def get_actual(guess):
"""
Returns the object the human player was thinking of
"""
global _objects
yn = interface.ask("My guess is %s. Was I right? " % guess.name)
if yn == "yes":
obj_name = guess.name
obj_id = guess.id
else:
if robot():
obj_name = robot().ask_object()
else:
print "\nObject names:\n"
for j in range(len(_objects)):
print _objects[j].name
obj_name = raw_input("\nWhat was your object? Remember to type it exactly as you saw above. ")
while True:
check = False
for i in range(len(_objects)):
if _objects[i].name == obj_name:
check = True
obj_id = _objects[i].id
break
if check == True:
break
obj_name = raw_input("It seems as though you mistyped. Please try typing the name of your object again. ")
return obj_id, obj_name
def get_actual(guess):
"""
Returns the object the human player was thinking of
"""
global _objects
answer = interface.ask("My guess is %s. Was I right? " % guess.name)
if answer:
obj_name = guess.name
obj_id = guess.id
else:
if robot():
obj_name = robot().ask_object()
else:
print "\nObject names:\n"
for j in range(len(_objects)):
print _objects[j].name
obj_name = raw_input("\nWhat was your object? Remember to type it exactly as you saw above. ")
while True:
check = False
for i in range(len(_objects)):
if _objects[i].name == obj_name:
check = True
obj_id = _objects[i].id
break
if check == True:
break
obj_name = raw_input("It seems as though you mistyped. Please try typing the name of your object again. ")
return obj_id, obj_name
def playGame(self, number_of_objects):
"""
Play the game
"""
log.info('Playing game %d', self.id)
start = time.time()
# Generate all image probabilities at once since this takes a little while
if config.args.imagemodels:
Pi = models.gen_image_probabilities(self, number_of_objects)
else:
#we use -1 to say that we'll skip that tag, so this skips all tags for all objects
Pi = [[-1] * 289] * 17
# Initialize game stats to empty
NoOfQuestions = 0
round_wins = 0
round_losses = 0
avg_win = 0
avg_lose = 0
question_answers = {}
questions_asked = {}
objlist = objects.get_all()
# For each object in the set
count = 0
for i in objlist:
if config.args.notsimulated:
for j in range(len(objlist)):
print objlist[j].name
interface.say("Choose an object. Don't tell me! ")
time.sleep(2.5)
#TODO: ask if the person is ready???
result, number_of_questions, answers, askedQuestions = i.playObject(self, Pi, number_of_objects)
log.info("Game %d, object %d complete, updating stats", self.id, i.id)
#averages are calculated later (in main) using these sums
if result == 0: # Loss
round_losses += 1
avg_lose += number_of_questions
else: # Win
round_wins += 1
avg_win += number_of_questions
NoOfQuestions += number_of_questions
# Save questions and answers for later
questions_asked[i.id] = askedQuestions
question_answers[i.id] = answers
count += 1
if config.args.notsimulated == True:
quit = interface.ask("Would you like to quit this game early? \nThere are %d rounds left. " % (17 - count))
if quit == "yes":
break
# Save results
self._record(round_wins, round_losses, NoOfQuestions, count)
end = time.time()
log.info("Game %d complete (Took %ds)", self.id, int(end - start))
return round_wins, round_losses, NoOfQuestions, avg_win, avg_lose, question_answers, questions_asked
def ask(question_id, object_we_play, game, answers, pO, Pi, objects,
number_of_objects, answer_data):
"""
Asks a question and updates probabilities based on the answer
"""
# gets probability that an answer will be yes for each of the 7 possible t-values of question/answer combos
probabilityD = get_tval()
# gets the correct question for the given tag
# This is where we'll update the code to include newly-generated questions; right now the code just pulls pre-generated questions from a database.
question_tag = tags.get(question_id)
questions = tags.get_questions()
question = questions[question_id - 1]
# in not simulated game, get answer from user
if config.args.notsimulated:
answer = interface.ask(question + " ")
# in simulated game, get answer from answer bank in database
else:
answer = answer_data[game.id - 1][object_we_play.id - 1][question_id -
1]
print question, answer
answers.append(answer)
multipliers = []
for objectID in range(17): # for all known objects
# count the number of that object's descriptions that contain the tag used to formulate that question
T = get_t(objectID + 1, question_id, number_of_objects)
# number of yes answers to this question/object combo
yes_answers = objects[objectID][question_id - 1][0]
# total number of answers given for this question/object combo
total_answers = objects[objectID][question_id - 1][1]
# TODO: see if Dr. Nielsen has any suggestions for improving the use of the
# image models since they reduce the accuracy at the moment
if answer:
K = probabilityD[T] + (yes_answers + 1) / (total_answers + 2.0)
# if Pi is -1 then it means we're skipping the image models for that tag
if Pi[0][question_id - 1] == -1:
multiplier = K / 2
multipliers.append(multiplier)
else:
multiplier = (K + Pi[objectID][question_id - 1]) / 3
multipliers.append(multiplier)
else:
K = (1 - probabilityD[T]) + (total_answers - yes_answers +
1) / (total_answers + 2.0)
if Pi[0][question_id - 1] == -1:
multiplier = K / 2
else:
multiplier = (K + 1 - Pi[objectID][question_id - 1]) / 3
pO[objectID] *= multiplier
for objectID in range(
17, number_of_objects
): # TODO: for all new objects/unknown tags for a new object
