validation_provider = DataProvider(valid_set, [org_suffix, lab_suffix],
is_pre_load=False,
processor=processor)
u_net = UNet3D(n_class=5, n_layer=4, root_filters=16, use_bn=True)
model = SimpleTFModel(u_net,
org_suffix,
lab_suffix,
dropout=0,
loss_function={'cross-entropy': 1.},
weight_function={'balance'})
optimizer = tf.keras.optimizers.Adam(args.learning_rate)
trainer = Trainer(model)
# train test
result = trainer.train(train_provider,
validation_provider,
epochs=args.epochs,
batch_size=args.batch_size,
mini_batch_size=args.minibatch_size,
output_path=output_path,
optimizer=optimizer,
eval_frequency=args.eval_frequency,
is_save_train_imgs=False,
is_save_valid_imgs=True,
is_rebuilt_path=True)
# eval test & pre load test
# loss_function should like {'name': weight}
# weight_function should like {'name': {'alpha': 1, 'beta':2}} or ['name'] / ('name') / {'name'}
model = SimpleTFModel(unet, org_suffix, lab_suffix, dropout=0, loss_function={'cross-entropy':1.0})
# set learning rate with step decay, every [decay_step] epoch, learning rate = [learning rate * decay_rate]
lr = StepDecayLearningRate(learning_rate=args.learning_rate,
decay_step=50,
decay_rate=0.5,
data_size=train_provider.size,
batch_size=args.batch_size)
# init optimizer with learning rate [lr]
optimizer = tf.keras.optimizers.Adam(lr)
# init trainer
trainer = Trainer(model)
# start training
result = trainer.train(train_provider, validation_provider,
epochs=args.epochs,
batch_size=args.batch_size,
mini_batch_size=args.minibatch_size,
output_path=output_path,
optimizer=optimizer,
learning_rate=lr,
eval_frequency=args.eval_frequency,
is_save_train_imgs=False,
is_save_valid_imgs=True)
# evaluate test data
org_suffix = '_brain.nii.gz'
lab_suffix = '_brain_restore.nii.gz'
train_set = glob.glob('data/distortion_data/train/*_brain.nii.gz')
valid_set = glob.glob('data/distortion_data/validation/*_brain.nii.gz')
test_set1 = glob.glob('data/distortion_data/test/*_brain.nii.gz')
test_set2 = glob.glob('data/distortion_data/new_test/*_brain.nii.gz')
test_result_set2 = glob.glob(
'data/distortion_data/new_test/*_brain_restore.nii.gz')
img_save_path = 'data/out_image/'
u_net = UNet3D(n_class=1, n_layer=3, root_filters=16, use_bn=True)
model = RegressionModel(u_net, org_suffix, lab_suffix, dropout=0)
trainer = Trainer(model)
trainer.restore('results/test3/ckpt/final')
pre = {org_suffix: [('channelcheck', 1)], lab_suffix: [('channelcheck', 1)]}
processor = SimpleImageProcessor(pre=pre)
train_provider = DataProvider(train_set, [org_suffix, lab_suffix],
is_pre_load=False,
processor=processor)
validation_provider = DataProvider(valid_set, [org_suffix, lab_suffix],
is_pre_load=False,
processor=processor)
test_provider = DataProvider(test_set1, [org_suffix, lab_suffix],
is_pre_load=False,
processor=processor)
new_test_provider = DataProvider(test_set2, [org_suffix, lab_suffix],
g_beta=g_beta,
g_lambda=g_lambda,
dropout=dropout)
gen_lr = StepDecayLearningRate(learning_rate=learning_rate,
decay_step=10,
decay_rate=0.8,
data_size=train_provider.size,
batch_size=batch_size)
disc_lr = StepDecayLearningRate(learning_rate=learning_rate,
decay_step=10,
decay_rate=0.8,
data_size=train_provider.size,
batch_size=batch_size)
gen_optimizer = tf.keras.optimizers.Adam(gen_lr)
disc_optimizer = tf.keras.optimizers.Adam(disc_lr)
trainer = Trainer(model)
# train
results = trainer.train(train_provider,
valid_provider,
epochs=epochs,
batch_size=batch_size,
mini_batch_size=mini_batch_size,
output_path=output_path,
optimizer=[gen_optimizer, disc_optimizer],
learning_rate=[gen_lr, disc_lr],
eval_frequency=eval_frequency)
# eval
test_provider = DataProvider(test_list, [org_suffix, age_suffix],
is_pre_load=False,
valid_provider = DataProvider(valid_list, [org_suffix, lab_suffix],
is_pre_load=False,
# temp_dir=output_path,
processor=processor)
# build model
vgg3d = VGG3D(n_layer=5, root_filters=16, use_bn=False)
model = ClsModel(vgg3d, org_suffix, lab_suffix, dropout=0.1)
lr = StepDecayLearningRate(learning_rate=args.learning_rate,
decay_step=10,
decay_rate=0.8,
data_size=train_provider.size,
batch_size=args.batch_size)
optimizer = tf.keras.optimizers.Adam(lr)
trainer = Trainer(model)
# train
results = trainer.train(train_provider, valid_provider,
epochs=args.epochs,
batch_size=args.batch_size,
mini_batch_size=args.minibatch_size,
output_path=output_path,
optimizer=optimizer,
learning_rate=lr,
eval_frequency=args.eval_batch_size)
# eval
test_provider = DataProvider(test_list, [org_suffix, lab_suffix],
is_pre_load=False,
processor=processor)