# coding: utf-8
from BackProp_Python_v2 import NN
from vrep_pioneer_simulation import VrepPioneerSimulation
from rdn import Pioneer # rdn pour ROS avec le pioneer
# import rospy
from online_trainer import OnlineTrainer
from offline_trainer import OfflineTrainer
import json
import threading
# On peut changer de robot mais il faut aussi changer la fonction de co�t du robot qui est adapt�e au mod�le
robot = VrepPioneerSimulation()
# robot = Pioneer(rospy)
HL_size = 9
# nbre neurons of Hiden layer
network = NN(3, HL_size, 2)
choice = input('Do you want to load previous network? (y/n) --> ')
if choice == 'y':
with open('last_w.json') as fp:
json_obj = json.load(fp)
for i in range(3):
for j in range(HL_size):
network.wi[i][j] = json_obj["input_weights"][i][j]
for i in range(HL_size):
for j in range(2):
network.wo[i][j] = json_obj["output_weights"][i][j]
def simulation(options):
"""Performs a learning session with the given option dictionnary"""
# Extract data from the option dictionnary
RealRobot = options["RealRobot"]
Verbose = options["Verbose"]
LogFileName = options["LogFileName"]
WeightFileName = options["WeightFileName"]
NeuronFileName = options["NeuronFileName"]
Description = options["Description"]
SingleSimulation = options["SingleSimulation"] == '1'
MaximumDuration = float(options["MaximumDuration"])
Tick = float(options["Tick"])
Prediction = options["Prediction"] == '1'
ExperimentalTheta = options["ExperimentalTheta"] == '1'
Gain = float(options["Gain"])
RestrictThetaShift = options["RestrictThetaShift"] == '1'
RestrictPropagation = options["RestrictPropagation"] == '1'
InvertRestriction = options["InvertRestriction"] == '1'
NewSigmoid = options["NewSigmoid"] == '1'
RandomRatio = float(options["RandomRatio"])
Load = options["Load"]
Learn = options["Learn"]
HiddenNeuronNumber = int(options["HiddenNeuronNumber"])
LearningStep = float(options["LearningStep"])
ThetaShiftRatio = float(options["ThetaShiftRatio"])
Size = float(options["Size"])
FixedStartingPosition = options["FixedStartingPosition"] == '1'
StartingPositionX = float(options["StartingPositionX"])
StartingPositionY = float(options["StartingPositionY"])
StartingPositionTheta = float(options["StartingPositionTheta"])
FixedTargetPosition = options["FixedTargetPosition"] == '1'
TargetPositionX = float(options["TargetPositionX"])
TargetPositionY = float(options["TargetPositionY"])
TargetPositionTheta = float(options["TargetPositionTheta"])
StopAutomatically = options["StopAutomatically"]
CriterionPercent = float(options["CriterionPercent"])
CriterionDuration = float(options["CriterionDuration"])
# Create the right robot model (pioneer API or vrep API)
if RealRobot == 'a':
RealRobot = ask('Is it a real robot ? (y/n) --> ')
if RealRobot == 'y':
robot = Pioneer(rospy)
else:
robot = VrepPioneerSimulation()
# Create the Neuron Network model
# and load its values from a save file if necessary
if WeightFileName:
if Load == 'a':
choice = ask('Do you want to load previous network ? (y/n) --> ')
else:
choice = Load
if choice == 'y':
try:
with open(WeightFileName) as fp:
json_obj = json.load(fp)
except:
print("Warning : could not find or open \"" + WeightFileName +
"\", starting from a random network...")
