def Cal_Angle(img, detection_net=_DETECTION, angle_net=_ANGLE):
if Decaf:
scores = detection_net.classify(img, False)
is_car = detection_net.top_k_prediction(scores, 1)
#print is_car
if is_car[1][0] == 'car':
car_conv3 = detection_net.feature("conv3_neuron_cudanet_out")
mid_convs = car_conv3.reshape((car_conv3.shape[0],-1))
scores = angle_net.classify(mid_convs)
angles = angle_net.top_k_prediction(scores, 3)
print angles
o = float(angles[1][0])-90 # *180/np.pi
o = o if o>0 else o+360
return int(o)
else:
scores = car_model.show_predictions(img[:,:,0])
print scores
if scores[-1][0] == 'car':
mid_convs = car_model.get_features(img[:,:,0])
scores = angle_net.classify(mid_convs)
angles = angle_net.top_k_prediction(scores, 3)
print angles
o = float(angles[1][0])-90 # *180/np.pi
o = o if o>0 else o+360
return int(o)
return None
def Cal_Angle(img, detection_net=_DETECTION, angle_net=_ANGLE):
if Decaf:
scores = detection_net.classify(img, False)
is_car = detection_net.top_k_prediction(scores, 1)
# print is_car
if is_car[1][0] == "car":
car_conv3 = detection_net.feature("conv3_neuron_cudanet_out")
mid_convs = car_conv3.reshape((car_conv3.shape[0], -1))
scores = angle_net.classify(mid_convs)
angles = angle_net.top_k_prediction(scores, 3)
print angles
o = float(angles[1][0]) - 90 # *180/np.pi
o = o if o > 0 else o + 360
return int(o)
else:
scores = car_model.show_predictions(img[:, :, 0])
print scores
if scores[-1][0] == "car":
mid_convs = car_model.get_features(img[:, :, 0])
scores = angle_net.classify(mid_convs)
angles = angle_net.top_k_prediction(scores, 3)
print angles
o = float(angles[1][0]) - 90 # *180/np.pi
o = o if o > 0 else o + 360
return int(o)
return None
def Cal_Angle(img, detection_net=_DETECTION, angle_net=_ANGLE):
# TODO: prob filter 0.95会不会太严格
if Decaf:
scores = detection_net.classify(img, False)
is_car = detection_net.top_k_prediction(scores, 1)
#print is_car
if is_car[1][0] == 'car' and is_car[0][0] >= 0.9:
car_conv3 = detection_net.feature("conv3_neuron_cudanet_out")
mid_convs = car_conv3.reshape((car_conv3.shape[0],-1))
scores = angle_net.classify(mid_convs)
angles = angle_net.top_k_prediction(scores, 3)
o = float(angles[1][0])-90 # *180/np.pi
o = o if o>0 else o+360
return int(o), is_car[0][0]
else:
is_car = car_model.show_predictions(img[:,:,0])
# print is_car
if is_car[-1][0] == 'car' and is_car[-1][1] >= 0.9:
mid_convs = car_model.get_features(img[:,:,0])
scores = angle_net.classify(mid_convs)
angles = angle_net.top_k_prediction(scores, 3)
#print angles
o = float(angles[1][0])-90 # *180/np.pi
o = o if o>0 else o+360
return int(o), is_car[-1][1]
return None
else:
print "Using " + sys.argv[1]
car = imread(sys.argv[1])
Decaf = False
if Decaf:
from kitnet import DecafNet as KitNet
kit_net = KitNet()
# print car.shape
car = car.reshape((40,40,1))
scores = kit_net.classify(car)
print 'Is car ? prediction:', kit_net.top_k_prediction(scores, 1)
car_conv3 = kit_net.feature("conv3_neuron_cudanet_out") #conv3_cudanet_out
mid_convs = car_conv3.reshape((car_conv3.shape[0],-1))
else:
os.chdir("E:/2013/cuda-convnet/trunk")
# sys.path.append("E:/2013/cuda-convnet/trunk")
from show_pred import model as car_model
scores = car_model.show_predictions(car)
print 'Is car ? prediction:', scores[-1]
mid_convs = car_model.get_features(car)
os.chdir(os.path.dirname(os.path.realpath(__file__)))
# visualize.draw_net_to_file(net._net, "decafnet.png") #bug!
# print 'Network structure written to decafnet.png'
net = DecafNet()
scores = net.classify(mid_convs)
print 'Direction ? prediction:', net.top_k_prediction(scores, 5)#scores*180