def test_spherical_op_take_gradient(self):
print >> sys.stderr
print >> sys.stderr, "================================================="
print >> sys.stderr, "====== TEST spherical_op_take_gradient ==========="
print >> sys.stderr, "================================================="
V_mat, UT_mat, Uinv_mat, QT_mat, omega_vec, w_bar_vec = self.generate_model_params(
)
HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec = self.generate_data(
)
V = shared(V_mat)
UT = shared(UT_mat)
Uinv = shared(Uinv_mat)
QT = shared(QT_mat)
omega = shared(omega_vec)
w_bar = shared(w_bar_vec)
# Fiddling to debug:
# grad_AT_mat.fill(0.)
# grad_q_vec.fill(0.)
# grad_s_vec.fill(0.)
# grad_AT_mat[0,0] = 10000.
grad_AT = shared(grad_AT_mat)
grad_q = shared(grad_q_vec)
grad_s = shared(grad_s_vec)
HT = T.matrix()
KindexesT = T.imatrix()
eta_var = T.scalar()
# eta = shared(self.eta)
# compute AT,q,s with numpy (unfactorized algo)
W_mat = spherical_op.numpy_recompute_W(V_mat, UT_mat, omega_vec)
# numpy_out = spherical_op.numpy_fprop(W_mat, HT_mat, KindexesT_mat)
numpy_grad_HT, numpy_new_W = spherical_op.numpy_bprop_update(
W_mat, HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec,
self.eta)
# compute grad_HT with theano (factorized algo)
AT, q, s, work_d, work_m = spherical_op.FactoredSphericalOp(
eta_var, invup_mode=self.invup_mode)(V, UT, Uinv, QT, omega, w_bar,
HT, KindexesT)
def z_loss(AT, q, s, eps=1e-12):
D = self.D
mu = s / D
sigma = T.sqrt((q / D) - mu**2)
c = T.nnet.softplus((mu - AT[:, 0]) / (sigma + eps))
return c.sum()
def simple_loss(AT, q, s, eps=1e-5):
D = self.D
mu = s / D
sigma2 = (q / D) - mu**2 + eps
c = sigma2 + 0 * AT[:, 0]
return c.sum()
def simplest_loss(AT, q, s, eps=1e-5):
L = -AT[:, 0] + 0.7 * s + 0.3 * q
return L.sum()
L = z_loss(AT, q, s)
L = simplest_loss(AT, q, s)
# grad_HT = theano.grad(L, [sub] + params)
grad_HT = T.grad(L, wrt=HT)
# linked_grad_s = grad_s + 1e-9*s
# grad_HT = spherical_op.FactoredSphericalOpBpropUpdate(invup_mode=self.invup_mode)(
# V, UT, Uinv, QT, omega, w_bar,
# HT, KindexesT,
# work_d, work_m,
# grad_AT, grad_q, linked_grad_s,
# eta )
# pdb.set_trace()
g = function([HT, KindexesT, eta_var], [AT, q, s, L, grad_HT])
# g = function_dump("my_theano_function.dump",[HT, KindexesT, eta_var],[grad_HT])
theano_AT, theano_q, theano_s, theano_L, theano_grad_HT = g(
HT_mat, KindexesT_mat, self.eta)
print >> sys.stderr, "******** KindexesT ********"
print >> sys.stderr, KindexesT_mat
print >> sys.stderr, "******** theano AT, q, s, L ********"
print >> sys.stderr, theano_AT, theano_q, theano_s, theano_L
#print >> sys.stderr, "******** numpy grad_HT ********"
#print >> sys.stderr, numpy_grad_HT
print >> sys.stderr, "******** theano grad_HT ********"
print >> sys.stderr, theano_grad_HT
print >> sys.stderr, "******** old W ********"
print >> sys.stderr, W_mat
#print >> sys.stderr, "******** numpy new W ********"
#print >> sys.stderr, numpy_new_W
print >> sys.stderr, "******** theano new W ********"
new_W_mat = spherical_op.numpy_recompute_W(V.get_value(),
UT.get_value(),
omega.get_value())
print >> sys.stderr, new_W_mat
print >> sys.stderr, "******** consistency of op's new w_bar and W ********"
print >> sys.stderr, "w_bar before update: ", w_bar_vec
print >> sys.stderr, "sum of old W: ", W_mat.sum(axis=0)
print >> sys.stderr, "w_bar after update:", w_bar.get_value()
print >> sys.stderr, "sum of numpy new W:", new_W_mat.sum(axis=0)
print >> sys.stderr, "******** consistency of op's new QT and W ********"
print >> sys.stderr, "QT before update:"
print >> sys.stderr, QT_mat
print >> sys.stderr, "QT after update: (consider only upper triangular part)"
print >> sys.stderr, QT.get_value()
print >> sys.stderr, "numpy new W.T W:"
print >> sys.stderr, np.dot(new_W_mat.T, new_W_mat)
def test_spherical_op_take_gradient(self):
print >> sys.stderr
print >> sys.stderr, "================================================="
print >> sys.stderr, "====== TEST spherical_op_take_gradient ==========="
print >> sys.stderr, "================================================="
V_mat, UT_mat, Uinv_mat, QT_mat, omega_vec, w_bar_vec = self.generate_model_params()
HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec = self.generate_data()
V = shared(V_mat)
UT = shared(UT_mat)
Uinv = shared(Uinv_mat)
QT = shared(QT_mat)
omega = shared(omega_vec)
w_bar = shared(w_bar_vec)
