def create_model(resize_dim_: Tuple[int, int],
lr_: float,
loss_function_name_: str,
object_storing_dir_: str,
num_filters_first_level_: int = 32,
num_classes_: int = 1,
dropout_rate_: float = 0.2,
use_batch_norm_: bool = True,
**loss_kwargs):
"""
Create 2D Unet model that will be used to predict segmentation masks on single slices of the 3D scan.
Returns a keras model that predicts single channel masks on single channel 2D images
:param resize_dim_: 2D dimensions of the images that the model cn predict on.
:param lr_: learning rate
:param loss_function_name_: name of loss function to use
:param object_storing_dir_: Directory where model objects will be stored. In this case, the model's summary.
:param num_filters_first_level_:
:param num_classes_:
:param dropout_rate_:
:param use_batch_norm_:
:param loss_kwargs: In case the loss function takes arguments, add them as a dictionary
:return: Keras.model 2D Unet
"""
# Build model with `custom_unet` library
model = custom_unet(input_shape=resize_dim_ + (1, ),
use_batch_norm=use_batch_norm_,
num_classes=num_classes_,
filters=num_filters_first_level_,
dropout=dropout_rate_,
output_activation='sigmoid')
with open(os.path.join(object_storing_dir_, 'model_summary.txt'),
'w') as f:
model.summary(print_fn=lambda x: f.write(x + '\n'))
# Get loss function to use
loss_function = training_utils.get_loss_function(loss_function_name_,
**loss_kwargs)
# Compile the model
model.compile(optimizer=Adam(learning_rate=lr_),
loss=loss_function,
metrics=[iou, iou_thresholded])
return model
def __init__(self,
inputShape=(512, 512, 3),
modelWeights=None,
useGPU=True,
dropout=0.05):
self.inputShape = inputShape
self.model = custom_unet(input_shape=inputShape,
use_batch_norm=False,
filters=64,
dropout=dropout,
num_layers=4,
output_activation='sigmoid')
self.useGPU = useGPU
if modelWeights is not None:
try:
self.model.load_weights(modelWeights)
self.trained = True
except IOError as e:
print(e)
self.trained = False
else:
self.trained = False
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
# load data
preprocess_input = None # sm.get_preprocessing('resnet18')
X_train_gen, Y_train_gen, X_test, Y_test = get_data(config.data_folder,
config.batch_size,
config.test_split_ratio,
preprocess_input)
# load model or create new
if config.load_model:
model = load_model(str(config.load_model), compile=False)
else:
model = custom_unet(input_shape=config.input_shape,
use_batch_norm=True,
num_classes=1,
filters=4,
dropout=0.2,
output_activation='sigmoid')
# model = sm.Unet('resnet18', input_shape=config.input_shape, classes=1, activation='sigmoid',
# decoder_filters=(256, 128, 64, 32, 16), encoder_freeze=False, encoder_weights='imagenet')
opt = Adam(lr=config.learning_rate)
model.compile(opt,
sm.losses.bce_jaccard_loss,
metrics=[
'mean_squared_error', sm.metrics.iou_score,
sm.losses.bce_jaccard_loss, sigmoid_iou_loss
])
# print network info
model.summary()
X_test = X_test / 255
#DISPLAYING THE IMAGE AND THE CORRESPONDING MASK
img_new = random.randint(0, len(train_ids))
io.imshow(X_train[img_new])
plt.show()
io.imshow(np.squeeze(Y_train[img_new]))
plt.show()
#IMPORTING KERAS UNET MODEL
from keras_unet.models import custom_unet
model = custom_unet(input_shape=(128, 128, 3),
use_batch_norm=False,
num_classes=1,
filters=128,
dropout=0.2,
output_activation='sigmoid')
model.summary()
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
results = model.fit(X_train, Y_train, validation_split=0.1, epochs=20)
