def main(argv):
parser = dan_run_loop.DANArgParser()
parser.set_defaults(data_dir='./data_dir',
model_dir='/output',
data_format='channels_last',
train_epochs=20,
epochs_per_eval=10,
batch_size=64)
flags = parser.parse_args(args=argv[1:])
mean_shape = None
imgs_mean = None
imgs_std = None
flags_trans = {
'train': tf.estimator.ModeKeys.TRAIN,
'eval': tf.estimator.ModeKeys.EVAL,
'predict': tf.estimator.ModeKeys.PREDICT
}
flags.mode = flags_trans[flags.mode]
if flags.mode == tf.estimator.ModeKeys.TRAIN:
mean_shape, imgs_mean, imgs_std = read_dataset_info(flags.data_dir)
def vgg16_model_fn(features, labels, mode, params):
return dan_run_loop.dan_model_fn(features=features,
groundtruth=labels,
mode=mode,
stage=params['dan_stage'],
num_lmark=params['num_lmark'],
model_class=VGG16Model,
mean_shape=mean_shape,
imgs_mean=imgs_mean,
imgs_std=imgs_std,
data_format=params['data_format'],
multi_gpu=params['multi_gpu'])
input_function = flags.use_synthetic_data and get_synth_input_fn(
) or vgg16_input_fn
if flags.mode == tf.estimator.ModeKeys.PREDICT:
faceset = '/media/morzh/ext4_volume/data/Faces/all_in_one/set_004_rect/'
file_img = '2UiNSKC3sGw.jpg'
file_rect = faceset + file_img + '.rect'
# rect = np.loadtxt(file_rect)
input_function = img_input_fn(faceset + file_img, (0, 0, 112, 112),
112, 74)
dan_run_loop.dan_main(flags, vgg16_model_fn, input_function)
def main(argv):
parser = dan_run_loop.DANArgParser()
parser.set_defaults(data_dir='./data_dir',
model_dir='./model_dirl2loss',
data_format='channels_last',
train_epochs=20,
epochs_per_eval=10,
batch_size=64)
flags = parser.parse_args(args=argv[1:])
mean_shape = None
imgs_mean = None
imgs_std = None
flags_trans = {
'train': tf.estimator.ModeKeys.TRAIN,
'eval': tf.estimator.ModeKeys.EVAL,
'predict': tf.estimator.ModeKeys.PREDICT
}
flags.mode = flags_trans[flags.mode]
if flags.mode == tf.estimator.ModeKeys.TRAIN:
mean_shape, imgs_mean, imgs_std = read_dataset_info(flags.data_dir)
def vgg16_model_fn(features, labels, mode, params):
return dan_run_loop.dan_model_fn(features=features,
groundtruth=labels,
mode=mode,
stage=params['dan_stage'],
num_lmark=params['num_lmark'],
model_class=VGG16Model,
mean_shape=mean_shape,
imgs_mean=imgs_mean,
imgs_std=imgs_std,
data_format=params['data_format'],
multi_gpu=params['multi_gpu'])
input_function = flags.use_synthetic_data and get_synth_input_fn(
) or vgg16_input_fn #use vgg16_input_fn
print(flags.use_synthetic_data)
if flags.mode == tf.estimator.ModeKeys.PREDICT:
input_function = video_input_fn(flags.data_dir, 112, flags.num_lmark)
dan_run_loop.dan_main(flags, vgg16_model_fn, input_function)