def examine_example(self, i, N, labels, kernel_results):
''' Check if the alpha_i can be optimized. It should satisfy the KKT condition.
If so, choose it as one parameter to be optimized.
Args:
i(int): index of the alpha to be checked
N(int): the number of features
labels(Expr): the labels of the training data
kernel_results(Expr): the result of the kernel function on the training data
'''
Ei = self.E[i, 0]
ai = self.alpha[i, 0]
r = labels[i, 0] * Ei
# check if satisfy KKT condition
if r < -self.tol and ai < self.C or r > self.tol and ai > self.tol:
alpha_expr = expr.lazify(self.alpha)
active_set_mask = (alpha_expr > self.tol) * (alpha_expr < self.C)
active_set = active_set_mask.glom().nonzero()[0]
#print 'actives:', active_set
# first check the jth example that maximize the |Ei - Ej|
idxj = -1
if active_set.shape[0] > 1:
active_E = expr.abs(expr.lazify(self.E) - Ei)[active_set_mask -
True]
idxj = np.argmax(active_E.glom())
if self.take_step(idxj, i, N, labels, kernel_results): return 1
# then check the examples in active_set
for j in active_set:
if j != idxj and self.take_step(j, i, N, labels,
kernel_results):
return 1
# finally check the other examples
for j in xrange(N):
if j not in active_set and self.take_step(
j, i, N, labels, kernel_results):
return 1
return 0
def examine_example(self, i, N, labels, kernel_results):
""" Check if the alpha_i can be optimized. It should satisfy the KKT condition.
If so, choose it as one parameter to be optimized.
Args:
i(int): index of the alpha to be checked
N(int): the number of features
labels(Expr): the labels of the training data
kernel_results(Expr): the result of the kernel function on the training data
"""
Ei = self.E[i, 0]
ai = self.alpha[i, 0]
r = labels[i, 0] * Ei
# check if satisfy KKT condition
if r < -self.tol and ai < self.C or r > self.tol and ai > self.tol:
alpha_expr = expr.lazify(self.alpha)
active_set_mask = (alpha_expr > self.tol) * (alpha_expr < self.C)
active_set = active_set_mask.glom().nonzero()[0]
# print 'actives:', active_set
# first check the jth example that maximize the |Ei - Ej|
idxj = -1
if active_set.shape[0] > 1:
active_E = expr.abs(expr.lazify(self.E) - Ei)[active_set_mask - True]
idxj = np.argmax(active_E.glom())
if self.take_step(idxj, i, N, labels, kernel_results):
return 1
# then check the examples in active_set
for j in active_set:
if j != idxj and self.take_step(j, i, N, labels, kernel_results):
return 1
# finally check the other examples
for j in xrange(N):
if j not in active_set and self.take_step(j, i, N, labels, kernel_results):
return 1
return 0
def predict_price(ask, bid, t):
# element-wise difference
spread = ask - bid
# element-wise average of ask and bid
midprice = (ask + bid) / 2
# slices allow for cheaply extracting parts of an array
d_spread = spread[t:] - spread[:-t]
# find prices `t` steps in the future of d_spread
d_spread = d_spread[:-t]
future_price = midprice[2*t:]
util.log_info('D: %s, M: %s', d_spread.shape, future_price.shape)
# compute a univariate linear predictor
regression = mean(future_price / d_spread)
prediction = regression * d_spread
error = mean(abs(prediction - future_price))
return error
def find_change(arr, threshold=0.5): diff = abs(arr[1:] - arr[:-1]) diff = eager(diff) return diff[diff > threshold]
def setUp(self):
if not hasattr(self, 'current'):
self.current = eager(expr.abs(10 + expr.randn(10)))
self.strike = eager(expr.abs(20 + expr.randn(10)))
def setUp(self):
if not hasattr(self, 'current'):
self.current = eager(expr.abs(10 + expr.randn(10)))
self.strike = eager(expr.abs(20 + expr.randn(10)))