from keras_unet.utils import plot_segm_history
plot_segm_history(results,
metrics=['accuracy', 'val_accuracy'],
print(xx.shape, yy.shape)
from keras_unet.utils import plot_imgs
plot_imgs(org_imgs=xx, mask_imgs=yy, nm_img_to_plot=2, figsize=6)
# %% [markdown]
# ## Initialize network
# %%
from keras_unet.models import custom_unet
input_shape = x_train[0].shape
model = custom_unet(input_shape,
filters=32,
use_batch_norm=True,
dropout=0.3,
dropout_change_per_layer=0.0,
num_layers=4)
# %%
model.summary()
# %% [markdown]
# ## Compile + train
# %%
from tensorflow.keras.callbacks import ModelCheckpoint
model_filename = 'segm_model_v3.h5'
callback_checkpoint = ModelCheckpoint(
model_filename,
def main():
#trainiing settings
#Multi GPU set up
strategy = tf.distribute.MirroredStrategy()
#load dataset
skin_path = 'Data/skin_lesions'
img_path = skin_path + '/image/'
msk_path = skin_path + '/mask/'
#get possible images and masks
imgs = os.listdir(img_path)
msks = os.listdir(msk_path)
#sort the images and masks
msks = sorted(msks)
imgs = sorted(imgs)
#read in the images
images,masks = read_in_skin_images(skin_path,msks,imgs)
#preprocess the dataset
for i in range(0,len(masks)):
m = masks[i]
m = m[:,:,0] #binarize the mask
m.reshape(m.shape[0],m.shape[1]) #reshape the binarized mask
m = cv2.resize(m,(256,256)) #resize the skin masks for memory contraints
masks[i] = m
#images
im = images[i]
images[i] = cv2.resize(im,(256,256)) #resize the images for memory constraints
#make Arrays
images = np.asarray(images)
masks = np.asarray(masks)
masks = masks / masks.max() #normalize the masks
masks = masks.reshape(masks.shape[0],masks.shape[1],masks.shape[2],1) #reshape to binarized mask array
#split the data
img_overall_train, img_test, mask_overall_train, mask_test = train_test_split(images, masks, test_size=0.16667, random_state=42)
img_train, img_val, mask_train, mask_val = train_test_split(img_overall_train, mask_overall_train, test_size = 0.166667, random_state = 32)
#data rotation values
data_aug_dict = dict(
# rotation_range=15.,
width_shift_range=0.05,
height_shift_range=0.05,
shear_range=0.05,
zoom_range=0.05,
horizontal_flip=True,
vertical_flip=True,
fill_mode='nearest'
)
#data generator
train_datagen = ImageDataGenerator(data_aug_dict)
train_generator = train_datagen.flow(
img_train, mask_train,
batch_size=16)
val_generator = train_datagen.flow(
img_val, mask_val)
#standardize steps per epoch for the dataset and batch size
STEPS_PER_EPOCH = len(img_train) // 16
#Get U Net
input_shape = img_train[0].shape
#use multi GPU strategy to run the custom U-Net
with strategy.scope():
#run the custom model with parameters you want to analyze
model = custom_unet(
input_shape,
filters=64,
use_batch_norm=False,
use_dropout_on_upsampling = False,
dropout=0.55,
activation = 'relu',
dropout_change_per_layer=0.00,
num_layers=4,
decoder_type = 'simple'
)
##Compile and Train
#save model with this name **Change model name for each run to save model**
model_filename = 'basic_600_BDice_LR001_BASIC_SKIN_UNET.h5'
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
opt = keras.optimizers.Adam(learning_rate=0.0001)
model.compile(
optimizer=opt,
#optimizer=SGD(lr=0.01, momentum=0.99),
loss= bce_dice_loss,
metrics=['mse',iou, iou_thresholded],
)
history = model.fit_generator(
train_generator,
steps_per_epoch=STEPS_PER_EPOCH,
epochs=600,
validation_data=(img_val, mask_val),
callbacks=[callback_checkpoint]
)
