predicted label: {}'.format(predictions[index]),
color="green")
else:
plt.suptitle('Predicted label: {}'.format(predictions[index]),
color="red")
else:
if predictions[index] == label[index]:
plt.suptitle('Predicted label: {} ({:.2f} %)'.format(
predictions[index],
np.max(percent[index]) * 100),
color="green")
else:
plt.suptitle('Predicted label: {} ({:.2f} %)'.format(
predictions[index],
np.max(percent[index]) * 100),
color="red")
make_folder('results/visualization')
plt.savefig('results/visualization/app.png', bbox_inches='tight')
def roc(x, y):
""" Plot ROC-AUC plot for random forest model.
Save the image in output folder.
Args:
x (numpy array): test images.
y (numpy array): test labels.
"""
model = joblib.load(os.path.join(ROOT_DIR, 'Random_model.sav'))
visualizer = yellowbrick.classifier.ROCAUC(model,
classes=[
'healthy', 'leaf_blight',
'ecsa', 'black rot',
'powdery mildew'
])
visualizer.score(x, y)
ax = visualizer.show()
make_folder('results/visualization')
ax.figure.savefig('results/visualization/auc_roc.png')
def plot(model):
""" Plot the accuracy and loss curve for the neural networks.
Save file in the output folder.
Args:
model (str): model for which visualization needs to be plot
"""
history_custom = Hist()
if model == "cnn_custom":
history_custom = pickle.load(
open(os.path.join(ROOT_DIR, 'custom_training_history.pkl'), 'rb'))
elif model == "vgg":
history_custom = pickle.load(
open(os.path.join(ROOT_DIR, 'vgg16_training_history.pkl'), 'rb'))
# Plot training & validation accuracy values
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=[15, 8])
ax1.plot(history_custom.history['acc'])
ax1.plot(history_custom.history['val_acc'])
ax1.set_title('Model accuracy')
ax1.set_ylabel('Accuracy')
ax1.set_xlabel('Epoch')
ax1.legend(['Train', 'Validation'], loc='lower right')
# Plot training & validation loss values
ax2.plot(history_custom.history['loss'])
ax2.plot(history_custom.history['val_loss'])
ax2.set_title('Model loss')
ax2.set_ylabel('Loss')
ax2.set_xlabel('Epoch')
ax2.legend(['Train', 'Validation'], loc='upper right')
make_folder('results/visualization')
plt.savefig('results/visualization/acc_loss_{}.png'.format(model))
def main():
""" Load images and json files from all folders and concatenate to form a single array.
Shuffle the array.
Split into test and train data sets.
"""
print('INFO: Extracting json data...')
# Accumulate data from json files
json_df_images = get_json_data(os.path.join(ROOT_DIR, 'images/'))
json_df_positive = get_json_data(os.path.join(ROOT_DIR, 'positive/'))
json_df_healthy = get_json_data(os.path.join(ROOT_DIR, 'healthy/'))
json_df_team4 = get_json_data(os.path.join(ROOT_DIR, 'team4/'))
json_df_team4_br = get_json_data(os.path.join(ROOT_DIR, 'team4_br/'))
json_df_leaf_blight = get_json_data(os.path.join(ROOT_DIR, 'leaf_blight/'))
# Accumulate data set from all folders
print('INFO: Extracting images and corresponding labels...')
