import _init_paths
from data_helper import readUNetData
from unet import UNet
from config import *
import numpy as np
import cv2
# Read data
(train_x, train_y), (test_x, test_y) = readUNetData()
model = UNet(img_height, img_width, save_path=model_path + unet_model_name)
train_x.astype('float32')
train_y.astype('float32')
# Train
model.compile()
model.train(train_x, train_y)
X, y = [], []
for i in range(1000):
img = np.zeros((IMG_WIDTH, IMG_HEIGHT, 3))
label = np.zeros((IMG_WIDTH, IMG_HEIGHT, 1))
cx, cy = np.random.randint(0, IMG_WIDTH), np.random.randint(0, IMG_HEIGHT)
w, h = np.random.randint(10, 30), np.random.randint(10, 30)
color = (np.random.random(3) * 256).astype(int)
cv2.rectangle(img, (cx, cy), (cx + w, cy + h), color, -1)
cx, cy = np.random.randint(0, IMG_WIDTH), np.random.randint(0, IMG_HEIGHT)
r = np.random.randint(10, 40)
color = (np.random.random(3) * 256).astype(int)
cv2.circle(img, (cx, cy), r, color, -1)
cv2.circle(label, (cx, cy), r, color, -1)
label[label > 0] = 1
X.append(img)
y.append(label)
X, y = np.asarray(X), np.asarray(y)
X = preprocess_input(X)
print X.shape, y.shape
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=42)
# Training.
model = UNet(input_shape=(IMG_WIDTH, IMG_HEIGHT, 3))
model.compile(optimizer=Adam(lr=1e-5),
loss=dice_coef_loss,
metrics=[dice_coef])
model.fit(X_train, y_train, validation_data=(X_test, y_test), nb_epoch=100)
# compile the model
# model.compile(loss='categorical_crossentropy', optimizer='adadelta', metrics=['accuracy'])
# model.compile(loss='categorical_crossentropy', optimizer=optimizers.Adadelta(lr=0.1, rho=0.95, epsilon=1e-08, decay=0.0005), metrics=['accuracy'])
# model.compile(loss = "categorical_crossentropy", optimizer = optimizers.SGD(lr=0.0001, momentum=0.9), metrics=["accuracy"], decay=0.0005)
# model.compile(loss='categorical_crossentropy',
# optimizer=optimizers.RMSprop(lr=0.00003, rho=0.9, epsilon=1e-08, decay=0.000001),
# metrics=['accuracy'])
# model.compile(loss = "categorical_crossentropy",
# optimizer = optimizers.SGD(lr=0.00003, momentum=0.9, nesterov=True, decay=1e-6),
# metrics=["accuracy"],
# decay=0.0005)
model.compile(loss='categorical_crossentropy',
optimizer=optimizers.Nadam(lr=0.000001),
metrics=['accuracy'])
# model.compile(loss = "mean_absolute_error",
# optimizer = optimizers.SGD(lr=0.0001, momentum=0.9, nesterov=True),
# metrics=["accuracy"],
# decay=0.0005)
print_summary(model)
plot_model(model, to_file='model.pdf', show_shapes=True)
history = model.fit_generator(train_generator,
steps_per_epoch=200,
epochs=epochs1,
validation_data=valid_generator,
validation_steps=50,
tblogdir = './logs/' + training_name + '{}'.format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
np.random.seed(19)
########################################
######### MODEL & SESSION ##########
########################################
tf.reset_default_graph()
sess = tf.Session()
graph = tf.get_default_graph()
set_session(sess)
model = UNet((patchsize, patchsize, 3), filters=features, blocks=blocks)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(adam),
metrics=['accuracy'])
########################################
############# CALLBACKS #############
########################################
os.makedirs(epoch_folder_name, exist_ok=True)
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=tblogdir,
write_graph=False)
tensorboard_callback.set_model(model)
skmetrics = SkMetrics()
skmetrics.set_file_writer(tblogdir)
Callbacks = [
filenames = os.listdir(train_df)
random.shuffle(filenames)
n_validation = 2560
train_filenames = filenames[n_validation:]
validation_filenames = filenames[:n_validation]
print('number of train samples=', len(train_filenames))
print('number of validation samples=', len(validation_filenames))
BATCH_SIZE = 16
IMAGE_SIZE = 320
N_EPOCH = 5
# build model
model = UNet(input_shape=(IMAGE_SIZE,IMAGE_SIZE,1))
model.compile(optimizer='adam',
loss=keras.losses.binary_crossentropy,
metrics=[iou_segmentation])
# visualize model
plot_model(UNet, 'model.png', show_shapes=True)
model.summary()
callbacks=[]
callback_lr = LearningRate(lr=0.0001)
callbacks = [callback_lr]
# training
train_df = os.path.join('../input/kaggle/stage_1_train_images')
train_gen = Generator(folder,
train_filenames,
box_locations,
valid_y = sorted(glob(os.path.join(valid_path, "mask", "*.jpg")))
H = 360
W = 640
batch_size = 2
lr = 1e-3
epochs = 200
model_path = "files/model.h5"
csv_path = "files/data.csv"
train_dataset = tf_dataset(train_x, train_y, batch=batch_size)
valid_dataset = tf_dataset(valid_x, valid_y, batch=batch_size)
model = UNet(H, W)
metrics = [dice_coef, iou, MeanIoU(num_classes=2), Recall(), Precision()]
model.compile(loss=dice_loss, optimizer=Adam(lr), metrics=metrics)
model.summary()
callbacks = [
ModelCheckpoint(model_path, verbose=1, save_best_only=True),
# ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=10, min_lr=1e-7, verbose=1),
CSVLogger(csv_path),
TensorBoard(),
EarlyStopping(monitor='val_loss',
patience=50,
restore_best_weights=False),
SGDRScheduler(min_lr=1e-6,
max_lr=1e-3,
steps_per_epoch=np.ceil(epochs / batch_size),
lr_decay=0.9,
cycle_length=5,
mask_list=glob.glob(TRAIN_PATH+'masks/*')
X_train=[]
Y_train=[]
width=1024
for n, id_ in tqdm(enumerate(im_list), total=len(im_list)):
im=cv2.imread(im_list[n])
im=cv2.resize(im,(width,width),interpolation = cv2.INTER_CUBIC)
X_train.append(im)
for n, id_ in tqdm(enumerate(mask_list), total=len(mask_list)):
mask=cv2.imread(mask_list[n],0)
mask=cv2.resize(mask,(width,width),interpolation = cv2.INTER_CUBIC)
Y_train.append(mask)
X_train=np.array(X_train)
Y_train=np.array(Y_train)
Y_train=Y_train.reshape(Y_train.shape+(1,))
model=UNet((width,width))
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['mse',dice_coef])
model.summary()
earlystopper = EarlyStopping(patience=5, verbose=1)
checkpointer = ModelCheckpoint('unet.{epoch:02d}-{val_loss:.2f}.h5',monitor='val_dice_coef', verbose=1, save_best_only=True)
results = model.fit(X_train, Y_train, validation_split=0.1, batch_size=16, epochs=100,
callbacks=[earlystopper, checkpointer])