# Fiddling to debug:
# grad_AT_mat.fill(0.)
# grad_q_vec.fill(0.)
# grad_s_vec.fill(0.)
# grad_AT_mat[0,0] = 10000.
grad_AT = shared(grad_AT_mat)
grad_q = shared(grad_q_vec)
grad_s = shared(grad_s_vec)
HT = T.matrix()
KindexesT = T.imatrix()
eta_var = T.scalar()
# eta = shared(self.eta)
# compute AT,q,s with numpy (unfactorized algo)
W_mat = spherical_op.numpy_recompute_W(V_mat, UT_mat, omega_vec)
# numpy_out = spherical_op.numpy_fprop(W_mat, HT_mat, KindexesT_mat)
numpy_grad_HT, numpy_new_W = spherical_op.numpy_bprop_update(W_mat, HT_mat, KindexesT_mat,
grad_AT_mat, grad_q_vec, grad_s_vec,
self.eta)
# compute grad_HT with theano (factorized algo)
AT, q, s, work_d, work_m = spherical_op.FactoredSphericalOp(eta_var, invup_mode=self.invup_mode)(V, UT, Uinv, QT, omega, w_bar,
HT, KindexesT)
def z_loss(AT, q, s, eps=1e-12):
D = self.D
mu = s / D
sigma = T.sqrt((q / D) - mu**2)
c = T.nnet.softplus((mu - AT[:, 0]) / (sigma + eps))
return c.sum()
def simple_loss(AT, q, s, eps=1e-5):
D = self.D
mu = s / D
sigma2 = (q / D) - mu**2 + eps
c = sigma2 + 0*AT[:,0]
return c.sum()
def simplest_loss(AT, q, s, eps=1e-5):
L = -AT[:,0]+0.7*s+0.3*q
return L.sum()
L = z_loss(AT,q,s)
L = simplest_loss(AT,q,s)
# grad_HT = theano.grad(L, [sub] + params)
grad_HT = T.grad(L, wrt=HT)
# linked_grad_s = grad_s + 1e-9*s
# grad_HT = spherical_op.FactoredSphericalOpBpropUpdate(invup_mode=self.invup_mode)(
# V, UT, Uinv, QT, omega, w_bar,
# HT, KindexesT,
# work_d, work_m,
# grad_AT, grad_q, linked_grad_s,
# eta )
# pdb.set_trace()
g = function([HT, KindexesT, eta_var],[AT,q,s,L,grad_HT])
# g = function_dump("my_theano_function.dump",[HT, KindexesT, eta_var],[grad_HT])
theano_AT, theano_q, theano_s, theano_L, theano_grad_HT = g(HT_mat, KindexesT_mat, self.eta)
print >> sys.stderr, "******** KindexesT ********"
print >> sys.stderr, KindexesT_mat
print >> sys.stderr, "******** theano AT, q, s, L ********"
print >> sys.stderr, theano_AT, theano_q, theano_s, theano_L
#print >> sys.stderr, "******** numpy grad_HT ********"
#print >> sys.stderr, numpy_grad_HT
print >> sys.stderr, "******** theano grad_HT ********"
print >> sys.stderr, theano_grad_HT
print >> sys.stderr, "******** old W ********"
print >> sys.stderr, W_mat
#print >> sys.stderr, "******** numpy new W ********"
#print >> sys.stderr, numpy_new_W
print >> sys.stderr, "******** theano new W ********"
new_W_mat = spherical_op.numpy_recompute_W(V.get_value(), UT.get_value(), omega.get_value())
print >> sys.stderr, new_W_mat
print >> sys.stderr, "******** consistency of op's new w_bar and W ********"
print >> sys.stderr, "w_bar before update: ", w_bar_vec
print >> sys.stderr, "sum of old W: ", W_mat.sum(axis=0)
print >> sys.stderr, "w_bar after update:", w_bar.get_value()
print >> sys.stderr, "sum of numpy new W:", new_W_mat.sum(axis=0)
print >> sys.stderr, "******** consistency of op's new QT and W ********"
print >> sys.stderr, "QT before update:"
print >> sys.stderr, QT_mat
print >> sys.stderr, "QT after update: (consider only upper triangular part)"
print >> sys.stderr, QT.get_value()
print >> sys.stderr, "numpy new W.T W:"
print >> sys.stderr, np.dot(new_W_mat.T, new_W_mat)
def test_spherical_op_grad(self):
print >> sys.stderr
print >> sys.stderr, "================================================="
print >> sys.stderr, "============ TEST spherical_op_grad ==========="
print >> sys.stderr, "================================================="
V_mat, UT_mat, Uinv_mat, QT_mat, omega_vec, w_bar_vec = self.generate_model_params(
)
HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec = self.generate_data(
)
V = shared(V_mat)
UT = shared(UT_mat)
Uinv = shared(Uinv_mat)
QT = shared(QT_mat)
omega = shared(omega_vec)
w_bar = shared(w_bar_vec)
