def prepare_model(self):
"""
first step prepare model
needs to be called by subclass in re-write process
Necessary: subclass needs to init
self._input_stream
"""
if self.config.MODEL_TYPE is 'od':
self.download_model()
self.load_frozen_graph()
self.load_category_index()
elif self.config.MODEL_TYPE is 'dl':
self.download_model()
self.load_frozen_graph()
self.fps = FPS(self.config.FPS_INTERVAL).start()
self._visualizer = Visualizer(self.config).start()
return self
def segmentation(model,config):
detection_graph = model.detection_graph
# fixed input sizes as model needs resize either way
vs = WebcamVideoStream(config.VIDEO_INPUT,640,480).start()
resize_ratio = 1.0 * 513 / max(vs.real_width,vs.real_height)
target_size = (int(resize_ratio * vs.real_width), int(resize_ratio * vs.real_height)) #(513, 384)
tf_config = model.tf_config
fps = FPS(config.FPS_INTERVAL).start()
print("> Starting Segmentaion")
with detection_graph.as_default():
with tf.Session(graph=detection_graph,config=tf_config) as sess:
while vs.isActive():
frame = vs.resized(target_size)
batch_seg_map = sess.run('SemanticPredictions:0',
feed_dict={'ImageTensor:0':
[cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)]})
seg_map = batch_seg_map[0]
#boxes = []
#labels = []
#ids = []
map_labeled = measure.label(seg_map, connectivity=1)
for region in measure.regionprops(map_labeled):
if region.area > config.MINAREA:
box = region.bbox
id = seg_map[tuple(region.coords[0])]
label = config.LABEL_NAMES[id]
#boxes.append(box)
#labels.append(label)
#ids.append(id)
if config.VISUALIZE:
draw_single_box_on_image(frame,box,label,id,config.DISCO_MODE)
vis = visualize_deeplab(frame,seg_map,fps._glob_numFrames,config.MAX_FRAMES,fps.fps_local(),
config.PRINT_INTERVAL,config.PRINT_TH,config.DL_DISPLAY_NAME,
config.VISUALIZE,config.VIS_FPS,config.DISCO_MODE,config.ALPHA)
if not vis:
break
fps.update()
fps.stop()
vs.stop()
def segmentation(model, config):
detection_graph = model.detection_graph
# fixed input sizes as model needs resize either way
vs = WebcamVideoStream(config.VIDEO_INPUT, 640, 480).start()
resize_ratio = 1.0 * 513 / max(vs.real_width, vs.real_height)
target_size = (int(resize_ratio * vs.real_width),
int(resize_ratio * vs.real_height)) #(513, 384)
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
fps = FPS(config.FPS_INTERVAL).start()
print("> Starting Segmentaion")
with detection_graph.as_default():
with tf.Session(graph=detection_graph, config=tf_config) as sess:
while vs.isActive():
frame = vs.resized(target_size)
batch_seg_map = sess.run(
'SemanticPredictions:0',
feed_dict={
'ImageTensor:0':
[cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)]
})
# visualization
if config.VISUALIZE:
seg_map = batch_seg_map[0]
seg_image = create_colormap(seg_map).astype(np.uint8)
cv2.addWeighted(seg_image, config.ALPHA, frame,
1 - config.ALPHA, 0, frame)
vis_text(frame, "fps: {}".format(fps.fps_local()),
(10, 30))
# boxes (ymin, xmin, ymax, xmax)
if config.BBOX:
map_labeled = measure.label(seg_map, connectivity=1)
for region in measure.regionprops(map_labeled):
if region.area > config.MINAREA:
box = region.bbox
p1 = (box[1], box[0])
p2 = (box[3], box[2])
cv2.rectangle(frame, p1, p2, (77, 255, 9), 2)
vis_text(
frame, config.LABEL_NAMES[seg_map[tuple(
region.coords[0])]],
(p1[0], p1[1] - 10))
cv2.imshow(config.DL_MODEL_NAME, frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
fps.update()
fps.stop()
vs.stop()
def detection(model,config):
print("> Building Graph")
# tf Session Config
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth=True
detection_graph = model.detection_graph
category_index = model.category_index
with detection_graph.as_default():
with tf.Session(graph=detection_graph,config=tf_config) as sess:
# start Videostream
vs = WebcamVideoStream(config.VIDEO_INPUT,config.WIDTH,config.HEIGHT).start()
# Define Input and Ouput tensors
tensor_dict = model.get_tensordict(['num_detections', 'detection_boxes', 'detection_scores','detection_classes', 'detection_masks'])