# serialize model to JSON
model_json = model.to_json()
with open("bcdloss_LGG_basic_UNetModel.json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("LGG_basic_UNetmodel.h5")
print("Saved model to disk")
def main():
train_x = np.empty(0, dtype='uint8')
train_y = np.empty(0, dtype='uint8')
batch_size = 4
x_dir = 'DataForSegmentator/input'
y_dir = 'DataForSegmentator/output_filters'
samples_num = len(os.listdir(x_dir))
for x_path, y_path in zip(sorted(os.listdir(x_dir)),
sorted(os.listdir(y_dir))):
train_x = np.append(train_x,
np.array(Image.open("%s/%s" % (x_dir, x_path))))
train_y = np.append(train_y,
np.array(Image.open("%s/%s" % (y_dir, y_path))))
train_x.shape = (samples_num, 256, 256, 3)
train_y.shape = (samples_num, 256, 256, 3)
if not os.path.isfile('train_unet_x.npy'):
np.save('train_unet_x', train_x)
train_x = np.memmap('train_unet_x.npy', shape=train_x.shape, offset=128)
x_generator = ImageDataGenerator(
shear_range=0.1,
zoom_range=0.1,
horizontal_flip=True,
vertical_flip=True,
) \
.flow(
x=train_x,
batch_size=batch_size,
seed=42
)
y_generator = ImageDataGenerator(
shear_range=0.1,
zoom_range=0.1,
horizontal_flip=True,
vertical_flip=True,
) \
.flow(
x=train_y,
batch_size=batch_size,
seed=42
)
train_generator = zip(x_generator, y_generator)
model = custom_unet(input_shape=train_x.shape[1:],
use_batch_norm=False,
num_classes=3,
filters=32,
dropout=0.2,
output_activation='relu')
# model = get_full_model(json_path='models/model_unet_70.json', h5_path='models/model_unet_70.h5')
model.compile(optimizer=Adam(lr=1e-8), loss='mae', metrics=['accuracy'])
callbacks = [
ModelCheckpoint("model_unet.h5",
monitor='acc',
verbose=True,
save_best_only=True),
# EarlyStopping(monitor='acc',
# patience=0,
# baseline=90,
# verbose=True,
# ),
]
full_history = {"acc": np.empty(0), "loss": np.empty(0)}
continue_train = True
epochs_sum = 0
while continue_train:
epochs = get_input_int("How many epochs?", 0, 100)
history = model.fit_generator(
generator=train_generator,
steps_per_epoch=int(samples_num / batch_size),
# validation_steps=train_x.shape[0] / batch_size,
epochs=epochs,
shuffle=True,
verbose=True,
callbacks=callbacks,
)
if epochs != 0:
full_history['acc'] = np.append(full_history['acc'],
history.history['accuracy'])
full_history['loss'] = np.append(full_history['loss'],
history.history['loss'])
epochs_sum = len(full_history['acc'])
#####################################################################
# ----------------------- evaluate model ----------------------------
#####################################################################
print("\nacc %.2f%%\n" %
(history.history['accuracy'][-1] * 100),
end='')
epochs = len(history.history['accuracy'])
plot_history_separate_from_dict(history_dict=full_history,
save_path_acc=None,
save_path_loss=None,
show=True,
save=False)
print("epochs: %d - %d" % (epochs_sum - epochs, epochs_sum))
#####################################################################
# ----------------------- CMD UI ------------------------------------
#####################################################################
if get_stdin_answer("Show image of prediction?"):
show_predict_on_window(
train_x, train_y,
np.array(model.predict(train_x), dtype='uint8'))
if get_stdin_answer(text='Save model?'):
save_model_to_json(model,
"models/model_unet_%d.json" % (epochs_sum))
model.save_weights('models/model_unet_%d.h5' % (epochs_sum))
continue_train = get_stdin_answer(text="Continue?")
