def get_Aopt(inX, iny):
X_train, y_train, X_test, y_test = ascdata.split_train_test(inX, iny)
X_train = np.concatenate((X_train, np.ones((X_train.shape[ 0 ], 1))), 1)
X_test = np.concatenate((X_test, np.ones((X_test.shape[ 0 ], 1))), 1)
X_train_less, s_train = ascdata.split_X_s(X_train)
X_test_less, s_test = ascdata.split_X_s(X_test)
s_train_phi = ascdata.generate_phi(s_train, d, A_phi, b_phi)
s_test_phi = ascdata.generate_phi(s_test, d, A_phi, b_phi)
nfeatures = X_train.shape[1] - 1
# Dimensions of phi(s)
nfeatures_phi = d
invT2 = 10
def logprob(inA, inX, iny, ins_phi):
RMS = 0
for i in range(len(iny)):
wi = np.dot(inA, inX[i])
RMS_current = (iny[i] - np.dot(wi, ins_phi[i]))**2
RMS += RMS_current
return -RMS
objective = lambda inA, t: -logprob(inA, X_train_less, y_train, s_train_phi)
LLHs = []
LLH_xs = []
def callback(params, t, g):
LLH = -objective(params, t)
LLHs.append(LLH)
LLH_xs.append(t)
print("Iteration {} log likelihood {}".format(t, LLH))
init_A = 0.00000000001*(np.ones((nfeatures_phi, nfeatures)))
# init_A = [[ -3.05236728e-04, -9.50015728e-04, -3.80139503e-04, 1.44010470e-04, -3.05236728e-04,
# -4.96117987e-04, -1.02736409e-04, -1.86416292e-04, -9.52628589e-04, -1.55023279e-03,
# 1.44717581e-04, 1.00000000e-11, -9.50028200e-04, -4.96117987e-04, 1.00000000e-11,
# -3.05236728e-04, 1.77416412e-06, -8.16665436e-06, 3.12622951e-05, -8.25700143e-04,
# 1.44627987e-04, 1.90211243e-05, -8.28273186e-04, -9.41349990e-04, -4.56671031e-04,
# 9.79097070e-03, -6.41866046e-04, -7.79274856e-05, 1.44539330e-04, -3.05236728e-04,
# -5.99188450e-04, -7.29470175e-04, -6.69558174e-04, -9.50028200e-04]]
init_A = np.array(init_A)
print("Optimizing network parameters...")
optimized_params = adam(grad(objective), init_A,
step_size=0.01, num_iters=1000, callback=callback)
Aopt = optimized_params
print "Aopt = ", Aopt
return Aopt, X_train_less, y_train, s_train, X_test_less, y_test, s_test, LLHs, LLH_xs
def get_predictions(Aopt, inX, s_bp):
s_bp_phi = ascdata.generate_phi(s_bp, d, A_phi, b_phi)
y_pred = []
for i in range(inX.shape[0]):
wi = np.dot(Aopt, inX[i])
y_pred.append(np.dot(wi.T, s_bp_phi[i]))
return y_pred