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
# Mask Transformations
fps = FPS(config.FPS_INTERVAL).start()
print('> Starting Detection')
while vs.isActive():
# Detection
if not (config.USE_TRACKER):
# default session
frame = vs.read()
output_dict = sess.run(tensor_dict, feed_dict={image_tensor: vs.expanded()})
num = output_dict['num_detections'][0]
classes = output_dict['detection_classes'][0]
boxes = output_dict['detection_boxes'][0]
scores = output_dict['detection_scores'][0]
if 'detection_masks' in output_dict:
masks = output_dict['detection_masks'][0]
else:
masks = None
# reformat detection
num = int(num)
boxes = np.squeeze(boxes)
classes = np.squeeze(classes).astype(np.uint8)
scores = np.squeeze(scores)
# Visualization
vis = vis_detection(frame, boxes, classes, scores, masks, category_index, fps.fps_local(),
config.VISUALIZE, config.DET_INTERVAL, config.DET_TH, config.MAX_FRAMES,
fps._glob_numFrames, config.OD_MODEL_NAME)
if not vis:
break
fps.update()
# End everything
vs.stop()
fps.stop()
def segmentation(model, config):
detection_graph = model.detection_graph
# fixed input sizes as model needs resize either way
vs = WebcamVideoStream(config.VIDEO_INPUT, 640, 480).start()
resize_ratio = 1.0 * 513 / max(vs.real_width, vs.real_height)
target_size = (int(resize_ratio * vs.real_width),
int(resize_ratio * vs.real_height)) #(513, 384)
tf_config = model.tf_config
fps = FPS(config.FPS_INTERVAL).start()
print("> Starting Segmentaion")
with detection_graph.as_default():
with tf.Session(graph=detection_graph, config=tf_config) as sess:
while vs.isActive():
frame = vs.resized(target_size)
batch_seg_map = sess.run(
'SemanticPredictions:0',
feed_dict={
'ImageTensor:0':
[cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)]
})
seg_map = batch_seg_map[0]
#boxes = []
#labels = []
map_labeled = measure.label(seg_map, connectivity=1)
for region in measure.regionprops(map_labeled):
if region.area > config.MINAREA:
box = region.bbox
label = config.LABEL_NAMES[seg_map[tuple(
region.coords[0])]]
#boxes.append(box)
#labels.append(label)
if config.VISUALIZE:
draw_single_box_on_image(frame, box, label)
vis = visualize_deeplab(
frame, seg_map, fps._glob_numFrames, config.MAX_FRAMES,
fps.fps_local(), config.PRINT_INTERVAL, config.PRINT_TH,
config.OD_MODEL_NAME + config._DEV + config._OPT,
config.VISUALIZE)
if not vis:
break
fps.update()
fps.stop()
vs.stop()
class Model(object):
"""
Base Tensorflow Inference Model Class
"""
def __init__(self,config):
self.config = config
self.detection_graph = tf.Graph()
self.category_index = None
self.masks = None
#self._tf_config = tf.compat.v1.ConfigProto(allow_soft_placement=True)
self._tf_config = tf.ConfigProto(allow_soft_placement=True)
self._tf_config.gpu_options.allow_growth=True
#self._tf_config.gpu_options.force_gpu_compatible=True
#self._tf_config.gpu_options.per_process_gpu_memory_fraction = 0.01
self._run_options = tf.RunOptions(trace_level=tf.RunOptions.NO_TRACE)
self._run_metadata = False
self._wait_thread = False
self._is_imageD = False
self._is_videoD = False
self._is_rosD = False
print ('> Model: {}'.format(self.config.MODEL_PATH))
def download_model(self):
"""
downlaods model from model_zoo
"""
if self.config.MODEL_TYPE == 'dl':
download_base = 'http://download.tensorflow.org/models/'
elif self.config.MODEL_TYPE == 'od':
download_base = 'http://download.tensorflow.org/models/object_detection/'
model_file = self.config.MODEL_NAME + '.tar.gz'
if not os.path.isfile(self.config.MODEL_PATH) and self.config.DOWNLOAD_MODEL:
print('> Model not found. Downloading it now.')
opener = urllib.request.URLopener()
opener.retrieve(download_base + model_file, model_file)
tar_file = tarfile.open(model_file)
for file in tar_file.getmembers():
file_name = os.path.basename(file.name)
if 'frozen_inference_graph.pb' in file_name:
tar_file.extract(file, os.getcwd() + '/models/')
os.remove(os.getcwd() + '/' + model_file)
else:
print('> Model found. Proceed.')