# Fiddling to debug:
# grad_AT_mat.fill(0.)
# grad_q_vec.fill(0.)
# grad_s_vec.fill(0.)
# grad_AT_mat[0,0] = 10000.
grad_AT = shared(grad_AT_mat)
grad_q = shared(grad_q_vec)
grad_s = shared(grad_s_vec)
HT = T.matrix()
KindexesT = T.imatrix()
eta_var = T.scalar()
# compute AT,q,s with numpy (unfactorized algo)
W_mat = spherical_op.numpy_recompute_W(V_mat, UT_mat, omega_vec)
# numpy_out = spherical_op.numpy_fprop(W_mat, HT_mat, KindexesT_mat)
numpy_grad_HT, numpy_new_W = spherical_op.numpy_bprop_update(
W_mat, HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec,
self.eta)
# compute grad_HT with theano (factorized algo)
AT, q, s, work_d, work_m = spherical_op.FactoredSphericalOp(
eta_var, invup_mode=self.invup_mode)(V, UT, Uinv, QT, omega, w_bar,
HT, KindexesT)
linked_grad_s = grad_s + 1e-9 * s
grad_HT = spherical_op.FactoredSphericalOpBpropUpdate(
invup_mode=self.invup_mode)(V, UT, Uinv, QT, omega, w_bar, HT,
KindexesT, work_d, work_m, grad_AT,
grad_q, linked_grad_s, eta_var)
# pdb.set_trace()
g = function([HT, KindexesT, eta_var], [grad_HT])
# g = function_dump("my_theano_function.dump",[HT, KindexesT, eta],[grad_HT])
theano_grad_HT = g(HT_mat, KindexesT_mat, self.eta)
# # compute AT,q,s with theano (factorized algo)
# f = function([HT, KindexesT],
# spherical_op.FactoredSphericalOp(eta)(V, UT, Uinv, QT, omega, w_bar,
# HT, KindexesT))
# theano_AT, theano_q, theano_s = f(HT_mat, KindexesT_mat)
# g = function([HT, KindexesT, grad_AT, grad_q, grad_s, eta ],
# spherical_op.FactoredSphericalOpBpropUpdate()(V, UT, Uinv, QT, omega, w_bar,
# HT, KindexesT,
# grad_AT, grad_q, grad_s,
# eta ))
# theano_grad_HT = g(HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec, self.eta)
print >> sys.stderr, "******** KindexesT ********"
print >> sys.stderr, KindexesT_mat
print >> sys.stderr, "******** numpy grad_HT ********"
print >> sys.stderr, numpy_grad_HT
print >> sys.stderr, "******** theano grad_HT ********"
print >> sys.stderr, theano_grad_HT
print >> sys.stderr, "******** old W ********"
print >> sys.stderr, W_mat
print >> sys.stderr, "******** numpy new W ********"
print >> sys.stderr, numpy_new_W
print >> sys.stderr, "******** theano new W ********"
new_W_mat = spherical_op.numpy_recompute_W(V.get_value(),
UT.get_value(),
omega.get_value())
print >> sys.stderr, new_W_mat
print >> sys.stderr, "******** consistency of op's new w_bar and W ********"
print >> sys.stderr, "w_bar before update: ", w_bar_vec
print >> sys.stderr, "sum of old W: ", W_mat.sum(axis=0)
print >> sys.stderr, "w_bar after update:", w_bar.get_value()
print >> sys.stderr, "sum of numpy new W:", new_W_mat.sum(axis=0)
print >> sys.stderr, "******** consistency of op's new QT and W ********"
print >> sys.stderr, "QT before update:"
print >> sys.stderr, QT_mat
print >> sys.stderr, "QT after update: (consider only upper triangular part)"
print >> sys.stderr, QT.get_value()
print >> sys.stderr, "numpy new W.T W:"
print >> sys.stderr, np.dot(new_W_mat.T, new_W_mat)
def test_spherical_op_grad(self):
print >> sys.stderr
print >> sys.stderr, "================================================="
print >> sys.stderr, "============ TEST spherical_op_grad ==========="
print >> sys.stderr, "================================================="
V_mat, UT_mat, Uinv_mat, QT_mat, omega_vec, w_bar_vec = self.generate_model_params()
HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec = self.generate_data()
V = shared(V_mat)
UT = shared(UT_mat)
Uinv = shared(Uinv_mat)
QT = shared(QT_mat)
omega = shared(omega_vec)
w_bar = shared(w_bar_vec)
