def worker_row(wkid, task_queue, res_dict):
while True:
st_time = time.time()
i = task_queue.get()
if i == -1: break # Break when stop token received
print "Worker #%d is working on transition of row #%d." % (wkid, i),
res = pv.KWTable()
row_vec = m_mat[i]
u = u_vec[i]
for j in range(len(row_vec)):
if row_vec[j] == 0.0: continue # No need to do 0-coeff aggregation
res.aggr_inplace(x_vec[j], 1.0, row_vec[j])
if not u == 0.0: res.aggr_inplace(y, 1.0, u)
ori_size = len(res)
res.rsvr_sample(RSVR_SIZE, in_place=True)
print " Size = %d -> %d" % (ori_size, len(res)),
el_time = time.time() - st_time
print " Elapsed time = %f" % (el_time)
res_dict[i] = res
return
def forecast(r):
# r is the # of steps to look ahead.
st_time = time.time()
ret = pv.KWTable()
l = x_vec[0]
b = x_vec[1]
s = x_vec[r + 1] # r-th seasonal component
ret.aggr_inplace(l, 1.0, 1.0)
ret.aggr_inplace(b, 1.0, float(r))
ret.aggr_inplace(s, 1.0, 1.0)
ret.rsvr_sample(RSVR_SIZE, in_place=True)
el_time = time.time() - st_time
print "Making %d-step forecast. Time stamp = %d. Elapsed time = %f" % (
R, TS_CURR + INTERVAL * (R - 1), el_time)
return ret
def read_rec(ts):
return pv.KWTable(filetype=FILETYPE,
fn=os.path.join(DATA_DIR,
str(ts) + ".rec"))
INTERVAL = 3600 # Seconds in a time slot
TS_START = 1181088000 # Starting timestamp (in seconds)
TS_END = TS_START + INTERVAL * 60 # Ending timestamp
FILETYPE = "flowbin"
PERIOD = 24 # # of time slots in a period
R = 1 # Forecast # of steps
ALPHA = 0.2
BETA = 0.2
GAMMA = 0.2
N_WORKERS = 4
# ---------- Global variables and objects ----------
TS_CURR = 0 # Current timestamp
x_vec = mp.Manager().dict(
) # State vector. Make it globally accessible (shared dictionary)
y = pv.KWTable() # Current time slot's observation
m_mat = [] # Transition matrix
u_vec = [] # U-Vector for transition
def make_trans_matrix():
ret = []
for i in range(PERIOD + 1):
ret.append([0.0] * (PERIOD + 1))
ret[0][0] = 1.0 - ALPHA
ret[0][1] = 1.0 - ALPHA
ret[0][2] = -1.0 * ALPHA
ret[1][0] = -1.0 * BETA
ret[1][1] = 1.0 - BETA
ret[1][2] = -1.0 * BETA
return pv.KWTable(filetype=FILETYPE,
fn=os.path.join(DATA_DIR,
str(ts) + ".rec"))
def worker_samp(i):
# Grab a KWTable from queue and sample it down to size
print "Worker is working on x_vec[%d]: KWTable(%x). Current size = %d" % (
i, id(x_vec[i]), len(x_vec[i]))
ret = x_vec[i].rsvr_sample(RSVR_SIZE, in_place=False)
# To do multiprocessing, I have no choice but turn off in_place option. But I still get large speedup.
return ret
if __name__ == "__main__":
l0 = pv.KWTable()
b0 = pv.KWTable()
s0_list = [pv.KWTable()
for i in range(PERIOD)] # create (w-1) empty tables
# note that s0_list[0] is not used
# indices 1..(PERIOD-1) are used
print "Using Python interpreter:", sys.executable
# ---------- First training period ----------
print "First training period"
for i in range(PERIOD):
st_time = time.time()
y = read_rec(TS_CURR)
b0.aggr_inplace(y, 1.0, -1.0)
print "Read time slot", TS_CURR, " b0 -= y",