# this assigns a probability to unknown objects based on the idea that
# an object is about as likely as the rest of the objects to have the quality
# of a certain tag - i.e. many objects are balls, but not many objects are polka dotted
# TODO: make more sophisticated - some tags are opposites, so if we know an opposite then we
# can take that into consideration when finding the multiplier
multiplier = np.mean(multipliers)
pO[objectID] *= multiplier
# Normalize the probabilities so that all object probabilities will sum to 1
pO /= np.sum(pO)
# Save the questions to each answer and the updated probabilities
with open("example.txt", "a") as myfile:
myfile.write(str(question_tag) + " -> " + str(answer) + " \n")
myfile.write(str(pO) + "\n")
return pO, answers
def playGame(self, number_of_objects):
"""
Play the game
"""
log.info('Playing game %d', self.id)
start = time.time()
# Generate all image probabilities at once since this takes a little while
if config.args.imagemodels:
Pi = models.gen_image_probabilities(self, number_of_objects)
else:
#we use -1 to say that we'll skip that tag, so this skips all tags for all objects
Pi = [[-1] * 289] * 17
# Initialize game stats to empty
NoOfQuestions = 0
round_wins = 0
round_losses = 0
avg_win = 0
avg_lose = 0
question_answers = {}
questions_asked = {}
objlist = objects.get_all()
count = 0
# quit variable passed to main that tells if user wants to quit
quit = False
# For each object in the set
for i in objlist:
if robot():
robot().initGazeCounts()
if config.args.notsimulated:
for j in range(len(objlist)):
print objlist[j].name
interface.say("Choose an object. Don't tell me! ")
time.sleep(2.5)
#TODO: ask if the person is ready???
result, number_of_questions, answers, askedQuestions = i.playObject(self, Pi, number_of_objects)
log.info("Game %d, object %d complete, updating stats", self.id, i.id)
#averages are calculated later (in main) using these sums
if result == 0: # Loss
round_losses += 1
avg_lose += number_of_questions
else: # Win
round_wins += 1
avg_win += number_of_questions
NoOfQuestions += number_of_questions
# Save questions and answers for later
questions_asked[i.id] = askedQuestions
question_answers[i.id] = answers
count += 1
if config.args.notsimulated == True:
quit = not interface.ask("Do you want to play again?")
# quit = interface.ask("Would you like to quit this game early? \nThere are %d rounds left. " % (17 - count))
if quit:
break
# Save results
self._record(round_wins, round_losses, NoOfQuestions, count)
end = time.time()
log.info("Game %d complete (Took %ds)", self.id, int(end - start))
return round_wins, round_losses, NoOfQuestions, avg_win, avg_lose, question_answers, questions_asked, quit
def ask(question_id, object_we_play, game, answers, pO, Pi, objects, number_of_objects, answer_data):
"""
Asks a question and updates probabilities based on the answer
"""
# gets probability that an answer will be yes for each of the 7 possible t-values of question/answer combos
probabilityD = get_tval()
# gets the correct question for the given tag
question_tag = tags.get(question_id)
questions = tags.get_questions()
question = questions[question_id - 1]
# in not simulated game, get answer from user
if config.args.notsimulated:
answer = interface.ask(question + " ")
# in simulated game, get answer from answer bank in database
else:
answer = answer_data[game.id-1][object_we_play.id-1][question_id-1]
print question, answer
answers.append(answer)
multipliers = []
for objectID in range(17): # for all known objects
# count the number of that object's descriptions that contain the tag used to formulate that question
T = get_t(objectID+1, question_id, number_of_objects)
# number of yes answers to this question/object combo
yes_answers = objects[objectID][question_id-1][0]
# total number of answers given for this question/object combo
total_answers = objects[objectID][question_id-1][1]
# TODO: see if Dr. Nielsen has any suggestions for improving the use of the
# image models since they reduce the accuracy at the moment
if answer:
K = probabilityD[T] + (yes_answers + 1)/(total_answers + 2.0)
# if Pi is -1 then it means we're skipping the image models for that tag
if Pi[0][question_id-1] == -1:
multiplier = K / 2
multipliers.append(multiplier)
else:
multiplier = (K + Pi[objectID][question_id-1]) / 3
multipliers.append(multiplier)
else:
K = (1 - probabilityD[T]) + (total_answers - yes_answers + 1)/(total_answers + 2.0)
if Pi[0][question_id-1] == -1:
multiplier = K / 2
else:
multiplier = (K + 1 - Pi[objectID][question_id-1]) / 3
pO[objectID] *= multiplier
for objectID in range(17,number_of_objects): # TODO: for all new objects/unknown tags for a new object
# this assigns a probability to unknown objects based on the idea that
# an object is about as likely as the rest of the objects to have the quality
# of a certain tag - i.e. many objects are balls, but not many objects are polka dotted
# TODO: make more sophisticated - some tags are opposites, so if we know an opposite then we
# can take that into consideration when finding the multiplier
multiplier = np.mean(multipliers)
pO[objectID] *= multiplier
# Normalize the probabilities so that all object probabilities will sum to 1
pO /= np.sum(pO)
# Save the questions to each answer and the updated probabilities
with open("example.txt", "a") as myfile:
myfile.write(str(question_tag) + " -> " + str(answer) + " \n")
myfile.write(str(pO) + "\n")
return pO, answers