# Fiddling to debug:
# grad_AT_mat.fill(0.)
# grad_q_vec.fill(0.)
# grad_s_vec.fill(0.)
# grad_AT_mat[0,0] = 10000.
grad_AT = shared(grad_AT_mat)
grad_q = shared(grad_q_vec)
grad_s = shared(grad_s_vec)
HT = T.matrix()
KindexesT = T.imatrix()
eta_var = T.scalar()
# compute AT,q,s with numpy (unfactorized algo)
W_mat = spherical_op.numpy_recompute_W(V_mat, UT_mat, omega_vec)
# numpy_out = spherical_op.numpy_fprop(W_mat, HT_mat, KindexesT_mat)
numpy_grad_HT, numpy_new_W = spherical_op.numpy_bprop_update(W_mat, HT_mat, KindexesT_mat,
grad_AT_mat, grad_q_vec, grad_s_vec,
self.eta)
# compute grad_HT with theano (factorized algo)
AT, q, s, work_d, work_m = spherical_op.FactoredSphericalOp(eta_var, invup_mode=self.invup_mode)(V, UT, Uinv, QT, omega, w_bar,
HT, KindexesT)
linked_grad_s = grad_s + 1e-9*s
grad_HT = spherical_op.FactoredSphericalOpBpropUpdate(invup_mode=self.invup_mode)(
V, UT, Uinv, QT, omega, w_bar,
HT, KindexesT,
work_d, work_m,
grad_AT, grad_q, linked_grad_s,
eta_var )
# pdb.set_trace()
g = function([HT, KindexesT, eta_var],[grad_HT])
# g = function_dump("my_theano_function.dump",[HT, KindexesT, eta],[grad_HT])
theano_grad_HT = g(HT_mat, KindexesT_mat, self.eta)
# # compute AT,q,s with theano (factorized algo)
# f = function([HT, KindexesT],
# spherical_op.FactoredSphericalOp(eta)(V, UT, Uinv, QT, omega, w_bar,
# HT, KindexesT))
# theano_AT, theano_q, theano_s = f(HT_mat, KindexesT_mat)
# g = function([HT, KindexesT, grad_AT, grad_q, grad_s, eta ],
# spherical_op.FactoredSphericalOpBpropUpdate()(V, UT, Uinv, QT, omega, w_bar,
# HT, KindexesT,
# grad_AT, grad_q, grad_s,
# eta ))
# theano_grad_HT = g(HT_mat, KindexesT_mat, grad_AT_mat, grad_q_vec, grad_s_vec, self.eta)
print >> sys.stderr, "******** KindexesT ********"
print >> sys.stderr, KindexesT_mat
print >> sys.stderr, "******** numpy grad_HT ********"
print >> sys.stderr, numpy_grad_HT
print >> sys.stderr, "******** theano grad_HT ********"
print >> sys.stderr, theano_grad_HT
print >> sys.stderr, "******** old W ********"
print >> sys.stderr, W_mat
print >> sys.stderr, "******** numpy new W ********"
print >> sys.stderr, numpy_new_W
print >> sys.stderr, "******** theano new W ********"
new_W_mat = spherical_op.numpy_recompute_W(V.get_value(), UT.get_value(), omega.get_value())
print >> sys.stderr, new_W_mat
print >> sys.stderr, "******** consistency of op's new w_bar and W ********"
print >> sys.stderr, "w_bar before update: ", w_bar_vec
print >> sys.stderr, "sum of old W: ", W_mat.sum(axis=0)
print >> sys.stderr, "w_bar after update:", w_bar.get_value()
print >> sys.stderr, "sum of numpy new W:", new_W_mat.sum(axis=0)
print >> sys.stderr, "******** consistency of op's new QT and W ********"
print >> sys.stderr, "QT before update:"
print >> sys.stderr, QT_mat
print >> sys.stderr, "QT after update: (consider only upper triangular part)"
print >> sys.stderr, QT.get_value()
print >> sys.stderr, "numpy new W.T W:"
print >> sys.stderr, np.dot(new_W_mat.T, new_W_mat)