def node_name(self,n):
if n.startswith("^"):
return n[1:]
else:
return n.split(":")[0]
def load_frozen_graph(self):
"""
loads graph from frozen model file
"""
print('> Loading frozen model into memory')
if (self.config.MODEL_TYPE == 'od' and self.config.SPLIT_MODEL):
# load a frozen Model and split it into GPU and CPU graphs
# Hardcoded split points for ssd_mobilenet
tf.reset_default_graph()
if self.config.SSD_SHAPE == 600:
shape = 7326
else:
shape = 1917
self.score = tf.placeholder(tf.float32, shape=(None, shape, self.config.NUM_CLASSES), name=self.config.SPLIT_NODES[0])
self.expand = tf.placeholder(tf.float32, shape=(None, shape, 1, 4), name=self.config.SPLIT_NODES[1])
#self.tofloat = tf.placeholder(tf.float32, shape=(None), name=self.config.SPLIT_NODES[2])
for node in tf.get_default_graph().as_graph_def().node:
if node.name == self.config.SPLIT_NODES[0]:
score_def = node
if node.name == self.config.SPLIT_NODES[1]:
expand_def = node
#if node.name == self.config.SPLIT_NODES[2]:
# tofloat_def = node
with self.detection_graph.as_default():
graph_def = tf.GraphDef()
with tf.gfile.GFile(self.config.MODEL_PATH, 'rb') as fid:
serialized_graph = fid.read()
graph_def.ParseFromString(serialized_graph)
edges = {}
name_to_node_map = {}
node_seq = {}
seq = 0
for node in graph_def.node:
n = self.node_name(node.name)
name_to_node_map[n] = node
edges[n] = [self.node_name(x) for x in node.input]
node_seq[n] = seq
seq += 1
for d in self.config.SPLIT_NODES:
assert d in name_to_node_map, "%s is not in graph" % d
nodes_to_keep = set()
next_to_visit = self.config.SPLIT_NODES[:]
while next_to_visit:
n = next_to_visit[0]
del next_to_visit[0]
if n in nodes_to_keep: continue
nodes_to_keep.add(n)
next_to_visit += edges[n]
nodes_to_keep_list = sorted(list(nodes_to_keep), key=lambda n: node_seq[n])
nodes_to_remove = set()
for n in node_seq:
if n in nodes_to_keep_list: continue
nodes_to_remove.add(n)
nodes_to_remove_list = sorted(list(nodes_to_remove), key=lambda n: node_seq[n])
keep = graph_pb2.GraphDef()
for n in nodes_to_keep_list:
keep.node.extend([copy.deepcopy(name_to_node_map[n])])
remove = graph_pb2.GraphDef()
remove.node.extend([score_def])
remove.node.extend([expand_def])
for n in nodes_to_remove_list:
remove.node.extend([copy.deepcopy(name_to_node_map[n])])
with tf.device('/gpu:0'):
tf.import_graph_def(keep, name='')
with tf.device('/cpu:0'):
tf.import_graph_def(remove, name='')
else:
# default model loading procedure
with self.detection_graph.as_default():
graph_def = tf.GraphDef()
with tf.gfile.GFile(self.config.MODEL_PATH, 'rb') as fid:
serialized_graph = fid.read()
graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(graph_def, name='')
def load_category_index(self):
"""
creates categorie_index from label_map
"""
print('> Loading label map')
label_map = tf_utils.load_labelmap(self.config.LABEL_PATH)
categories = tf_utils.convert_label_map_to_categories(label_map, max_num_classes=self.config.NUM_CLASSES, use_display_name=True)
self.category_index = tf_utils.create_category_index(categories)
def get_tensor_dict(self, outputs):
"""
returns tensordict for given tensornames list
"""
ops = self.detection_graph.get_operations()
all_tensor_names = {output.name for op in ops for output in op.outputs}
self.tensor_dict = {}
for key in outputs:
tensor_name = key + ':0'
if tensor_name in all_tensor_names:
self.tensor_dict[key] = self.detection_graph.get_tensor_by_name(tensor_name)
return self.tensor_dict
def prepare_model(self):
"""
first step prepare model
needs to be called by subclass in re-write process
Necessary: subclass needs to init
self._input_stream
"""
if self.config.MODEL_TYPE is 'od':
self.download_model()
self.load_frozen_graph()
self.load_category_index()
elif self.config.MODEL_TYPE is 'dl':
self.download_model()
self.load_frozen_graph()
self.fps = FPS(self.config.FPS_INTERVAL).start()
self._visualizer = Visualizer(self.config).start()
return self
def isActive(self):
"""
checks if stream and visualizer are active
"""
