def snd_derivative_pdf(self, y, m, s=None):
s = sigmoid(y * m, self.epsilon)
# return y * np.square(1 - s) * s - y * np.square(s) * (1 - s)
res = s * (1-s) * (1 - 2*s)
# print(np.any(np.isnan(res)))
return res
def fst_derivative_log_pdf(self, y, f, s=None):
#sigma = sigmoid(f, 1e-10)
sigma = sigmoid(f)
A = poisson_link(f)
res = sigma * (np.divide(y, A) - 1.0)
res = np.nan_to_num(res)
#res = np.maximum(res, 1e-8)
return res
def fst_derivative_log_pdf(self, y, m, s=None):
# ym = y * m
# print("ym: ", ym.shape)
# sym = sigmoid(-ym)
# print("sym: ", sym.shape)
# res = y * (1. - sigmoid(y * m))
# res = y * sigmoid(y * m)
res = np.multiply(y, sigmoid(-np.multiply(y, m)))
return res
def snd_derivative_log_pdf(self, y, f, s=None):
sigma = sigmoid(f)
A = poisson_link(f)
temp1 = sigma * (1 - sigma) * (np.divide(y, A) - 1)
temp2 = y * np.square(sigma) / np.square(A)
res = temp1 - temp2
res = np.nan_to_num(res)
#return np.minimum(0, res)
return res
def fst_derivative_pdf(self, y, m, S=None):
# pdf * sigmoid * (y/A(f) - 1)
A = poisson_link(m)
pdf = self.pdf(y, m, S)
sigm = sigmoid(m)
assert(sigm.shape == m.shape)
assert(A.shape == m.shape)
temp = np.subtract(np.divide(y, A), 1)
res = np.multiply(pdf, np.multiply(sigm, temp))
assert(not np.any(pds.isnull(res)))
return res
def snd_derivative_pdf(self, y, m, s=None):
pdf = self.pdf(y, poisson_link(m))
sigm = sigmoid(m)
A = poisson_link(m)
pmf_prime = self.fst_derivative_pdf(y, m)
assert(sigm.shape == m.shape)
assert(A.shape == m.shape)
# temp1 = (y/A - 1) * (pmf_prime * sigm + sigm*(1-sigm)*pmf)
# temp2 = pmf * sigm * (-y)/np.square(A) * sigm
# return 1/factorial(y) * (temp1 + temp2)
temp1 = np.multiply(pmf_prime, np.multiply(sigm, np.divide(y, A) - 1))
temp = np.multiply(-np.divide(y, np.square(A)), np.square(sigm)) \
+ np.multiply(y / A - 1, np.multiply(sigm, 1 - sigm))
temp2 = np.multiply(pdf, temp)
res = temp1 + temp2
assert(not np.any(pds.isnull(res)))
return res
def snd_derivative_log_pdf(self, y, m, s=None):
# return -np.square(y) * sigmoid(y * m) * sigmoid(-y * m)
return np.multiply(-sigmoid(np.multiply (y, m)), sigmoid(-np.multiply(y, m)))
def fst_derivative_pdf(self, y, m, s=None):
sigm = sigmoid(np.multiply(y,m), self.epsilon)
temp = np.multiply(y, sigm)
res = np.multiply(temp, np.subtract(1, sigm))
return res
def pdf(self, y, m, s=None):
res = sigmoid(np.multiply(y, m), self.epsilon)
# print(res)
# print(np.any(np.iszero(res)))
# print(not np.any(res))
return res
def sample(self, m, s=None, k=1):
return np.random.binomial(1, sigmoid(m), size=k)