from keras.models import Sequential
from keras.layers import Conv2D, MaxPool2D, UpSampling2D
from keras.models import load_model
import cv2
import numpy as np
#edit with your model
IMAGE_SIZE = (480, 480)
from keras_unet.models import custom_unet
model = custom_unet(input_shape=(480, 480, 3),
use_batch_norm=True,
num_classes=1,
filters=16,
dropout=0,
num_layers=4,
output_activation='sigmoid')
model.load_weights('my_model_final_in_480_f_16.h5')
import glob
path = glob.glob("Test/Input/*.jpg")
for myfile in path:
test_im = cv2.imread(myfile)
true_size = test_im.shape
imshow_size = (512, round(true_size[0] * 512 / true_size[1]))
#cv2.imshow('Input',cv2.resize(test_im, imshow_size))
rotation_range=180,
width_shift_range=0.3,
height_shift_range=0.3,
shear_range=0.1,
zoom_range=0.1,
horizontal_flip=True,
vertical_flip=True,
fill_mode='constant',
))
from keras_unet.models import custom_unet
model = custom_unet(
x_train[0].shape,
use_batch_norm=True,
num_classes=1,
filters=64,
dropout=0.2,
output_activation='sigmoid',
)
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
callbacks = [
EarlyStopping(monitor='val_loss', patience=3, verbose=1, min_delta=1e-4),
ModelCheckpoint('./lesson2-model/model.h5',
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='auto'),
]
def main():
#trainiing settings
#Multi GPU strategy
strategy = tf.distribute.MirroredStrategy()
#load dataset
image_paths, mask_paths = load_images(
"Data/lgg-mri-segmentation/kaggle_3m/")
#check to see if each image has a mask if so, read in the images
if image_mask_check(image_paths, mask_paths):
masks, images = read_in_images(image_paths, mask_paths)
#reshape mask image to (256,256) -- binarize
for i in range(0, len(masks)):
m = masks[i]
m = m[:, :, 0]
m.reshape(m.shape[0], m.shape[1])
masks[i] = m
#make images and masks into numpy arrays and normalize the masks
images = np.asarray(images)
masks = np.asarray(masks)
masks = masks / 255
masks = masks.reshape(masks.shape[0], masks.shape[1], masks.shape[2], 1)
#split the data
img_overall_train, img_test, mask_overall_train, mask_test = train_test_split(
images, masks, test_size=0.16667, random_state=42)
img_train, img_val, mask_train, mask_val = train_test_split(
img_overall_train,
mask_overall_train,
test_size=0.166667,
random_state=32)
#create data generator for running model
#batch size of 16
train_datagen = ImageDataGenerator()
train_generator = train_datagen.flow(img_train, mask_train, batch_size=16)
val_generator = train_datagen.flow(img_val, mask_val)
#steps within each epoch is the total number of training images divided by the batch size
STEPS_PER_EPOCH = len(img_train) // 16
#Get U Net
input_shape = img_train[0].shape
#allow model to be trained over multiple GPUs
with strategy.scope():
#load in the custom u net from the models section
model = custom_unet(input_shape,
filters=32,
use_batch_norm=True,
dropout=0.3,
dropout_change_per_layer=0.0,
num_layers=4)
print(model.summary())
##Compile and Train
#save each model if validation loss improves
model_filename = 'UNet_Model'
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
opt = keras.optimizers.Adam(learning_rate=0.001)
#set compile and run for model
model.compile(
optimizer=opt,
#loss='binary_crossentropy',
loss=bce_dice_loss,
metrics=[iou, iou_thresholded])
#use fit_generator because using data generator
#fit the model for
history = model.fit_generator(train_generator,
steps_per_epoch=STEPS_PER_EPOCH,
epochs=300,
validation_data=(img_val, mask_val),
callbacks=[callback_checkpoint])
# serialize model to JSON
#save model as JSON and weight file
model_json = model.to_json()
with open("UNetModel.json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("model.h5")
print("Saved model to disk")
def main():
#trainiing settings
#Multi GPU strategy
strategy = tf.distribute.MirroredStrategy()
#load dataset
image_paths, mask_paths = load_images(
"Data/lgg-mri-segmentation/kaggle_3m/")
#sort images and masks
image_paths = sorted(image_paths)
mask_paths = sorted(mask_paths)
#read in images and masks
if image_mask_check(image_paths, mask_paths):
masks, images = read_in_MR_images(image_paths, mask_paths)
#basic pre-processing with MR additives