return self._input_stream.isActive() and self._visualizer.isActive()
def stop(self):
"""
stops all Model sub classes
"""
self._input_stream.stop()
self._visualizer.stop()
self.fps.stop()
if self.config.SPLIT_MODEL and self.config.MODEL_TYPE is 'od':
self._gpu_worker.stop()
self._cpu_worker.stop()
def detect(self):
"""
needs to be written by subclass
"""
self.detection = None
def run(self):
"""
runs detection loop on video or image
listens on isActive()
"""
print("> starting detection")
self.start()
while self.isActive():
# detection
self.detect()
# Visualization
if not self._wait_thread:
self.visualize_detection()
self.fps.update()
self.stop()
def start(self):
"""
starts fps and visualizer class
"""
self.fps.start()
self._visualizer = Visualizer(self.config).start()
def visualize_detection(self):
self.detection = self._visualizer.visualize_detection(self.frame,self.boxes,
self.classes,self.scores,
self.masks,self.fps.fps_local(),
self.category_index,self._is_imageD)
def prepare_ros(self,node):
"""
prepares ros Node and ROSInputstream
only in ros branch usable due to ROS realted package stuff
"""
assert node in ['detection_node','deeplab_node'], "only 'detection_node' and 'deeplab_node' supported"
import rospy
from ros import ROSStream, DetectionPublisher, SegmentationPublisher
self._is_rosD = True
rospy.init_node(node)
self._input_stream = ROSStream(self.config.ROS_INPUT)
if node is 'detection_node':
self._ros_publisher = DetectionPublisher()
if node is 'deeplab_node':
self._ros_publisher = SegmentationPublisher()
# check for frame
while True:
self.frame = self._input_stream.read()
time.sleep(1)
print("...waiting for ROS image")
if self.frame is not None:
self.stream_height,self.stream_width = self.frame.shape[0:2]
break
def prepare_timeliner(self):
"""
prepares timeliner and sets tf Run options
"""
self._run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
self._run_metadata = tf.RunMetadata()
self.timeliner = TimeLiner()
def prepare_tracker(self):
"""
prepares KCF tracker
"""
sys.path.append(os.getcwd()+'/rod/kcf')
import KCF
self._tracker = KCF.kcftracker(False, True, False, False)
self._tracker_counter = 0
self._track = False
def run_tracker(self):
"""
runs KCF tracker on videoStream frame
!does not work on images, obviously!
"""
self.frame = self._input_stream.read()
if self._first_track:
self._trackers = []
self._tracker_boxes = self.boxes
num_tracked = 0
for box in self.boxes[~np.all(self.boxes == 0, axis=1)]:
self._tracker.init(conv_detect2track(box,self._input_stream.real_width,
self._input_stream.real_height),self.tracker_frame)
self._trackers.append(self._tracker)
num_tracked += 1
if num_tracked <= self.config.NUM_TRACKERS:
break
self._first_track = False
for idx,self._tracker in enumerate(self._trackers):
tracker_box = self._tracker.update(self.frame)
self._tracker_boxes[idx,:] = conv_track2detect(tracker_box,
self._input_stream.real_width,
self._input_stream.real_height)
self._tracker_counter += 1
self.boxes = self._tracker_boxes
# Deactivate Tracker
if self._tracker_counter >= self.config.TRACKER_FRAMES:
self._track = False
self._tracker_counter = 0
def activate_tracker(self):
"""
activates KCF tracker
deactivates mask detection
"""
#self.masks = None
self.tracker_frame = self.frame
self._track = True
self._first_track = True
def detection(model,config):
# Tracker
if config.USE_TRACKER:
import sys
sys.path.append(os.getcwd()+'/stuff/kcf')
import KCF
tracker = KCF.kcftracker(False, True, False, False)
tracker_counter = 0
track = False
print("> Building Graph")
# tf Session Config
tf_config = model.tf_config
detection_graph = model.detection_graph
category_index = model.category_index
with detection_graph.as_default():
with tf.Session(graph=detection_graph,config=tf_config) as sess:
# start Videostream
vs = WebcamVideoStream(config.VIDEO_INPUT,config.WIDTH,config.HEIGHT).start()
# Define Input and Ouput tensors
tensor_dict = model.get_tensordict(['num_detections', 'detection_boxes', 'detection_scores','detection_classes', 'detection_masks'])