array1, disease1 = get_images(os.path.join(ROOT_DIR, 'Grape/Black_rot/'),
name='black rot')
array2, disease2 = get_images(os.path.join(ROOT_DIR, 'Grape/Esca/'),
name='ecsa')
array3, disease3 = get_images(os.path.join(ROOT_DIR, 'Grape/Leaf_blight/'),
name='leaf_blight')
array4, disease4 = get_images(os.path.join(ROOT_DIR, 'Grape/healthy/'),
name='healthy')
array5, disease5 = get_images(os.path.join(ROOT_DIR, 'images/'),
js=json_df_images)
array6, disease6 = get_images(os.path.join(ROOT_DIR, 'positive/'),
js=json_df_positive)
array7, disease7 = get_images(os.path.join(ROOT_DIR, 'healthy/'),
js=json_df_healthy)
array8, disease8 = get_images(os.path.join(ROOT_DIR, 'team4/'),
js=json_df_team4)
array9, disease9 = get_images(os.path.join(ROOT_DIR, 'team4_br/'),
js=json_df_team4_br)
array10, disease10 = get_images(os.path.join(ROOT_DIR, 'leaf_blight/'),
js=json_df_leaf_blight)
# Concatenate data
disease_arr = np.concatenate(
(disease1, disease2, disease3, disease4, disease5, disease6, disease7,
disease8, disease9, disease10),
axis=0)
print('=== TOTAL DATA ===')
print(disease_arr.shape)
img_arr = np.concatenate((array1, array2, array3, array4, array5, array6,
array7, array8, array9, array10),
axis=0)
print(img_arr.shape)
# Shuffle data
img_arr, disease_arr = shuffle(img_arr, disease_arr, random_state=42)
print(np.unique(disease_arr))
# split train set and test set
X_train, X_test, y_train, y_test = train_test_split(img_arr,
disease_arr,
test_size=0.2,
random_state=42)
print('=== TRAIN TEST SPLIT ===')
print(X_test.shape)
print(X_train.shape)
# Save data
make_folder('../data/test')
make_folder('../data/intermediate')
np.save('../data/test/ImageTest_input.npy', X_test)
np.save('../data/test/DiseaseTest_input.npy', y_test)
np.save('../data/intermediate/ImageTrain_input.npy', X_train)
np.save('../data/intermediate/DiseaseTrain_input.npy', y_train)
def main():
""" Load data.
Normalize and encode.
Train custom CNN model.
Print accuracy on test data.
"""
# Load stored data
X_train = np.load('../data/augment/ImageAugment_input.npy')
y_train = np.load('../data/augment/DiseaseAugment_input.npy')
print("=== TRAIN DATA ===")
print(X_train.shape)
print(y_train.shape)
X_test = np.load('../data/test/ImageTest_input.npy')
y_test = np.load('../data/test/DiseaseTest_input.npy')
print("=== TEST DATA ===")
print(X_test.shape)
print(y_test.shape)
# hot encoding of labels
y_train = encoder(y_train, NUM_CLASSES)
y_test = encoder(y_test, NUM_CLASSES)
# Input normalization
X_train = (X_train / 255.0).astype(np.float32)
X_test = (X_test / 255.0).astype(np.float32)
# Custom CNN model
model_custom = Sequential(
(layers.Conv2D(32,
kernel_size=(3, 3),
padding='same',
input_shape=(180, 180, 3)), layers.Activation('relu'),
layers.Conv2D(32, kernel_size=(3, 3), padding='same'),
layers.Activation('relu'), layers.MaxPooling2D(pool_size=(2, 2)),
layers.Conv2D(32, kernel_size=(3, 3),
padding='same'), layers.Activation('relu'),
layers.Conv2D(32, kernel_size=(3, 3), padding='same'),
layers.Activation('relu'), layers.MaxPooling2D(pool_size=(2, 2)),
layers.Conv2D(32, kernel_size=(3, 3),
padding='same'), layers.Activation('relu'),
layers.Conv2D(32, kernel_size=(3, 3), padding='same'),
layers.Activation('relu'), layers.MaxPooling2D(pool_size=(2, 2)),
layers.Conv2D(32, kernel_size=(3, 3),
padding='same'), layers.Activation('relu'),
layers.Conv2D(32, kernel_size=(3, 3), padding='same'),
layers.Activation('relu'), layers.MaxPooling2D(pool_size=(2, 2)),
layers.Conv2D(32, kernel_size=(3, 3),
padding='same'), layers.Activation('relu'),
