if TRAIN:
for fold in range(1, folds):
train, test = dl.get_k_split(fold)
tr, val = train_test_split(train, test_size=.15, random_state=43)
gen, test_gen = create_gens(tr, val)
model = UNet3D_Extra(input_shape=(1, 128, 128, 128), n_labels=1)
model.compile(optimizer=keras.optimizers.adam(.001), loss=loss_func)
callbacks = get_callbacks(model_file=main_dr + 'saved_models/Leak-' +
str(fold) + '.h5',
initial_learning_rate=5e-3,
learning_rate_drop=.5,
learning_rate_epochs=None,
learning_rate_patience=10,
verbosity=1,
early_stopping_patience=30)
model.fit_generator(generator=gen,
validation_data=test_gen,
use_multiprocessing=True,
workers=8,
epochs=epochs,
callbacks=callbacks)
#model.save_weights(main_dr + 'saved_models/Leak-' + str(fold) + '.h5')
if EVAL:
model = CNN_3D_AE(input_dims, 200)
model.compile(loss='binary_crossentropy',
optimizer=keras.optimizers.adam(initial_lr))
if load:
model.load_weights(model_path)
print('loaded weights')
model.summary()
gen, test_gen = create_gens(train, test)
callbacks = get_callbacks(model_file=model_path,
initial_learning_rate=initial_lr,
learning_rate_drop=.5,
learning_rate_epochs=None,
learning_rate_patience=100,
verbosity=1,
early_stopping_patience=100)
model.fit_generator(generator=gen,
validation_data=test_gen,
use_multiprocessing=True,
workers=8,
epochs=epochs,
callbacks=callbacks)
preds = model.predict_generator(test_gen)
for i in range(5):
ex = preds[i]
n_classes=1,
shuffle=True,
augment=True,
label_size=5)
test_gen = RN_Generator(data_points=val,
dim=config['RN_input_size'],
batch_size=bs,
n_classes=1,
shuffle=False,
augment=False,
label_size=5)
callbacks = get_callbacks(model_file=main_dr + 'saved_models/RN.h5',
initial_learning_rate=initial_lr,
learning_rate_drop=.5,
learning_rate_epochs=None,
learning_rate_patience=10,
verbosity=1,
early_stopping_patience=50)
#model = load_RN_model('/home/sage/GenDiagFramework/saved_models/RN.h5')
model = RetinaNet_Train()
model.fit_generator(generator=train_gen,
validation_data=test_gen,
use_multiprocessing=True,
workers=8,
epochs=25,
callbacks=callbacks)