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
# Mask Transformations
if 'detection_masks' in tensor_dict:
# Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], [0])
detection_masks = tf.squeeze(tensor_dict['detection_masks'], [0])
real_num_detection = tf.cast(tensor_dict['num_detections'][0], tf.int32)
detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1])
detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1])
detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(
detection_masks, detection_boxes, vs.real_height, vs.real_width)
detection_masks_reframed = tf.cast(tf.greater(detection_masks_reframed, 0.5), tf.uint8)
# Follow the convention by adding back the batch dimension
tensor_dict['detection_masks'] = tf.expand_dims(detection_masks_reframed, 0)
if config.SPLIT_MODEL:
score_out = detection_graph.get_tensor_by_name('Postprocessor/convert_scores:0')
expand_out = detection_graph.get_tensor_by_name('Postprocessor/ExpandDims_1:0')
score_in = detection_graph.get_tensor_by_name('Postprocessor/convert_scores_1:0')
expand_in = detection_graph.get_tensor_by_name('Postprocessor/ExpandDims_1_1:0')
# Threading
score = model.score
expand = model.expand
gpu_worker = SessionWorker("GPU",detection_graph,tf_config)
cpu_worker = SessionWorker("CPU",detection_graph,tf_config)
gpu_opts = [score_out, expand_out]
cpu_opts = [tensor_dict['detection_boxes'], tensor_dict['detection_scores'], tensor_dict['detection_classes'], tensor_dict['num_detections']]
gpu_counter = 0
cpu_counter = 0
fps = FPS(config.FPS_INTERVAL).start()
print('> Starting Detection')
while vs.isActive():
# Detection
if not (config.USE_TRACKER and track):
if config.SPLIT_MODEL:
# split model in seperate gpu and cpu session threads
masks = None # No Mask Detection possible yet
if gpu_worker.is_sess_empty():
# read video frame, expand dimensions and convert to rgb
frame = vs.read()
# put new queue
gpu_feeds = {image_tensor: vs.expanded()}
if config.VISUALIZE:
gpu_extras = frame # for visualization frame
else:
gpu_extras = None
gpu_worker.put_sess_queue(gpu_opts,gpu_feeds,gpu_extras)
g = gpu_worker.get_result_queue()
if g is None:
# gpu thread has no output queue. ok skip, let's check cpu thread.
gpu_counter += 1
else:
# gpu thread has output queue.
gpu_counter = 0
score,expand,frame = g["results"][0],g["results"][1],g["extras"]
if cpu_worker.is_sess_empty():
# When cpu thread has no next queue, put new queue.
# else, drop gpu queue.
cpu_feeds = {score_in: score, expand_in: expand}
cpu_extras = frame
cpu_worker.put_sess_queue(cpu_opts,cpu_feeds,cpu_extras)
c = cpu_worker.get_result_queue()
if c is None:
# cpu thread has no output queue. ok, nothing to do. continue
cpu_counter += 1
continue # If CPU RESULT has not been set yet, no fps update
else:
cpu_counter = 0
boxes, scores, classes, num, frame = c["results"][0],c["results"][1],c["results"][2],c["results"][3],c["extras"]
else:
# default session
frame = vs.read()
output_dict = sess.run(tensor_dict, feed_dict={image_tensor: vs.expanded()})
num = output_dict['num_detections'][0]
classes = output_dict['detection_classes'][0]
boxes = output_dict['detection_boxes'][0]
scores = output_dict['detection_scores'][0]
if 'detection_masks' in output_dict:
masks = output_dict['detection_masks'][0]
else:
masks = None
# reformat detection
num = int(num)
boxes = np.squeeze(boxes)
classes = np.squeeze(classes).astype(np.uint8)
scores = np.squeeze(scores)
# Visualization
vis = vis_detection(frame, boxes, classes, scores, masks, category_index, fps.fps_local(),
config.VISUALIZE, config.DET_INTERVAL, config.DET_TH, config.MAX_FRAMES,
fps._glob_numFrames, config.OD_MODEL_NAME)
if not vis:
break
# Activate Tracker
if config.USE_TRACKER and num <= config.NUM_TRACKERS:
tracker_frame = frame
track = True
first_track = True
# Tracking
else:
frame = vs.read()
if first_track:
trackers = []
tracker_boxes = boxes
for box in boxes[~np.all(boxes == 0, axis=1)]:
tracker.init(conv_detect2track(box,vs.real_width, vs.real_height), tracker_frame)
trackers.append(tracker)