layers.Conv2D(32, kernel_size=(3, 3),
padding='same'), layers.Activation('relu'),
layers.MaxPooling2D(pool_size=(2, 2)), layers.Flatten(),
layers.Dropout(0.5), layers.Dense(128), layers.Activation('relu'),
layers.Dense(NUM_CLASSES, activation='softmax')))
model_custom.compile(optimizer=Adam(),
loss="categorical_crossentropy",
metrics=['accuracy'])
# Learning rate decay
reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
patience=5,
min_lr=0.00001)
history_custom = model_custom.fit(X_train,
y_train,
batch_size=8,
epochs=1,
verbose=1,
validation_split=.1,
callbacks=[reduce_lr])
scores = model_custom.evaluate(X_test, y_test, verbose=0)
print("========================")
print("TEST SET: %s: %.2f%%" %
(model_custom.metrics_names[1], scores[1] * 100))
print("========================")
print(model_custom.summary())
# save model
make_folder('../results/models/')
model_custom.save('../results/models/custom.h5')
history = dict()
history['acc'] = history_custom.history['acc']
history['val_acc'] = history_custom.history['val_acc']
history['loss'] = history_custom.history['loss']
history['val_loss'] = history_custom.history['val_loss']
hist = Hist()
setattr(hist, 'history', history)
pickle.dump(hist,
open('../results/models/custom_training_history.pkl', 'wb'))
def main():
""" Load data.
Normalize and encode.
Train CNN-VGG16 model.
Print accuracy on test data.
"""
# Load stored data
X_train = np.load('../data/augment/ImageAugment_input.npy')
y_train = np.load('../data/augment/DiseaseAugment_input.npy')
print("=== TRAIN DATA ===")
print(X_train.shape)
print(y_train.shape)
X_test = np.load('../data/test/ImageTest_input.npy')
y_test = np.load('../data/test/DiseaseTest_input.npy')
print("=== TEST DATA ===")
print(X_test.shape)
print(y_test.shape)
# hot encoding of labels
y_train = encoder(y_train, NUM_CLASSES)
y_test = encoder(y_test, NUM_CLASSES)
# Input normalization
X_train = (X_train / 255.0).astype(np.float32)
X_test = (X_test / 255.0).astype(np.float32)
# VGG16 CNN model
IMG_SHAPE = (180, 180, 3)
VGG16_MODEL = tf.keras.applications.VGG16(input_shape=IMG_SHAPE,
include_top=False,
weights='imagenet')
VGG16_MODEL.trainable = False
global_average_layer = tf.keras.layers.GlobalAveragePooling2D()
prediction_layer = Dense(NUM_CLASSES, activation='softmax')
model_vgg16 = Sequential([
VGG16_MODEL,
Conv2D(512, kernel_size=(3, 3), padding='same'),
Activation('relu'),
Conv2D(1024, kernel_size=(3, 3), padding='same'), global_average_layer,
prediction_layer
])
model_vgg16.compile(optimizer=Adam(),
loss="categorical_crossentropy",
metrics=['accuracy'])
# Learning rate decay
reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
patience=5,
min_lr=0.00001)
history_custom = model_vgg16.fit(X_train,
y_train,
batch_size=8,
epochs=20,
verbose=1,
validation_split=.1,
callbacks=[reduce_lr])
scores = model_vgg16.evaluate(X_test, y_test, verbose=0)
print("========================")
print("TEST SET: %s: %.2f%%" %
(model_vgg16.metrics_names[1], scores[1] * 100))
print("========================")
print(model_vgg16.summary())
print("=== BASE MODEL SUMMARY ===")
print(VGG16_MODEL.summary())
# save model
make_folder('../results/models')
model_vgg16.save('../results/models/vgg16.h5')
history = dict()
history['acc'] = history_custom.history['acc']
history['val_acc'] = history_custom.history['val_acc']
history['loss'] = history_custom.history['loss']
history['val_loss'] = history_custom.history['val_loss']
hist = Hist()
setattr(hist, 'history', history)
pickle.dump(hist, open('../results/models/vgg16_training_history.pkl',
'wb'))