network = NN(3 + len(robot.activated_sensors),
HiddenNeuronNumber, 2)
else:
# Check the size of the saved network
n = len(json_obj["output_weights"])
if n != HiddenNeuronNumber:
print(
"Warning : the config.cfg file wanted ",
HiddenNeuronNumber,
" hidden neurons, but the saved network has ",
n,
" hidden neurons. The saved network will still be used...",
sep='')
HiddenNeuronNumber = n
network = NN(3 + len(robot.activated_sensors),
HiddenNeuronNumber, 2)
# Load the weight values
for i in range(3):
for j in range(HiddenNeuronNumber):
network.wi[i][j] = json_obj["input_weights"][i][j]
for i in range(HiddenNeuronNumber):
for j in range(2):
network.wo[i][j] = json_obj["output_weights"][i][j]
else:
network = NN(3 + len(robot.activated_sensors), HiddenNeuronNumber,
2)
else:
network = NN(3 + len(robot.activated_sensors), HiddenNeuronNumber, 2)
if NewSigmoid:
network.newSigmoid = True
# The trainer will monitor the lesson until one of the stopping criterion is met
# It works on a separate thread, started by its start() method
# See online_trainer.py for more details
trainer = OnlineTrainer(robot, network)
if Learn == 'a':
choice = ask('Do you want to learn ? (y/n) --> ')
else:
choice = Learn
if choice == 'y':
trainer.training = True
else: # choice = 'n'
trainer.training = False
# Move the robot to its starting location
# This is only possible with a vrep simulation
# since the real robot always starts at [0, 0, 0]
if not (RealRobot == 'y'):
if FixedStartingPosition:
location = [
StartingPositionX, StartingPositionY, StartingPositionTheta
]
robot.set_position2(location)
else:
location = []
while len(location) != 3:
location = input(
"Enter the starting location : x y radian --> ")
location = location.split()
for i in range(len(location)):
location[i] = float(location[i])
robot.set_position2(location)
if FixedTargetPosition:
target = [TargetPositionX, TargetPositionY, TargetPositionTheta]
else:
target = []
while len(target) != 3:
target = input("Enter the first target : x y radian --> ")
target = target.split()
for i in range(len(target)):
target[i] = float(target[i])
print('New target : [%d, %d, %d]' % (target[0], target[1], target[2]))
if StopAutomatically == 'a':
StopAutomatically = ask(
"Do you want to stop the simulation when the ending criterion is met ? --> "
)
StopAutomatically = StopAutomatically == 'y'
if Verbose == 'a':
Verbose = ask(
"Do you want to display data while the trainer is running ? (y/n) --> "
)
Verbose = Verbose == 'y'
continue_running = True
# Each iteration correspond to one lesson
# If the SingleSimulation parameter is not set to 1 in config.cfg,
# Several successive lessons can be done
while (continue_running):
if StopAutomatically:
stop_criterion = [CriterionPercent, CriterionDuration]
else:
stop_criterion = None
thread = threading.Thread(
target=trainer.train,
args=(target, options, Tick, Prediction, ExperimentalTheta, Gain,
RestrictThetaShift, RestrictPropagation, InvertRestriction,
RandomRatio, LearningStep, ThetaShiftRatio, Size,
MaximumDuration, stop_criterion, Verbose, LogFileName,
NeuronFileName, Description))
thread.daemon = True # Kill the current simulation when this script is killed
trainer.running = True
begin_date = time.time()
thread.start() # Start the simulation on another thread
if StopAutomatically or MaximumDuration > 0:
while trainer.running:
time.sleep(0.1)
else: # StopAutomatically=='n' and MaximumDuration==0:
input("Press Enter to stop the current training\n")
elapsed_time = time.time() - begin_date
print(" " * 60, end='\r')
print("Finished in " + str(round(elapsed_time, 1)) + " sec.")
trainer.running = False
if SingleSimulation:
choice = 'n'
else:
choice = ask("Do you want to continue ? (y/n) --> ")
if choice == 'y':
if Learn == 'a':
choice_learning = ask('Do you want to learn ? (y/n) --> ')
else:
choice_learning = Learn
if choice_learning == 'y':
trainer.training = True
elif choice_learning == 'n':
trainer.training = False
if FixedStartingPosition:
location = [
StartingPositionX, StartingPositionY, StartingPositionTheta
]
robot.set_position2(location)
else:
choice_reset_position = ask(
"Do you want to change the robot location ? (y/n) --> ")
if choice_reset_position == 'y':
location = []
while len(location) != 3:
location = input(
"Enter the new location : x y radian --> ")
location = location.split()
for i in range(len(location)):
location[i] = float(location[i])
robot.set_position2(location)
if FixedTargetPosition:
target = [
TargetPositionX, TargetPositionY, TargetPositionTheta
]
else:
target = []
while len(target) != 3:
target = input("Enter the new target : x y radian --> ")
target = target.split()
for i in range(len(target)):
target[i] = float(target[i])
print('New target : [%d, %d, %d]' %
(target[0], target[1], target[2]))
else: # choice = 'n'
continue_running = False
if WeightFileName:
# Save the current Neuron Network
with open(WeightFileName, 'w') as fp:
json_obj = {
"input_weights": network.wi,
"output_weights": network.wo
}
json.dump(json_obj, fp)
fp.close()
while not trainer.ready_to_exit:
time.sleep(0.1)
return elapsed_time