def compute_flow_and_disp(save_path, data_path, model_path_disp,
model_path_flow):
if not os.path.exists(save_path):
os.makedirs(save_path)
file_list = sorted(os.listdir(data_path + 'image_2'))
for i in range(len(file_list)):
import netdef_slim as nd
img_t0_L = data_path + 'image_2/' + file_list[i]
img_t0_R = data_path + 'image_3/' + file_list[i]
dispnet_controller = nd.load_module(model_path_disp +
'controller.py').Controller()
data = dispnet_controller.net_actions(net_dir=model_path_disp).eval(
img_t0_L, img_t0_R)
disp = data['disp.L'][0, 0, :, :]
writeFloat(save_path + file_list[i] + '.disp.L.float3', disp)
flownet_controller = nd.load_module(model_path_flow +
'controller.py').Controller()
if i < len(file_list) - 1:
img_t1_L = data_path + 'image_2/' + file_list[i + 1]
flow_fwd = flownet_controller.net_actions(
net_dir=model_path_flow).eval(img_t0_L, img_t1_L)
for key, value in flow_fwd.items():
if 'flow' in key:
writeFlow(save_path + file_list[i] + '.' + key + '.flo',
value[0, :, :, :].transpose(1, 2, 0))
else:
writeFloat(
save_path + file_list[i] + '.' + key + '.float3',
value[0, :, :, :].transpose(1, 2, 0))
flow_bwd = flownet_controller.net_actions(
net_dir=model_path_flow).eval(img_t1_L, img_t0_L)
for key, value in flow_bwd.items():
if 'flow' in key:
writeFlow(
save_path + file_list[i] + '.' +
key.replace('fwd', 'bwd') + '.flo',
value[0, :, :, :].transpose(1, 2, 0))
else:
writeFloat(
save_path + file_list[i] + '.' +
key.replace('fwd', 'bwd') + '.float3',
value[0, :, :, :].transpose(1, 2, 0))
def __init__(self, path=None, gpu_id=0, quiet=False):
if path is None:
path = self.base_path
self._gpu_id = gpu_id
self._path = path
self._quiet = quiet
parts = list(os.path.normpath(self._path).split('/'))
self._name = parts[-1]
self._train_dir = self.path('training')
self._scratch_dir = 'scratch/%s' % uuid.uuid4()
self._scratch_log_file = self._scratch_dir + '/log.txt'
self._net_config_file = self.path('config.py')
os.environ['NETDEF_QUIET'] = str(quiet)
nd.set_quiet(quiet)
nd.load_module(self._net_config_file)
def __init__(self):
c = nd.load_module(
'/home/netdef/netdef_models/DispNet3/css/controller.py'
).Controller()
self.disp_net = c.net_actions(net_dir=c.base_path)
self.bridge = CvBridge()
self.images_queue = []
self.disparity_image_publisher = rospy.Publisher('/nn/debug', Image)
self.disparity_publisher = rospy.Publisher('/nn/depth', DisparityImage)
self.right_camera_info_subscriber = rospy.Subscriber(
'/camera/infra2/camera_info', CameraInfo,
self.on_camera_info_received)
self.right_camera_info_publisher = rospy.Publisher(
'/camera/infra2/modified_camera_info', CameraInfo)
def __init__(self, net_dir, save_snapshots=True, save_summaries=True):
self._check_evo_manager_init()
self.net = nd.load_module(os.path.join(net_dir, 'net.py'))
self.net_dir = net_dir
height = 480
width = 640
tf.reset_default_graph()
last_evo, current_evo = nd.evo_manager.get_status()
self.env = self.net.get_env()
print('Evolution: ' + last_evo.path())
self.eval_out = self.env.make_eval_graph(
width=width,
height=height,
)
self.session = self._create_session()
self.trainer = SimpleTrainer(session=self.session,
train_dir=last_evo.path())
self.session.run(tf.global_variables_initializer())
ignore_vars = []
if len(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope="copy")) > 0:
ignore_vars = [
tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope="copy")[0]
]
state = last_evo.last_state()
self.trainer.load_checkpoint(state.path(), ignore_vars=ignore_vars)
placeholders = tf.get_collection('placeholders')
self.img0 = placeholders[0]
self.img1 = placeholders[1]
def __init__(self, path=None, net_actions=NetActions):
if path is not None:
self.base_path = path
nd.load_module(os.path.join(self.base_path, 'config.py'))
self.net_actions = net_actions
def __init__(self, net_dir, save_snapshots=True, save_summaries=True):
self._check_evo_manager_init()
self.net = nd.load_module(os.path.join(net_dir, 'net.py'))
self.net_dir = net_dir
im = cv2.imread(im_paths[0])# HxWxC
im_h, im_w, im_c = im.shape
x0,y0,w,h = youtube_to_rec(labels[0],im_h,im_w)
init_rbox = [x0,y0,w,h]
#[cx, cy], [w, h] = rect_2_cxy_wh(init_rbox)
#target_pos, target_sz = np.array([cx, cy]), np.array([w, h])
#state = SiamRPN_init(im, target_pos, target_sz, net)
# tracker init
#target_pos, target_sz = np.array([cx, cy]), np.array([w, h])
cv2.rectangle(im, (x0,y0), (x0+w, y0+h), (0, 255, 0), 2)
warped_images = []
c = nd.load_module('/home/jianingq/netdef_models/FlowNetH/Pred-Merged-SS/controller.py').Controller()
for i in range(0,2):
c_net = c.net_actions(net_dir='/home/jianingq/netdef_models/FlowNetH/Pred-Merged-SS')
im1 = im_paths[i]
im2 = im_paths[i+1]
if i == 0:
eval_out, session = c_net.init_eval(im2, im1)
out = c_net.simple_eval(eval_out, session, im2, im1)
for k,v in out.items():
if(k =='flow[0][1].fwd'):
flow_data.append(v[0,:,:,:].transpose(1,2,0))
elif(k == 'iul_entropy[0][1].fwd'):
entropy_data.append(v[0,:,:,:].transpose(1,2,0))
if(k == 'iul_scale[0][1].fwd'):