def extract_weigths():
from model import create_model_infant_seg
from main import prepare_dirs, setup_tensorflow
prepare_dirs(delete_train_dir=False)
sess, summary_writer = setup_tensorflow()
(tf_t1_input, tf_t2_input, tf_label, aux1_pred, aux2_pred, main_pred,
aux1_loss, aux2_loss, main_loss, final_loss, gene_vars,
main_possibility) = create_model_infant_seg(train_phase=False)
saver = tf.train.Saver()
model_path = tf.train.latest_checkpoint(FLAGS.last_trained_checkpoint)
print('saver restore from:%s' % model_path)
saver.restore(sess, model_path)
print '** after resotre..'
hdf5_data = h5py.File(FLAGS.model_saved_hdf5, 'w')
for op in tf.trainable_variables():
para_name = op.name.split('/')[-1]
if para_name.startswith('weights') or para_name.startswith('biases'):
# layer_name = op.name.split('/')[0]
# para_name = para_name[:-2]
print '>> create data value:', op.name[:-2]
value = sess.run(op)
dset = hdf5_data.create_dataset(op.name[:-2],
value.shape,
dtype=np.float32)
dset[...] = value
hdf5_data.close()
def train():
prepare_dirs(delete_train_dir=True)
sess, summary_writer = setup_tensorflow()
(tf_t1_input, tf_t2_input, tf_label, aux1_pred, aux2_pred, main_pred,
aux1_loss, aux2_loss, main_loss, final_loss, gene_vars,
main_possibility) = create_model_infant_seg(train_phase=True)
train_minimize, learning_rate, global_step = create_optimizers(final_loss)
train_data = TrainData(locals())
train_model(train_data)
def test():
prepare_dirs(delete_train_dir=False)
sess, summary_writer = setup_tensorflow()
# here for test, batch_size of tf_input is 1
(tf_t1_input, tf_t2_input, tf_label, aux1_pred, aux2_pred, main_pred,
aux1_loss, aux2_loss, main_loss, final_loss, gene_vars,
main_possibility) = create_model_infant_seg(train_phase=False)
saver = tf.train.Saver()
model_path = tf.train.latest_checkpoint(FLAGS.last_trained_checkpoint)
print('saver restore from:%s' % model_path)
saver.restore(sess, model_path)
test_data = TestData(locals())
predict_multi_modality_test_images_in_nifti(test_data)