def test_bernoulli_derivatives():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# 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 = hidden.Model([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units, ))
b = be.randn((num_hidden_units, ))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].int_params.loc[:] = a
rbm.layers[1].int_params.loc[:] = b
rbm.weights[0].int_params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
vdata_scaled = rbm.layers[0].rescale(vdata)
# compute the mean of the hidden layer
rbm.layers[1].update([vdata], [rbm.weights[0].W()])
hid_mean = rbm.layers[1].mean()
hid_mean_scaled = rbm.layers[1].rescale(hid_mean)
# compute the derivatives
d_visible_loc = -be.mean(vdata, axis=0)
d_hidden_loc = -be.mean(hid_mean_scaled, axis=0)
d_W = -be.batch_outer(vdata, hid_mean_scaled) / len(vdata)
# compute the derivatives using the layer functions
vis_derivs = rbm.layers[0].derivatives(vdata, [hid_mean_scaled],
[rbm.weights[0].W()])
hid_derivs = rbm.layers[1].derivatives(hid_mean, [vdata_scaled],
[rbm.weights[0].W_T()])
weight_derivs = rbm.weights[0].derivatives(vdata, hid_mean_scaled)
assert be.allclose(d_visible_loc, vis_derivs.loc), \
"derivative of visible loc wrong in bernoulli-bernoulli rbm"
assert be.allclose(d_hidden_loc, hid_derivs.loc), \
"derivative of hidden loc wrong in bernoulli-bernoulli rbm"
assert be.allclose(d_W, weight_derivs.matrix), \
"derivative of weights wrong in bernoulli-bernoulli rbm"
def test_state_for_grad_DrivenSequentialMC():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# 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])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units, ))
b = be.randn((num_hidden_units, ))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.weights[0].params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
data_state = State.from_visible(vdata, rbm)
dropout_scale = State.dropout_rescale(rbm)
# since we set no dropout, dropout_scale should be None
assert dropout_scale is None
for u in [
'markov_chain', 'mean_field_iteration', 'deterministic_iteration'
]:
# set up the sampler
sampler = fit.DrivenSequentialMC(rbm, updater=u, clamped=[0])
sampler.set_state(data_state)
# update the state of the hidden layer
grad_state = sampler.state_for_grad(1, dropout_scale)
assert be.allclose(data_state.units[0], grad_state.units[0]), \
"visible layer is clamped, and shouldn't get updated: {}".format(u)
assert not be.allclose(data_state.units[1], grad_state.units[1]), \
"hidden layer is not clamped, and should get updated: {}".format(u)
# compute the conditional mean with the layer function
ave = rbm.layers[1].conditional_mean(
rbm._connected_rescaled_units(1, data_state, dropout_scale),
rbm._connected_weights(1))
assert be.allclose(ave, grad_state.units[1]), \
"hidden layer of grad_state should be conditional mean: {}".format(u)
def test_random_grad():
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
gu.random_grad(rbm)
def test_zero_grad():
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 = BoltzmannMachine([vis_layer, hid_layer])
# create a gradient object filled with zeros
gu.zero_grad(rbm)
def test_find_k_nearest_neighbors():
n = 20
shp = (20, n)
perm = be.rand_int(0, 20, (20, ))
k = 1
be.set_seed()
y = be.randn(shp)
x = y[perm]
indices, _distances = math_utils.find_k_nearest_neighbors(x, y, k)
assert be.allclose(indices, perm)
assert be.allclose(_distances, be.zeros((20, )), 1e-2, 1e-2)
def test_grad_accumulate():
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 = BoltzmannMachine([vis_layer, hid_layer])
# create a gradient object filled with random numbers
grad = gu.random_grad(rbm)
gu.grad_accumulate(be.norm, grad)
def test_clamped_SequentialMC():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
steps = 1
# 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])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units, ))
b = be.randn((num_hidden_units, ))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.weights[0].params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
data_state = State.from_visible(vdata, rbm)
dropout_scale = State.dropout_rescale(rbm)
# since we set no dropout, dropout_scale should be None
assert dropout_scale is None
for u in [
