def test_grad_apply():
num_visible_units = 100
num_hidden_units = 50
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.BernoulliLayer(num_visible_units)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = model.Model([vis_layer, hid_layer])
# create a gradient object filled with random numbers
grad = gu.random_grad(rbm)
gu.grad_apply(be.square, grad)
def test_bernoulli_GFE_derivatives():
# Tests that the GFE derivative update increases GFE versus 100
# random update vectors
num_units = 5
layer_1 = layers.BernoulliLayer(num_units)
layer_2 = layers.BernoulliLayer(num_units)
layer_3 = layers.BernoulliLayer(num_units)
rbm = BoltzmannMachine([layer_1, layer_2, layer_3])
for i in range(len(rbm.connections)):
rbm.connections[i].weights.params.matrix[:] = \
0.01 * be.randn(rbm.connections[i].shape)
for lay in rbm.layers:
lay.params.loc[:] = be.rand_like(lay.params.loc)
state, cop1_GFE = rbm.compute_StateTAP(init_lr=0.1, tol=1e-7, max_iters=50)
grad = rbm._grad_gibbs_free_energy(state)
gu.grad_normalize_(grad)
for i in range(100):
lr = 1.0
gogogo = True
random_grad = gu.random_grad(rbm)
gu.grad_normalize_(random_grad)
while gogogo:
cop1 = deepcopy(rbm)
lr_mul = partial(be.tmul, lr)
cop1.parameter_update(gu.grad_apply(lr_mul, grad))
cop1_state, cop1_GFE = cop1.compute_StateTAP(init_lr=0.1, tol=1e-7, max_iters=50)
cop2 = deepcopy(rbm)
cop2.parameter_update(gu.grad_apply(lr_mul, random_grad))
cop2_state, cop2_GFE = cop2.compute_StateTAP(init_lr=0.1, tol=1e-7, max_iters=50)
regress = cop2_GFE - cop1_GFE < 0.0
if regress:
if lr < 1e-6:
assert False, \
"TAP FE gradient is not working properly for Bernoulli models"
break
else:
lr *= 0.5
else:
break
def test_gaussian_GFE_derivatives_gradient_descent():
num_units = 5
layer_1 = layers.GaussianLayer(num_units)
layer_2 = layers.BernoulliLayer(num_units)
rbm = BoltzmannMachine([layer_1, layer_2])
for i in range(len(rbm.connections)):
rbm.connections[i].weights.params.matrix[:] = \
0.01 * be.randn(rbm.connections[i].shape)
for lay in rbm.layers:
lay.params.loc[:] = be.rand_like(lay.params.loc)
state, GFE = rbm.compute_StateTAP(use_GD=False, tol=1e-7, max_iters=50)
grad = rbm._grad_gibbs_free_energy(state)
gu.grad_normalize_(grad)
for i in range(100):
lr = 0.001
gogogo = True
random_grad = gu.random_grad(rbm)
gu.grad_normalize_(random_grad)
while gogogo:
cop1 = deepcopy(rbm)
lr_mul = partial(be.tmul, lr)
cop1.parameter_update(gu.grad_apply(lr_mul, grad))
cop1_state, cop1_GFE = cop1.compute_StateTAP(use_GD=False, tol=1e-7, max_iters=50)
cop2 = deepcopy(rbm)
cop2.parameter_update(gu.grad_apply(lr_mul, random_grad))
cop2_state, cop2_GFE = cop2.compute_StateTAP(use_GD=False, tol=1e-7, max_iters=50)
regress = cop2_GFE - cop1_GFE < 0
if regress:
if lr < 1e-6:
assert False, \
"TAP FE gradient is not working properly for Gaussian models"
break
else:
lr *= 0.5
else:
break
def test_bernoulli_GFE_derivatives():
num_units = 500
layer_1 = layers.BernoulliLayer(num_units)
layer_2 = layers.BernoulliLayer(num_units)
layer_3 = layers.BernoulliLayer(num_units)
rbm = model.Model([layer_1, layer_2, layer_3])
for i in range(len(rbm.weights)):
rbm.weights[i].params.matrix[:] = \
0.01 * be.randn(rbm.weights[i].shape)
for lay in rbm.layers:
lay.params.loc[:] = be.rand_like(lay.params.loc)
state = rbm.compute_StateTAP(init_lr=0.1, tol=1e-7, max_iters=50)
GFE = rbm.gibbs_free_energy(state)
lr = 0.1
gogogo = True
grad = rbm.grad_TAP_free_energy(0.1, 1e-7, 50)
while gogogo:
cop = deepcopy(rbm)
lr_mul = partial(be.tmul, -lr)
delta = gu.grad_apply(lr_mul, grad)
cop.parameter_update(delta)
cop_state = cop.compute_StateTAP(init_lr=0.1, tol=1e-7, max_iters=50)
cop_GFE = cop.gibbs_free_energy(cop_state)
regress = cop_GFE - GFE < 0.0
print(lr, cop_GFE, GFE, cop_GFE - GFE, regress)
if regress:
if lr < 1e-6:
assert False, \
"TAP FE gradient is not working properly for Bernoulli models"
break
else:
lr *= 0.5
else:
break