first_track = False
for idx,tracker in enumerate(trackers):
tracker_box = tracker.update(frame)
tracker_boxes[idx,:] = conv_track2detect(tracker_box, vs.real_width, vs.real_height)
vis = vis_detection(frame, tracker_boxes, classes, scores, masks, category_index, fps.fps_local(),
config.VISUALIZE, config.DET_INTERVAL, config.DET_TH, config.MAX_FRAMES,
fps._glob_numFrames, config.OD_MODEL_NAME)
if not vis:
break
tracker_counter += 1
#tracker_frame = frame
if tracker_counter >= config.TRACKER_FRAMES:
track = False
tracker_counter = 0
fps.update()
# End everything
vs.stop()
fps.stop()
if config.SPLIT_MODEL:
gpu_worker.stop()
cpu_worker.stop()
class Model(object):
"""
Base Tensorflow Inference Model Class
"""
def __init__(self, config):
self.config = config
self.detection_graph = tf.Graph()
self.category_index = None
self.masks = None
self._tf_config = tf.ConfigProto(allow_soft_placement=True)
self._tf_config.gpu_options.allow_growth = True
self._run_options = tf.RunOptions(trace_level=tf.RunOptions.NO_TRACE)
self._run_metadata = False
self._wait_thread = False
print('> Model: {}'.format(self.config.MODEL_PATH))
def download_model(self):
"""
downlaods model from model_zoo
"""
if self.config.MODEL_TYPE == 'dl':
download_base = 'http://download.tensorflow.org/models/'
elif self.config.MODEL_TYPE == 'od':
download_base = 'http://download.tensorflow.org/models/object_detection/'
model_file = self.config.MODEL_NAME + '.tar.gz'
if not os.path.isfile(
self.config.MODEL_PATH) and self.config.DOWNLOAD_MODEL:
print('> Model not found. Downloading it now.')
opener = urllib.request.URLopener()
opener.retrieve(download_base + model_file, model_file)
tar_file = tarfile.open(model_file)
for file in tar_file.getmembers():
file_name = os.path.basename(file.name)
if 'frozen_inference_graph.pb' in file_name:
tar_file.extract(file, os.getcwd() + '/models/')
os.remove(os.getcwd() + '/' + model_file)
else:
print('> Model found. Proceed.')
def _node_name(self, n):
if n.startswith("^"):
return n[1:]
else:
return n.split(":")[0]
def load_frozenmodel(self):
"""
loads graph from frozen model file
"""
print('> Loading frozen model into memory')
if (self.config.MODEL_TYPE == 'od' and self.config.SPLIT_MODEL):
# load a frozen Model and split it into GPU and CPU graphs
# Hardcoded split points for ssd_mobilenet
input_graph = tf.Graph()
with tf.Session(graph=input_graph, config=self._tf_config):
if self.config.SSD_SHAPE == 600:
shape = 7326
else:
shape = 1917
self.score = tf.placeholder(tf.float32,
shape=(None, shape,
self.config.NUM_CLASSES),
name=self.config.SPLIT_NODES[0])
self.expand = tf.placeholder(tf.float32,
shape=(None, shape, 1, 4),
name=self.config.SPLIT_NODES[1])
for node in input_graph.as_graph_def().node:
if node.name == self.config.SPLIT_NODES[0]:
score_def = node
if node.name == self.config.SPLIT_NODES[1]:
expand_def = node
with self.detection_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile(self.config.MODEL_PATH, 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
edges = {}
name_to_node_map = {}
node_seq = {}
seq = 0
for node in od_graph_def.node:
n = self._node_name(node.name)
name_to_node_map[n] = node
edges[n] = [self._node_name(x) for x in node.input]
node_seq[n] = seq
seq += 1
for d in self.config.SPLIT_NODES:
assert d in name_to_node_map, "%s is not in graph" % d
nodes_to_keep = set()
next_to_visit = self.config.SPLIT_NODES[:]
while next_to_visit:
n = next_to_visit[0]
del next_to_visit[0]
if n in nodes_to_keep: continue
nodes_to_keep.add(n)
next_to_visit += edges[n]
nodes_to_keep_list = sorted(list(nodes_to_keep),
key=lambda n: node_seq[n])
nodes_to_remove = set()
for n in node_seq:
if n in nodes_to_keep_list: continue
nodes_to_remove.add(n)
nodes_to_remove_list = sorted(list(nodes_to_remove),
key=lambda n: node_seq[n])
keep = graph_pb2.GraphDef()
for n in nodes_to_keep_list:
keep.node.extend([copy.deepcopy(name_to_node_map[n])])
remove = graph_pb2.GraphDef()
remove.node.extend([score_def])
remove.node.extend([expand_def])
for n in nodes_to_remove_list:
