def train_validation_set_2(model_input=None, model_output=None):
data_v = DataSources.load_validation_dataset2()
data = DataSources.load_naive_augmented_dataset(DataSources.DataSources.VALIDATION_SET2_NG)
# data = data[:]
run_train(data, data_v, model_input=model_input, model_output=model_output, epochs=50)
def train_300w_3d_helen_naive_augmentations(data_sources: [DataSources.DataSources],
model_input, model_output,
limit=-1):
log.info('train_300w_3d_helen_naive_augmentations::')
data_v: [Data] = DataSources.load_validation_dataset2(recalc_pose=True)
data: [Data] = []
for data_source in data_sources:
data += DataSources.load_naive_augmented_dataset(data_source, limit=limit)
if limit > -1:
np.random.shuffle(data)
data = data[:limit]
run_train(data, data_v, model_input=model_input, model_output=model_output, epochs=30)
def validate_pose_vs_landmarks():
import DataSources
import GenerateTrainingSet
data = DataSources.load_naive_augmented_dataset(
DataSources.DataSources.VALIDATION_SET2_NG)
data = DataSources.load_validation_dataset2(
DataSources.DataSources.VALIDATION_2, recalc_pose=False)
total_theta = 0
for data_ in data:
face_model = GenerateTrainingSet.get_face_model()
rot_mat_orig, _ = cv2.Rodrigues(data_.pose[:3])
rotation_vecs, translation_vecs = GenerateTrainingSet.solve_pnp(
data_.landmarks_2d, face_model.model_TD, face_model)
rot_mat_land, _ = cv2.Rodrigues(rotation_vecs)
theta = Utils.get_theta_between_rot_mats(rot_mat_orig, rot_mat_land)
total_theta += theta
print(np.rad2deg(total_theta / len(data)))
def test_300w_3d_helen1(model_input=None, limit=-1, is_6pos=False, model_name='c2_net'):
log.info('run_validation_set2::')
data: [Data] = DataSources.load_naive_augmented_dataset(DataSources.DataSources._300W_3D_HELEN_NG1, limit=limit)
predict(data, model_input, limit, is_6pos, model_name=model_name)