for i in range(0, len(masks)):
m = masks[i]
m = m[:, :, 0] #binarize the masks
m.reshape(m.shape[0], m.shape[1]) #set up binarize shape for masks
masks[i] = m
#MRI Values
im = images[i]
images[i] = normalize_MRIvolume(
im) #normalize the MR images with histogram normalization
#make Arrays
images = np.asarray(images)
masks = np.asarray(masks)
masks = masks / masks.max() #normalize masks
masks = masks.reshape(masks.shape[0], masks.shape[1], masks.shape[2],
1) #set up binarize shape for masks
#split the data
img_overall_train, img_test, mask_overall_train, mask_test = train_test_split(
images, masks, test_size=0.16667, random_state=42)
img_train, img_val, mask_train, mask_val = train_test_split(
img_overall_train,
mask_overall_train,
test_size=0.166667,
random_state=32)
#data generator
train_datagen = ImageDataGenerator()
train_generator = train_datagen.flow(img_train, mask_train, batch_size=16)
val_generator = train_datagen.flow(img_val, mask_val)
#standardize steps per epoch based on image dataset and batch size
STEPS_PER_EPOCH = len(img_train) // 16
# STEPS_PER_EPOCH = 250
#Get U Net
input_shape = img_train[0].shape #input shape for the U-Net
#multiGPU processing
with strategy.scope():
#train custom U-Net model based on parameters you want to train
model = custom_unet(input_shape,
filters=64,
use_batch_norm=False,
dropout=0.55,
dropout_change_per_layer=0.00,
num_layers=4,
decoder_type='simple',
use_dropout_on_upsampling=False,
activation='relu')
##Compile and Train
#save model name for evaluation
model_filename = 'Basic_LGG_aug_UNET_01LR.h5'
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='loss',
save_best_only=True,
)
opt = keras.optimizers.Adam(learning_rate=0.0001)
model.compile(
optimizer=opt,
#optimizer=SGD(lr=0.01, momentum=0.99),
loss=bce_dice_loss,
metrics=['mse', iou, iou_thresholded],
)
history = model.fit_generator(train_generator,
steps_per_epoch=STEPS_PER_EPOCH,
epochs=600,
validation_data=(img_val, mask_val),
callbacks=[callback_checkpoint])
# serialize model to JSON
model_json = model.to_json()
with open("bcdloss_LGG_basic_UNetModel.json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("LGG_basic_UNetmodel.h5")
print("Saved model to disk")
# Add a per-pixel classification layer
outputs = layers.Conv2D(num_classes,
3,
activation="softmax",
padding="same")(x)
# Define the model
model = keras.Model(inputs, outputs)
return model
# Free up RAM in case the model definition cells were run multiple times
#keras.backend.clear_session()
model = custom_unet(input_shape=img_size + (3, ),
use_batch_norm=True,
num_classes=2,
filters=32,
dropout=0.2,
output_activation='softmax')
# Build model
# model = get_model(img_size, num_classes)
model.summary()
#%% Train the model
# Configure the model for training.
# We use the "sparse" version of categorical_crossentropy
# because our target data is integers.
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy")
callbacks = [
keras.callbacks.ModelCheckpoint("insulators-augmented.h5",
def main():
#trainiing settings
#load dataset
image_paths = 'Data/Lung Segmentation/CXR_png/'
mask_paths = 'Data/Lung Segmentation/masks/'
images, masks = load_lung_images(image_paths, mask_paths)
masks, images = read_in_images(images, masks)
for i in range(0, len(masks)):
m = masks[i]
im = images[i]
#resize
m = cv2.resize(masks[i], (1024, 1024))
im = cv2.resize(images[i], (1024, 1024))
#reshape mask
m = m[:, :, 0]
m.reshape(m.shape[0], m.shape[1])
masks[i] = m
images[i] = im
#make Arrays
images = np.asarray(images)
masks = np.asarray(masks)
masks = masks / 255
masks = masks.reshape(masks.shape[0], masks.shape[1], masks.shape[2], 1)
#split the data
img_overall_train, img_test, mask_overall_train, mask_test = train_test_split(
images, masks, test_size=0.16667, random_state=42)
img_train, img_val, mask_train, mask_val = train_test_split(
img_overall_train,
mask_overall_train,
test_size=0.166667,
random_state=32)
#data generator
train_datagen = ImageDataGenerator()
train_generator = train_datagen.flow(img_train, mask_train, batch_size=16)
val_generator = train_datagen.flow(img_val, mask_val)
STEPS_PER_EPOCH = len(img_train) // 16
#Get U Net
