def time(self, session_index, avatar_index,\
tree_nb, reduc_traj = 1, threads = 1 ):
# Method called many times in parallel to compute the time
# at which avatar made decision. reduc_traj is a boolean
# to decide if we introduce some dichotomic search to find
# the time of choice. Threads is the number of threads
# used.
# tree_nb is the number of the look up tree allocated to
# the current subprocess.
trajectory = hp.Trajectory_points(\
session_index,\
avatar_index)
# I am only interested in the trajectory after
# earthquake
time_earthquake = hp.earthquake_time(session_index)
time_token = 0
for point in trajectory:
if point[3] < time_earthquake:
time_token += 1
else:
break
trajectory = trajectory[time_token:]
# An avatar can make more than one choice. Going for
# the train, being refused a seat, going then to a
# car.
choices = []
# Memory to store the decision_times corresponding to
# the choices
decision_time_mem = []
i = 0
if reduc_traj == 1:
# Reduces greatly the span in which to search for the
# time of decision. Makes the search much faster.
ind = self.good_part(trajectory, tree_nb)
low = ind[0]
high = ind[1]
trajectory = trajectory[low : high]
# memories used to apply a filter on the series of
# ratios within the neighbours. removing noise without
# smoothing the transition corresponding to the
# decision time.
car_ratio_mem1 = -1
time_mem1 = -1
car_ratio_mem2 = -1
time_mem1 = -1
car_ratio = -1
time_mem = -1
#find interesting subpart
if threads != 1:
# I use the word thread because it is short but the
# correct term is subprocess. multi-threading in
# python still uses only one interpreter which is
# limited to one process on the computer and doesn't
# much improvement on computation time.
# multi-processing is much more powerfull as it
# duplicates the python interpreter itselfs into many
# processes.
# Shared memory across processes.
ratios = Queue()
arguments = []
for position in trajectory:
# prepares the arguments to pass to subprocesses
# for computation
arguments.append([position[0] , position[1],\
position[3], ratios])
# array used as memory to store subprocesses
p = []
for i in range(threads):
# Giving the good length to the array
p.append(-1)
while not len(arguments) == 0:
# While there is some arguments to be passed for
# computation...
cnt = 0
for process in p:
# iterate over the processes
if process == -1 or not process.is_alive():
# If one is idled or not yet instantiated
if p[cnt] != -1:
# finish it properly if idled
p[cnt].join()
if len(arguments) == 0:
# if no more arguments to pass, then
# go next
continue
# else pass one argument to the
# subprocess for computation
argument = arguments.pop()
argument.append(cnt)
p[cnt] = Process(\
target=self.car_ratio_threaded,\
args = (argument[0],\
argument[1],\
argument[2],\
argument[3],\
argument[4]))
p[cnt].start()
cnt += 1
# When no more arguments, it can take a while to
# wait for all the currently running subprocesses
# to finish, let's wait a little
for process in p:
process.join()
results = []
while not ratios.empty():
# The computed ratios where stored in shared
# memory, lets unpack this memory
results.append(ratios.get())
# The computing being done in parallel, the ratios
# are not in the good order, lets order it.
results = sorted(results,key=itemgetter(0))
car_ratios = []
for result in results:
car_ratios.append(result[1])
else:
# if the procedure is not done in parallel, it is much
# simpler ...
car_ratios = self.car_ratios(trajectory, tree_nb)
# index of position in trajectory during the loop
j = 0
# used for applying a filter
switch_token = 0
for position in trajectory:
# Applying filter to smooth noise without
# smoothing the transition corresponding to the
# choice
if position[3] < time_earthquake:
continue
if i == 0:
i = 1
car_ratio_mem2 = car_ratio_mem1
time_mem2 = time_mem1
car_ratio_mem1 = car_ratio
time_mem1 = time_mem
time_mem = position[3]
car_ratio = car_ratios[j]
if car_ratio == -1:
continue
if car_ratio_mem2 == -1:
continue
le = abs(car_ratio - car_ratio_mem2)
li = min(abs(car_ratio - car_ratio_mem1), \
abs(car_ratio_mem1 - car_ratio_mem2))
if li > le:
car_ratio_mem1 = \
(car_ratio_mem2 + car_ratio)/2.0
if car_ratio_mem2 > 0.90 and switch_token != 1:
# seems to have choosen car
choices.append([time_mem2, 1])
switch_token = 1
decision_time_mem.append(time_mem2)
elif car_ratio_mem2 < 0.10 and switch_token != 2:
# seems to have choosen train
choices.append([time_mem2, 2])
switch_token = 2
decision_time_mem.append(time_mem2)
elif car_ratio_mem2 < 0.80 and car_ratio_mem2 >0.2:
# no obvious choice
switch_token = 0
#we are loosing the 2 last value of car_ratio,
#never mind, a choice in the last two seconds is
#not relevant
j += 1
return choices
def __init__(self, nb_trees = 4):
#builds the KD-tree extracting people from choice
#'train' on one hand, on choice 'car' (A or B) on the
#other hand, concatening the points on each side, and
#then concatening the two sides remembering the
#junction indice.