remove.node.extend(
[copy.deepcopy(name_to_node_map[n])])
with tf.device('/gpu:0'):
tf.import_graph_def(keep, name='')
with tf.device('/cpu:0'):
tf.import_graph_def(remove, name='')
else:
# default model loading procedure
with self.detection_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile(self.config.MODEL_PATH, 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def, name='')
def load_labelmap(self):
"""
creates categorie_index from label_map
"""
print('> Loading label map')
label_map = tf_utils.load_labelmap(self.config.LABEL_PATH)
categories = tf_utils.convert_label_map_to_categories(
label_map,
max_num_classes=self.config.NUM_CLASSES,
use_display_name=True)
self.category_index = tf_utils.create_category_index(categories)
def get_tensordict(self, outputs):
"""
returns tensordict for given tensornames list
"""
ops = self.detection_graph.get_operations()
all_tensor_names = {output.name for op in ops for output in op.outputs}
self.tensor_dict = {}
for key in outputs:
tensor_name = key + ':0'
if tensor_name in all_tensor_names:
self.tensor_dict[
key] = self.detection_graph.get_tensor_by_name(tensor_name)
return self.tensor_dict
def prepare_model(self):
"""
first step prepare model
needs to be called by subclass in re-write process
Necessary: subclass needs to init
self._input_stream
"""
if self.config.MODEL_TYPE is 'od':
self.download_model()
self.load_frozenmodel()
self.load_labelmap()
elif self.config.MODEL_TYPE is 'dl':
self.download_model()
self.load_frozenmodel()
self.fps = FPS(self.config.FPS_INTERVAL).start()
self._visualizer = Visualizer(self.config).start()
return self
def isActive(self):
"""
checks if stream and visualizer are active
"""
return self._input_stream.isActive() and self._visualizer.isActive()
def stop(self):
"""
stops all sub classes
"""
self._input_stream.stop()
self._visualizer.stop()
self.fps.stop()
if self.config.SPLIT_MODEL and self.config.MODEL_TYPE is 'od':
self._gpu_worker.stop()
self._cpu_worker.stop()
def detect(self):
"""
needs to be written by subclass
"""
self.detection = None
def run(self):
"""
runs detection loop on video or image
listens on isActive()
"""
print("> starting detection")
self.start()
while self.isActive():
# detection
self.detect()
# Visualization
if not self._wait_thread:
self.visualize_detection()
self.fps.update()
self.stop()
def start(self):
"""
starts fps and visualizer class
"""
self.fps.start()
self._visualizer = Visualizer(self.config).start()
def visualize_detection(self):
self.detection = self._visualizer.visualize_detection(
self.frame, self.boxes, self.classes, self.scores, self.masks,
self.fps.fps_local(), self.category_index)
def detection(model,config):
# Tracker
if config.USE_TRACKER:
import sys
sys.path.append(os.getcwd()+'/rod/kcf')
import KCF
tracker = KCF.kcftracker(False, True, False, False)
tracker_counter = 0
track = False
print("> Building Graph")
# tf Session Config
tf_config = model.tf_config
detection_graph = model.detection_graph
category_index = model.category_index
with detection_graph.as_default():
with tf.Session(graph=detection_graph,config=tf_config) as sess:
# start Videostream
vs = WebcamVideoStream(config.VIDEO_INPUT,config.WIDTH,config.HEIGHT).start()
# Define Input and Ouput tensors
tensor_dict = model.get_tensordict(['num_detections', 'detection_boxes', 'detection_scores','detection_classes', 'detection_masks'])
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
# Mask Transformations
if 'detection_masks' in tensor_dict:
# Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], [0])
detection_masks = tf.squeeze(tensor_dict['detection_masks'], [0])
real_num_detection = tf.cast(tensor_dict['num_detections'][0], tf.int32)
detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1])
detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1])
detection_masks_reframed = reframe_box_masks_to_image_masks(
detection_masks, detection_boxes, vs.real_height, vs.real_width)
detection_masks_reframed = tf.cast(tf.greater(detection_masks_reframed, 0.5), tf.uint8)