input_shape = img_train[0].shape
model = custom_unet(input_shape,
filters=32,
use_batch_norm=True,
dropout=0.3,
dropout_change_per_layer=0.0,
num_layers=4)
print(model.summary())
##Compile and Train
model_filename = 'chest_segm_model_v3.h5'
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
model.compile(
optimizer='adam',
#optimizer=SGD(lr=0.01, momentum=0.99),
loss='binary_crossentropy',
#loss=jaccard_distance,
metrics=[iou, iou_thresholded])
history = model.fit_generator(train_generator,
steps_per_epoch=STEPS_PER_EPOCH,
epochs=75,
validation_data=(img_val, mask_val),
callbacks=[callback_checkpoint])
# serialize model to JSON
model_json = model.to_json()
with open("UNetChestModel.json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("basic_chest_model.h5")
print("Saved model to disk")
def main():
#trainiing settings
#Multi GPU set up
strategy = tf.distribute.MirroredStrategy()
#load dataset
lung_path = 'Data/lung-masks'
img_path = lung_path + '/2d_images/'
msk_path = lung_path + '/2d_masks/'
#list all images and masks
imgs = os.listdir(img_path)
msks = os.listdir(msk_path)
#sort images and masks
msks = sorted(msks)
imgs = sorted(imgs)
#read in images and masks
images, masks = read_in_lung_images(lung_path, msks, imgs)
#set up basic preprocessing
for i in range(0, len(masks)):
m = masks[i]
m = m[:, :, 0] #binarize the masks
m.reshape(m.shape[0], m.shape[1]) #reshape binary masks
m = cv2.resize(m,
(256, 256)) #resize the masks due to memory constraints
masks[i] = m
#images
im = images[i]
images[i] = cv2.resize(
im, (256, 256)) #resize the images due to memory constraints
#make Arrays
images = np.asarray(images)
masks = np.asarray(masks)
masks = masks / masks.max() #normalize the masks
masks = masks.reshape(masks.shape[0], masks.shape[1], masks.shape[2],
1) #reshape the masks to binary set up
#split the data
img_overall_train, img_test, mask_overall_train, mask_test = train_test_split(
images, masks, test_size=0.16667, random_state=42)
img_train, img_val, mask_train, mask_val = train_test_split(
img_overall_train,
mask_overall_train,
test_size=0.166667,
random_state=32)
#data generator
train_datagen = ImageDataGenerator()
train_generator = train_datagen.flow(img_train, mask_train, batch_size=16)
val_generator = train_datagen.flow(img_val, mask_val)
#standardize steps per epoch based on the dataset
STEPS_PER_EPOCH = len(img_train) // 16
#Get U Net
input_shape = img_train[0].shape #get input shape for the U-Net model
#use multiGPU training strategy
with strategy.scope():
#change parameters based on model you want to train
#set up custom u net model
model = custom_unet(input_shape,
filters=64,
use_batch_norm=False,
use_dropout_on_upsampling=False,
dropout=0.55,
activation='relu',
dropout_change_per_layer=0.00,
num_layers=4,
decoder_type='simple')
##Compile and Train
#save model name based on model you are training and want to evaluate
model_filename = 'traditional_600_LR0001_BASIC_LUNG_UNET.h5'
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
opt = keras.optimizers.Adam(learning_rate=0.0001)
model.compile(
optimizer=opt,
#optimizer=SGD(lr=0.01, momentum=0.99),
loss=bce_dice_loss,
metrics=['mse', iou, iou_thresholded],
)
history = model.fit_generator(train_generator,
steps_per_epoch=STEPS_PER_EPOCH,
epochs=600,
validation_data=(img_val, mask_val),
callbacks=[callback_checkpoint])
# serialize model to JSON
model_json = model.to_json()
with open("bcdloss_LGG_basic_UNetModel.json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("LGG_basic_UNetmodel.h5")
print("Saved model to disk")
test_size=0.2,
shuffle=False)
# data augmentation with horizontal and vertical flips
train_gen = get_augmented(X_train,
y_train,
batch_size=2,
data_gen_args=dict(
horizontal_flip=True,
vertical_flip=True,
))
# initilaize network
model = custom_unet(input_shape=X_train[0].shape,
filters=32,
use_batch_norm=True,
num_classes=1,
dropout=0.3,
num_layers=4)
# compile and train
callback_checkpoint = ModelCheckpoint(
MODEL_PATH,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
model.compile(optimizer=Adam(), loss='binary_crossentropy', metrics=[iou])
# train model
history = model.fit_generator(train_gen,