'markov_chain', 'mean_field_iteration', 'deterministic_iteration'
]:
# set up the sampler with the visible layer clamped
sampler = fit.SequentialMC(rbm, updater=u, clamped=[0])
sampler.set_state(data_state)
# update the sampler state and check the output
sampler.update_state(steps, dropout_scale)
assert be.allclose(data_state.units[0], sampler.state.units[0]), \
"visible layer is clamped, and shouldn't get updated: {}".format(u)
assert not be.allclose(data_state.units[1], sampler.state.units[1]), \
"hidden layer is not clamped, and should get updated: {}".format(u)
def test_exponential_conditional_params():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.ExponentialLayer(num_visible_units)
hid_layer = layers.ExponentialLayer(num_hidden_units)
rbm = model.Model([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
# for the exponential layers, we need a > 0, b > 0, and W < 0
a = be.rand((num_visible_units, ))
b = be.rand((num_hidden_units, ))
W = -be.rand((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.weights[0].params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
hdata = rbm.layers[1].random((batch_size, num_hidden_units))
# compute conditional parameters
hidden_rate = -be.dot(vdata, W) # (batch_size, num_hidden_units)
hidden_rate += be.broadcast(b, hidden_rate)
visible_rate = -be.dot(hdata,
be.transpose(W)) # (batch_size, num_visible_units)
visible_rate += be.broadcast(a, visible_rate)
# compute the conditional parameters using the layer functions
hidden_rate_func = rbm.layers[1]._conditional_params([vdata],
[rbm.weights[0].W()])
visible_rate_func = rbm.layers[0]._conditional_params(
[hdata], [rbm.weights[0].W_T()])
assert be.allclose(hidden_rate, hidden_rate_func), \
"hidden rate wrong in exponential-exponential rbm"
assert be.allclose(visible_rate, visible_rate_func), \
"visible rate wrong in exponential-exponential rbm"
def test_bernoulli_conditional_params():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# 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])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units, ))
b = be.randn((num_hidden_units, ))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.weights[0].params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
hdata = rbm.layers[1].random((batch_size, num_hidden_units))
# compute conditional parameters
hidden_field = be.dot(vdata, W) # (batch_size, num_hidden_units)
hidden_field += b
visible_field = be.dot(hdata,
be.transpose(W)) # (batch_size, num_visible_units)
visible_field += a
# compute conditional parameters with layer funcitons
hidden_field_layer = rbm.layers[1]._conditional_params(
[vdata], [rbm.weights[0].W()])
visible_field_layer = rbm.layers[0]._conditional_params(
[hdata], [rbm.weights[0].W_T()])
assert be.allclose(hidden_field, hidden_field_layer), \
"hidden field wrong in bernoulli-bernoulli rbm"
assert be.allclose(visible_field, visible_field_layer), \
"visible field wrong in bernoulli-bernoulli rbm"
def test_grad_normalize_():
num_visible_units = 10
num_hidden_units = 10
# 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 = BoltzmannMachine([vis_layer, hid_layer])
# create a gradient object filled with random numbers
grad = gu.random_grad(rbm)
gu.grad_normalize_(grad)
nrm = gu.grad_norm(grad)
assert nrm > 1-1e-6
assert nrm < 1+1e-6
def test_exponential_update():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.ExponentialLayer(num_visible_units)
hid_layer = layers.ExponentialLayer(num_hidden_units)
rbm = hidden.Model([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
# for the exponential layers, we need a > 0, b > 0, and W < 0
a = be.rand((num_visible_units, ))
b = be.rand((num_hidden_units, ))
W = -be.rand((num_visible_units, num_hidden_units))
rbm.layers[0].int_params['loc'] = a
rbm.layers[1].int_params['loc'] = b
rbm.weights[0].int_params['matrix'] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
hdata = rbm.layers[1].random((batch_size, num_hidden_units))
# compute extrinsic parameters
hidden_rate = -be.dot(vdata, W) # (batch_size, num_hidden_units)
hidden_rate += be.broadcast(b, hidden_rate)
visible_rate = -be.dot(hdata,
be.transpose(W)) # (batch_size, num_visible_units)
visible_rate += be.broadcast(a, visible_rate)