self.nb_trees = nb_trees
car_array = []
train_array = []
choice_hashes = {}
con = lite.connect('everscape.db')
with con:
# Loading the choices of avatars
cur = con.cursor()
cur.execute("\
SELECT * FROM Path\
")
rows = cur.fetchall()
for row in rows:
# Stores and sorts those choices in hashes
try:
# Ensures the hash corresponding to the sessions
# is created, else creates it.
choice_hashes[row[0]]
except:
choice_hashes[row[0]] = {}
# For each session hash, the key is the avatar,
# the value is its choice.
# Here all the sessions and avatar numbers are
# real numbers, not the indices
choice_hashes[row[0]][row[1]]=row[3]
for session in choice_hashes:
# Iterating over the sessions, here the session
# variable is a hash (iterating over an array of
# hashes
# Getting the index of the session in order to be
# able to get the hash containing the choices made
# by the avatars during the session
session_index = sessions.index(session)
# Train logs. Not only people who choose train
# went to the train station. Some other people
# choose to go to the station but were not granted
# a seat. In the scope of this analysis we
# consider they made the choice to go to the
# station.
logs = hp.train_logs(session_index)
# Getting the hash containing the choices of the
# avatars in the session.
avatars_choice = choice_hashes[session]
# Usefull to retrieve the index of the avatars
avatar_array = hp.Avatar_of_session(session_index)
for avatar in avatars_choice:
# Iterating in this hash
# Usefull to call the method from helpers file
avatar_index = avatar_array.index(avatar)
# Getting the trajectory of the considered
# avatar.
trajectory = hp.Trajectory_points(session_index,\
avatar_index)
# We only take the part of the trajectory
# after the earthquake. If the avatar went by
# car and returned by train, mingling the two
# parts of the trajectory is problematic.
time_earthquake =\
hp.earthquake_time(session_index)
if (avatars_choice[avatar] == 1 or \
avatars_choice[avatar] == 2) and\
not hp.requested_train(logs, avatar):
# Carefull, some people who took a car
# went first to the train station. We
# don't want to take those people into
# account to build a tree against which we
# will look up trajectories.
for point in trajectory:
if point[3] > time_earthquake:
# Only the part of trajectory after
# earthquake is usefull.
car_array.append([point[0],\
point[1]])
# all_array.append([point[0],\
# point[1]])
elif avatars_choice[avatar] == 3:
# print("classified train")
for point in trajectory:
if point[3] > time_earthquake:
# Only the part of trajectory after
# earthquake is usefull.
train_array.append (\
[point[0], point[1]])
# When I will ask the tree for neighbours, the tree
# will give back indices of the data points
# corresponding to neighbours. As I will concatene
# the data points from car user with data points from
# train users, I have to keep track of the limit
# indice between both dataset in order to trace which
# indices correspond to which dataset
self.limit = len(car_array)
# Here we concatene the to datasets
data = car_array
for position in train_array:
data.append(position)
# Building a KDTree is recursive and here we have some
# volume
sys.setrecursionlimit(10000)
# Building many times the same tree (it is fast) so
# that later, for slow parts of the program we can
# make parallel computing (to compute the time of
# decision).
self.trees = []
for i in range(self.nb_trees):
self.trees.append(-1)
tree = sp.KDTree(data)
for i in range(self.nb_trees):
self.trees[i] = tree