def test_init(self):
unet_config = UnetConfig(input_size=(16, 16, 3),
filters=10,
dropout=0.6,
batchnorm=False)
unet = Unet(config=unet_config)
unet.compile(loss="binary_crossentropy", metrics=["accuracy"])
unet.summary()
def test_default_init(self):
unet = Unet()
unet.compile()
unet.summary()
# Hyper parameters
smooth = 1e-15
BATCH_SIZE = 32
LR = 1e-3
EPOCHS = 10
MOMENTUM = 0.9
NUM_CLASS = 1
# Data Loader
train_generator = DataLoader(META_DATA_PATH, batch_size=BATCH_SIZE,
abs_image_path=TRAINING_DATA_PATH, abs_mask_path=TRAINING_MASK_PATH,
phase='train', input_size=224, output_size=224)
test_generator = DataLoader(META_DATA_PATH, batch_size=BATCH_SIZE,
abs_image_path=TESTING_DATA_PATH, abs_mask_path=TESTING_MASK_PATH,
phase='test', input_size=224, output_size=224)
# Build model using Unet class
model = Unet(input_shape=(224, 224, 1)).build()
losses = tf.keras.losses.BinaryCrossentropy()
optimizer = tf.keras.optimizers.Adam(learning_rate=LR)
callback = tf.keras.callbacks.LearningRateScheduler(scheduler)
model.compile(optimizer=optimizer, loss=losses, metrics=iou)
# Training model with my custom generator
model.fit_generator(train_generator,
steps_per_epoch=len(train_generator),
epochs=EPOCHS,
callbacks=[callback],
validation_data=test_generator,
validation_steps=len(test_generator))
def lr_scheduler(epoch):
lr = learning_rate
new_lr = lr * 0.1**(epoch // 10)
return max(new_lr, 1e-10)
with tf.device('/device:GPU:0'):
unet = Unet().build(IMAGE_SHAPE)
model_json = unet.to_json()
with open(os.path.join(SEGMENT_RESULT_PATH, 'model.json'), 'w') as f:
f.write(json.dumps(model_json))
unet.compile(loss=loss_func, optimizer=optim, metrics=[monitors])
unet.summary()
# with open(os.path.join(SEGMENT_RESULT_PATH,'model.json'), 'r') as f:
# model_json = json.loads(f.read())
# unet = keras.models.model_from_json(model_json)
augm = {
"gamma": True,
"rotate": True,
"flip": True,
"hiseq": False,
"normal": False,
"invert": False,
"crop": True
}