# update the extrinsic parameter using the layer functions
rbm.layers[1].update(vdata, rbm.weights[0].W())
rbm.layers[0].update(hdata, be.transpose(rbm.weights[0].W()))
assert be.allclose(hidden_rate, rbm.layers[1].ext_params['rate']), \
"hidden rate wrong in exponential-exponential rbm"
assert be.allclose(visible_rate, rbm.layers[0].ext_params['rate']), \
"visible rate wrong in exponential-exponential rbm"
def test_bernoulli_update():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# 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 = hidden.Model([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units, ))
b = be.randn((num_hidden_units, ))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].int_params.loc[:] = a
rbm.layers[1].int_params.loc[:] = b
rbm.weights[0].int_params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
hdata = rbm.layers[1].random((batch_size, num_hidden_units))
# compute extrinsic parameters
hidden_field = be.dot(vdata, W) # (batch_size, num_hidden_units)
hidden_field += be.broadcast(b, hidden_field)
visible_field = be.dot(hdata,
be.transpose(W)) # (batch_size, num_visible_units)
visible_field += be.broadcast(a, visible_field)
# update the extrinsic parameter using the layer functions
rbm.layers[0].update([hdata], [rbm.weights[0].W_T()])
rbm.layers[1].update([vdata], [rbm.weights[0].W()])
assert be.allclose(hidden_field, rbm.layers[1].ext_params.field), \
"hidden field wrong in bernoulli-bernoulli rbm"
assert be.allclose(visible_field, rbm.layers[0].ext_params.field), \
"visible field wrong in bernoulli-bernoulli rbm"
def test_pdist():
n = 500
a_shape = (1000, n)
b_shape = (1000, n)
# distance distributions
a_mean, a_scale = 1, 1
b_mean, b_scale = -1, 1
be.set_seed()
a = a_mean + a_scale * be.randn(a_shape)
b = b_mean + b_scale * be.randn(b_shape)
dists = math_utils.pdist(a, b)
dists_t = math_utils.pdist(b, a)
assert be.shape(dists) == (1000, 1000)
assert be.allclose(be.transpose(dists_t), dists)
assert be.mean(dists) > 2 * math.sqrt(n) and be.mean(
dists) < 3 * math.sqrt(n)
def test_grad_norm():
num_visible_units = 1000
num_hidden_units = 1000
# 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 = BoltzmannMachine([vis_layer, hid_layer])
# create a gradient object filled with random numbers
grad = gu.random_grad(rbm)
nrm = gu.grad_norm(grad)
assert nrm > math.sqrt(be.float_scalar(num_hidden_units + \
num_visible_units + num_visible_units*num_hidden_units)/3) - 1
assert nrm < math.sqrt(be.float_scalar(num_hidden_units + \
num_visible_units + num_visible_units*num_hidden_units)/3) + 1
def test_unclamped_DrivenSequentialMC():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
steps = 1
# 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 = BoltzmannMachine([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units,))
b = be.randn((num_hidden_units,))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.connections[0].weights.params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
data_state = State.from_visible(vdata, rbm)
for u in ['markov_chain', 'mean_field_iteration', 'deterministic_iteration']:
# set up the sampler with the visible layer clamped
sampler = samplers.SequentialMC(rbm, updater=u)
sampler.set_state(data_state)
# update the sampler state and check the output
sampler.update_state(steps)
assert not be.allclose(data_state[0], sampler.state[0]), \
"visible layer is not clamped, and should get updated: {}".format(u)
assert not be.allclose(data_state[1], sampler.state[1]), \
"hidden layer is not clamped, and should get updated: {}".format(u)
def test_gaussian_derivatives():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.GaussianLayer(num_visible_units)
hid_layer = layers.GaussianLayer(num_hidden_units)
rbm = hidden.Model([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units, ))
b = be.randn((num_hidden_units, ))
log_var_a = 0.1 * be.randn((num_visible_units, ))
log_var_b = 0.1 * be.randn((num_hidden_units, ))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].int_params.loc[:] = a
rbm.layers[1].int_params.loc[:] = b
rbm.layers[0].int_params.log_var[:] = log_var_a
rbm.layers[1].int_params.log_var[:] = log_var_b
rbm.weights[0].int_params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
visible_var = be.exp(log_var_a)
