def initialize_RPi_Stuff():
# note: global variables start with a little "g"
global g_getter
global g_camera
global g_steerstats
global g_graph
global g_image_data
global g_stop_event
global g_lock
g_getter = DataGetter()
g_camera = picamera.PiCamera()
g_camera.resolution = (128, 96) #final image size
g_camera.framerate = 10 #<---- framerate (fps) determines speed of data recording
g_steerstats = np.load(g_pi_training_steerstats_file)['arr_0']
model.load_weights(g_pi_training_weights_file)
model._make_predict_function()
g_graph = tf.get_default_graph()
g_image_data = np.zeros((36, 128, 3), dtype=np.uint8)
g_stop_event = threading.Event()
g_lock = threading.Lock()
# dazzle them with Night Rider LED show...
for x in range(0, 3):
for i in range(0, 6):
displayBinaryOnLEDs(2**i)
time.sleep(.125)
for i in range(5, -1, -1):
displayBinaryOnLEDs(2**i)
time.sleep(.125) # dazzle them with Night Rider LED show...
# turn off all LEDs for initialization
turn_OFF_all_LEDs()
def initialize_service():
#initialize the serial port: if the first port fails, we try the other one
global g_serial
g_serial = 0
try:
g_serial = serial.Serial('/dev/ttyACM1')
except serial.SerialException:
try:
g_serial = serial.Serial('/dev/ttyACM0')
except serial.SerialException:
print('Cannot connect to serial port')
#initialize the camera
global g_camera
g_camera = picamera.PiCamera()
g_camera.resolution = (128, 96)
g_camera.framerate = 10
#initialize the data collector object
global g_collector
g_collector = DataCollector(g_serial, "/home/pi/autonomous/data")
#initialize the image frame to be shared in autonomous mode
global g_image_data
g_image_data = np.zeros((36, 128, 3), dtype=np.uint8)
#initialize some stuff needed for network thread
global g_stop_event
g_stop_event = threading.Event()
global g_lock
g_lock = threading.Lock()
#this is the object the camera writes to in autonomous mode
global g_getter
g_getter = DataGetter()
#this stuff sets up the network
global g_graph
g_graph = tf.get_default_graph()
#model.load_weights('weights_2018-02-24_14-00-35_epoch_40.h5')
model.load_weights('weights_2018-04-05_02-43-30_epoch_66.h5')
model._make_predict_function()
global g_steerstats
g_steerstats = np.load('steerstats.npz')['arr_0']
global g_ip_thread
g_ip_thread = 0
def initialize_service():
#initialize the serial port: if the first port fails, we try the other one
global g_serial
try:
g_serial = serial.Serial('/dev/ttyACM1')
except serial.SerialException:
try:
g_serial = serial.Serial('/dev/ttyACM0')
except serial.SerialException:
logging.debug("error: cannot connect to serial port")
#initialize the camera
global g_camera
g_camera = picamera.PiCamera()
g_camera.resolution = (128, 96)
g_camera.framerate = FRAME_RATE
#initialize the data collector object
global g_collector
g_collector = DataCollector(g_serial, COLLECT_DIR)
#initialize the image frame to be shared in autonomous mode
global g_image_data
g_image_data = np.zeros((36, 128, 3), dtype=np.uint8)
#initialize some stuff needed for network thread
global g_stop_event
g_stop_event = threading.Event()
global g_lock
g_lock = threading.Lock()
#this is the object the camera writes to in autonomous mode
global g_getter
g_getter = DataGetter()
#this stuff sets up the network
global g_graph
g_graph = tf.get_default_graph()
#model.load_weights('weights_2018-02-24_14-00-35_epoch_40.h5')
model.load_weights(WEIGHTS_FILE)
model._make_predict_function()
global g_steerstats
g_steerstats = np.load(STEERSTATS_FILE)['arr_0']
global g_ip_thread
g_ip_thread = 0
for image in imdata[0:data_lengths[i]-args.delay]: #note that we compensate for delay here
crop_image=image[row_offset:row_offset+nrows, :]
#we have to normalize each image to zero mean, unit variance:
image_mean=crop_image.mean()
image_std=crop_image.std()
training_images[n]=(crop_image-image_mean)/image_std
n+=1#this increments for each image, and keeps track of where to put the images in the training_images array
i+=1#this increments for each file, and keeps track of where to get the number of data frames from data_lengths
#this loads the predefined network architecture from dropout_model.py
from dropout_model import model
num_epochs=args.epochs#number of epochs to train over
save_epochs=args.save_frequency#number of epochs between weight file saves
#if the user inputs a weight file for initial state, load it:
if args.init_weights!="":
model.load_weights(args.init_weights)
for n in range(num_epochs):
print("starting epoch {0}".format(n))
h=model.fit([training_images], [(steer-steerSampleMean)/steerSampleSTD],
batch_size=25, epochs=1, verbose=1, validation_split=0.1, shuffle=True)
if n%save_epochs ==0 :
savename='%s_%s_epoch_%d.h5'%(args.weight_filename, time_string, n)
print("Saving epoch {0} to {1}".format(n, savename))
model.save_weights(savename, overwrite=True)
savename='%s_%s_complete.h5'%(args.weight_filename, time_string)
model.save_weights(savename, overwrite=True)
