horizontal_flip=True,
vertical_flip=True,
fill_mode='constant'))
#Plot images with mask overlays
batch = next(train_gen)
xx, yy = batch
print(xx.shape, yy.shape)
plot_imgs(org_imgs=xx, mask_imgs=yy, nm_img_to_plot=5, figsize=6)
# Importing Satellite Unet model from the keras_unet package
input_shape = x_train[0].shape
model = satellite_unet(input_shape,
num_classes=1,
output_activation='sigmoid',
num_layers=4)
#Creating Checkpoint
model_filename = '/model.h5'
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
#Compiling the model with appropriate optimizer
model.compile(optimizer=SGD(lr=0.01, momentum=0.99),
loss='binary_crossentropy',
width_shift_range=0,
height_shift_range=0,
shear_range=0,
channel_shift_range=channel_shift_range,
zoom_range=zoom_range,
horizontal_flip=True,
vertical_flip=True,
fill_mode='constant'))
###########################
# SETUP AND TRAIN NETWORK #
###########################
# Initialize network
input_shape = images_train[0].shape
model = satellite_unet(input_shape)
model.summary()
# Compile & train
if val_ratio != 0:
callback_checkpoint = ModelCheckpoint(model_filename_best,
verbose=1,
monitor='val_loss',
save_best_only=True)
else:
callback_checkpoint = ModelCheckpoint(model_filename_best,
verbose=1,
monitor='loss',
save_best_only=True)
model.compile(
def main():
keras.backend.clear_session()
# The original images should be downloaded to ../../ONLINE_RESULTS_DIRECTORY so just download the annotation_results folder from BOKU-drive and place it correspondingly into your FS
if not os.path.exists(VAL_IMAGE_PATH) or not os.path.exists(
TRAIN_MASK_PATH):
# crop images according to our hierarchy: only ears in plots are considered
prepare_images(new_image_size=NEW_IMAGE_SIZE_2d,
path=os.path.join("..", "..", ONLINE_RESULTS_DIRECTORY),
mask_path=EAR_LABEL_IMAGE)
# get all image-names
all_images = get_all_resized_image_names(parent_directory=os.path.join(
DATA_PATH, RESIZED),
file_ending_original=".jpg")
#split the images into train / val-set
train_set, test_set = split_train_set(all_images, 0.2)
save_train_and_test_set(train_set, test_set)
print("restart the program")
# don't know why but isnt it funny?
return
# create 4 image generators, because it only runs like this. Could be improved but...
train_image_Generator = kunet.utils.ImageDataGenerator(
horizontal_flip=True, rescale=(1.0 / 255))
train_mask_Generator = kunet.utils.ImageDataGenerator(horizontal_flip=True)
val_image_Generator = kunet.utils.ImageDataGenerator(rescale=(1.0 / 255))
val_mask_Generator = kunet.utils.ImageDataGenerator()
train_image_generator = train_image_Generator.flow_from_directory(
os.path.join(IMG_PATH, "train"),
batch_size=BATCH_SIZE,
target_size=NEW_IMAGE_SIZE_2d,
seed=412,
class_mode='input')
train_mask_generator = train_mask_Generator.flow_from_directory(
os.path.join(MASK_PATH, "train"),
batch_size=BATCH_SIZE,
target_size=NEW_IMAGE_SIZE_2d,
color_mode="grayscale",
seed=412,
class_mode='input')
val_image_generator = val_image_Generator.flow_from_directory(
os.path.join(IMG_PATH, "val"),
batch_size=BATCH_SIZE,
target_size=NEW_IMAGE_SIZE_2d,
seed=42,
class_mode='input')
val_mask_generator = val_mask_Generator.flow_from_directory(
os.path.join(MASK_PATH, "val"),
batch_size=BATCH_SIZE,
target_size=NEW_IMAGE_SIZE_2d,
color_mode="grayscale",
seed=42,
class_mode='input')
### look into the generator:
""" x, y = train_image_generator.next()
plt.imshow(x[0])
plt.show()
return"""
train_generator = zip(train_image_generator, train_mask_generator)
val_generator = zip(val_image_generator, val_mask_generator)
# creates a new model.h5-file (and overwrites the old!!!!)
if not CONTINUE_LEARNING:
print("attention I will start learning from 0!")
model = satellite_unet(NEW_IMAGE_SIZE, num_classes=1)
# model = unet()
# model = vanilla_unet(NEW_IMAGE_SIZE, num_classes=1, filters=FILTER_SIZE, num_layers=NUM_LAYERS)
opt = Adam(lr=1E-5, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
"""Arguments
lr: float >= 0. Learning rate.
rho: float >= 0.
epsilon: float >= 0. Fuzz factor.
decay: float >= 0. Learning rate decay over each update."""
# my_optimizer = rmsprop(lr=1, rho=0.9, epsilon=1e-08, decay=0.05)
model.compile(optimizer=opt,
loss=jaccard_distance,
metrics=jaccard_coef)
else:
print("I WILL USE THE ALREADY LEARNED WEIGHTS from output.h5")
model = keras.models.load_model(os.path.join(DATA_PATH,
"weights/output.h5"),
custom_objects={
'jaccard_distance':
jaccard_distance,
'jaccard_coef': jaccard_coef
})
model.load_weights(os.path.join(DATA_PATH, "weights/output.h5"))
#backup after each epoch
checkpoint = keras.callbacks.ModelCheckpoint(os.path.join(
WEIGHTS_PATH, "output.h5"),
monitor='loss',
verbose=1)
#nice to have
csv_logger = keras.callbacks.CSVLogger(os.path.join(
WEIGHTS_PATH, "log.out"),
append=True,
separator=';')
#never reached this condition but could become handy. But probably one has to change min_delta and mode?
earlystopping = keras.callbacks.EarlyStopping(monitor='loss',
verbose=1,
min_delta=0.01,
patience=3,
mode='min')
# mode= min for mse, max for non-error-functions
model.summary()
#prints a summary
model.fit(train_generator,
epochs=NO_OF_EPOCHS,
steps_per_epoch=(NO_OF_TRAINING_IMAGES // BATCH_SIZE),
validation_data=val_generator,
validation_steps=(NO_OF_VAL_IMAGES // BATCH_SIZE),
callbacks=[checkpoint, csv_logger, earlystopping])
# when learning is finished, model gets saved in model.h5
model.save(os.path.join(DATA_PATH, "model.h5"))
return
from keras_unet.models import satellite_unet
from keras.callbacks import ModelCheckpoint
from keras.optimizers import Adam, SGD
from keras_unet.metrics import iou, iou_thresholded
from keras import metrics
from keras.preprocessing.image import load_img, img_to_array
from keras.models import Model
import tensorflow as tf
#Define model here
input_shape = (512, 512, 3)
model = satellite_unet(
input_shape,
#use_batch_norm=False,
num_classes=1,
#filters=64,
#dropout=0.2,
output_activation='sigmoid',
num_layers=4)
#Model checkpoints
model_filename = ''
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
#Compile model here