vdata_scaled = vdata / be.broadcast(visible_var, vdata)
# compute the mean of the hidden layer
rbm.layers[1].update([vdata_scaled], [rbm.weights[0].W()])
hidden_var = be.exp(log_var_b)
hid_mean = rbm.layers[1].mean()
hid_mean_scaled = rbm.layers[1].rescale(hid_mean)
# compute the derivatives
d_vis_loc = -be.mean(vdata_scaled, axis=0)
d_vis_logvar = -0.5 * be.mean(be.square(be.subtract(a, vdata)), axis=0)
d_vis_logvar += be.batch_dot(
hid_mean_scaled, be.transpose(W), vdata, axis=0) / len(vdata)
d_vis_logvar /= visible_var
d_hid_loc = -be.mean(hid_mean_scaled, axis=0)
d_hid_logvar = -0.5 * be.mean(
be.square(hid_mean - be.broadcast(b, hid_mean)), axis=0)
d_hid_logvar += be.batch_dot(vdata_scaled, W, hid_mean,
axis=0) / len(hid_mean)
d_hid_logvar /= hidden_var
d_W = -be.batch_outer(vdata_scaled, hid_mean_scaled) / len(vdata_scaled)
# compute the derivatives using the layer functions
rbm.layers[1].update([vdata_scaled], [rbm.weights[0].W()])
rbm.layers[0].update([hid_mean_scaled], [rbm.weights[0].W_T()])
vis_derivs = rbm.layers[0].derivatives(vdata, [hid_mean_scaled],
[rbm.weights[0].W()])
hid_derivs = rbm.layers[1].derivatives(hid_mean, [vdata_scaled],
[rbm.weights[0].W_T()])
weight_derivs = rbm.weights[0].derivatives(vdata_scaled, hid_mean_scaled)
assert be.allclose(d_vis_loc, vis_derivs.loc), \
"derivative of visible loc wrong in gaussian-gaussian rbm"
assert be.allclose(d_hid_loc, hid_derivs.loc), \
"derivative of hidden loc wrong in gaussian-gaussian rbm"
assert be.allclose(d_vis_logvar, vis_derivs.log_var, rtol=1e-05, atol=1e-01), \
"derivative of visible log_var wrong in gaussian-gaussian rbm"
assert be.allclose(d_hid_logvar, hid_derivs.log_var, rtol=1e-05, atol=1e-01), \
"derivative of hidden log_var wrong in gaussian-gaussian rbm"
assert be.allclose(d_W, weight_derivs.matrix), \
"derivative of weights wrong in gaussian-gaussian rbm"
def test_gaussian_update():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.GaussianLayer(num_visible_units)
hid_layer = layers.GaussianLayer(num_hidden_units)
rbm = hidden.Model([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units, ))
b = be.randn((num_hidden_units, ))
log_var_a = 0.1 * be.randn((num_visible_units, ))
log_var_b = 0.1 * be.randn((num_hidden_units, ))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].int_params.loc[:] = a
rbm.layers[1].int_params.loc[:] = b
rbm.layers[0].int_params.log_var[:] = log_var_a
rbm.layers[1].int_params.log_var[:] = log_var_b
rbm.weights[0].int_params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
hdata = rbm.layers[1].random((batch_size, num_hidden_units))
# compute the variance
visible_var = be.exp(log_var_a)
hidden_var = be.exp(log_var_b)
# rescale the data
vdata_scaled = vdata / be.broadcast(visible_var, vdata)
hdata_scaled = hdata / be.broadcast(hidden_var, hdata)
# test rescale
assert be.allclose(vdata_scaled, rbm.layers[0].rescale(vdata)),\
"visible rescale wrong in gaussian-gaussian rbm"
assert be.allclose(hdata_scaled, rbm.layers[1].rescale(hdata)),\
"hidden rescale wrong in gaussian-gaussian rbm"
# compute the mean
hidden_mean = be.dot(vdata_scaled, W) # (batch_size, num_hidden_units)
hidden_mean += be.broadcast(b, hidden_mean)
visible_mean = be.dot(hdata_scaled,
be.transpose(W)) # (batch_size, num_hidden_units)
visible_mean += be.broadcast(a, visible_mean)
# update the extrinsic parameters using the layer functions
rbm.layers[0].update([hdata_scaled], [rbm.weights[0].W_T()])
rbm.layers[1].update([vdata_scaled], [rbm.weights[0].W()])
assert be.allclose(visible_var, rbm.layers[0].ext_params.variance),\
"visible variance wrong in gaussian-gaussian rbm"
assert be.allclose(hidden_var, rbm.layers[1].ext_params.variance),\
"hidden variance wrong in gaussian-gaussian rbm"
assert be.allclose(visible_mean, rbm.layers[0].ext_params.mean),\
"visible mean wrong in gaussian-gaussian rbm"
assert be.allclose(hidden_mean, rbm.layers[1].ext_params.mean),\
"hidden mean wrong in gaussian-gaussian rbm"
def test_conditional_sampling():
"""
Test sampling from one layer conditioned on the state of another layer.
Note:
This test compares values estimated by *sampling* to values computed
analytically. It can fail for small batch_size, or strict tolerances,
even if everything is working propery.