# Follow the convention by adding back the batch dimension
tensor_dict['detection_masks'] = tf.expand_dims(detection_masks_reframed, 0)
if config.SPLIT_MODEL:
score_out = detection_graph.get_tensor_by_name('Postprocessor/convert_scores:0')
expand_out = detection_graph.get_tensor_by_name('Postprocessor/ExpandDims_1:0')
score_in = detection_graph.get_tensor_by_name('Postprocessor/convert_scores_1:0')
expand_in = detection_graph.get_tensor_by_name('Postprocessor/ExpandDims_1_1:0')
# Threading
score = model.score
expand = model.expand
gpu_worker = SessionWorker("GPU",detection_graph,tf_config)
cpu_worker = SessionWorker("CPU",detection_graph,tf_config)
gpu_opts = [score_out, expand_out]
cpu_opts = [tensor_dict['detection_boxes'], tensor_dict['detection_scores'], tensor_dict['detection_classes'], tensor_dict['num_detections']]
fps = FPS(config.FPS_INTERVAL).start()
masks = None
print('> Starting Detection')
while vs.isActive():
# Detection
if not (config.USE_TRACKER and track):
if config.SPLIT_MODEL:
# split model in seperate gpu and cpu session threads
if gpu_worker.is_sess_empty():
# read video frame, expand dimensions and convert to rgb
frame = vs.read()
# put new queue
gpu_feeds = {image_tensor: vs.expanded()}
if config.VISUALIZE:
gpu_extras = frame # for visualization frame
else:
gpu_extras = None
gpu_worker.put_sess_queue(gpu_opts,gpu_feeds,gpu_extras)
g = gpu_worker.get_result_queue()
if g is None:
# gpu thread has no output queue. ok skip, let's check cpu thread.
pass
else:
# gpu thread has output queue.
score,expand,frame = g["results"][0],g["results"][1],g["extras"]
if cpu_worker.is_sess_empty():
# When cpu thread has no next queue, put new queue.
# else, drop gpu queue.
cpu_feeds = {score_in: score, expand_in: expand}
cpu_extras = frame
cpu_worker.put_sess_queue(cpu_opts,cpu_feeds,cpu_extras)
c = cpu_worker.get_result_queue()
if c is None:
# cpu thread has no output queue. ok, nothing to do. continue
continue # If CPU RESULT has not been set yet, no fps update
else:
boxes, scores, classes, num, frame = c["results"][0],c["results"][1],c["results"][2],c["results"][3],c["extras"]
else:
# default session
frame = vs.read()
output_dict = sess.run(tensor_dict, feed_dict={image_tensor: vs.expanded()})
num = output_dict['num_detections'][0]
classes = output_dict['detection_classes'][0]
boxes = output_dict['detection_boxes'][0]
scores = output_dict['detection_scores'][0]
if 'detection_masks' in output_dict:
masks = output_dict['detection_masks'][0]
# reformat detection
num = int(num)
boxes = np.squeeze(boxes)
classes = np.squeeze(classes).astype(np.uint8)
scores = np.squeeze(scores)
# Visualization
vis = visualize_objectdetection(frame,boxes,classes,scores,masks,category_index,fps._glob_numFrames,
config.MAX_FRAMES,fps.fps_local(),config.PRINT_INTERVAL,config.PRINT_TH,
config.OD_DISPLAY_NAME,config.VISUALIZE,config.VIS_FPS,config.DISCO_MODE,config.ALPHA)
if not vis:
break
# Activate Tracker
if config.USE_TRACKER and num <= config.NUM_TRACKERS:
tracker_frame = frame
track = True
first_track = True
# Tracking
else:
frame = vs.read()
if first_track:
trackers = []
tracker_boxes = boxes
for box in boxes[~np.all(boxes == 0, axis=1)]:
tracker.init(conv_detect2track(box,vs.real_width, vs.real_height), tracker_frame)
trackers.append(tracker)
first_track = False
for idx,tracker in enumerate(trackers):
tracker_box = tracker.update(frame)
tracker_boxes[idx,:] = conv_track2detect(tracker_box, vs.real_width, vs.real_height)
vis = visualize_objectdetection(frame,tracker_boxes,classes,scores,masks,category_index,fps._glob_numFrames,
config.MAX_FRAMES,fps.fps_local(),config.PRINT_INTERVAL,config.PRINT_TH,
config.OD_DISPLAY_NAME,config.VISUALIZE,config.VIS_FPS,config.DISCO_MODE,config.ALPHA)
if not vis:
break
tracker_counter += 1
#tracker_frame = frame
if tracker_counter >= config.TRACKER_FRAMES:
track = False
tracker_counter = 0
fps.update()
# End everything
vs.stop()
fps.stop()
if config.SPLIT_MODEL:
gpu_worker.stop()
cpu_worker.stop()