"""
num_visible_units = 20
num_hidden_units = 10
steps = 1000
mean_tol = 0.1
# set a seed for the random number generator
be.set_seed()
layer_types = [
layers.BernoulliLayer,
layers.GaussianLayer]
for layer_type in layer_types:
# set up some layer and model objects
vis_layer = layer_type(num_visible_units)
hid_layer = layer_type(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.rand((num_visible_units,))
b = be.rand((num_hidden_units,))
W = 10 * be.rand((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.connections[0].weights.params.matrix[:] = W
if layer_type == layers.GaussianLayer:
log_var_a = be.randn((num_visible_units,))
log_var_b = be.randn((num_hidden_units,))
rbm.layers[0].params.log_var[:] = log_var_a
rbm.layers[1].params.log_var[:] = log_var_b
# initialize a state
state = State.from_model(1, rbm)
# set up a calculator for the moments
moments = mu.MeanVarianceArrayCalculator()
for _ in range(steps):
moments.update(rbm.layers[0].conditional_sample(
rbm._connected_rescaled_units(0, state),
rbm._connected_weights(0)))
model_mean = rbm.layers[0].conditional_mean(
rbm._connected_rescaled_units(0, state),
rbm._connected_weights(0))
ave = moments.mean
close = be.allclose(ave, model_mean[0], rtol=mean_tol, atol=mean_tol)
assert close, "{} conditional mean".format(layer_type)
if layer_type == layers.GaussianLayer:
model_mean, model_var = rbm.layers[0].conditional_params(
rbm._connected_rescaled_units(0, state),
rbm._connected_weights(0))
close = be.allclose(be.sqrt(moments.var), be.sqrt(model_var[0]), rtol=mean_tol, atol=mean_tol)
assert close, "{} conditional standard deviation".format(layer_type)
def test_tap_machine(paysage_path=None):
num_hidden_units = 10
batch_size = 100
num_epochs = 5
learning_rate = schedules.PowerLawDecay(initial=0.1, coefficient=1.0)
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'examples', 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'examples', 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
samples = pre.binarize_color(
be.float_tensor(
pandas.read_hdf(shuffled_filepath,
key='train/images').as_matrix()[:10000]))
samples_train, samples_validate = batch.split_tensor(samples, 0.95)
data = batch.Batch({
'train':
batch.InMemoryTable(samples_train, batch_size),
'validate':
batch.InMemoryTable(samples_validate, batch_size)
})
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = ProgressMonitor(generator_metrics = \
[ReconstructionError(), TAPLogLikelihood(10), TAPFreeEnergy(10)])
untrained_performance = perf.epoch_update(data,
rbm,
store=True,
show=False)
# set up the optimizer and the fit method
opt = optimizers.Gradient(stepsize=learning_rate, tolerance=1e-5)
tap = fit.TAP(True, 0.1, 0.01, 25, True, 0.5, 0.001, 0.0)
solver = fit.SGD(rbm, data)
solver.monitor.generator_metrics.append(TAPLogLikelihood(10))
solver.monitor.generator_metrics.append(TAPFreeEnergy(10))
# fit the model
print('training with stochastic gradient ascent')
solver.train(opt, num_epochs, method=tap.tap_update)
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = solver.monitor.memory[-1]
assert (trained_performance['TAPLogLikelihood'] >
untrained_performance['TAPLogLikelihood']), \
"TAP log-likelihood did not increase"
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_onehot_derivatives():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.OneHotLayer(num_visible_units)
hid_layer = layers.OneHotLayer(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units,))
b = be.randn((num_hidden_units,))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.connections[0].weights.params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
vdata_scaled = rbm.layers[0].rescale(vdata)
# compute the conditional mean of the hidden layer
hid_mean = rbm.layers[1].conditional_mean([vdata], [rbm.connections[0].W()])
hid_mean_scaled = rbm.layers[1].rescale(hid_mean)
# compute the derivatives
d_visible_loc = -be.mean(vdata, axis=0)
d_hidden_loc = -be.mean(hid_mean_scaled, axis=0)
d_W = -be.batch_outer(vdata, hid_mean_scaled) / len(vdata)
# compute the derivatives using the layer functions
vis_derivs = rbm.layers[0].derivatives(vdata, [hid_mean_scaled],
[rbm.connections[0].W(trans=True)])
hid_derivs = rbm.layers[1].derivatives(hid_mean, [vdata_scaled],
[rbm.connections[0].W()])
weight_derivs = rbm.connections[0].weights.derivatives(vdata, hid_mean_scaled)
# compute simple weighted derivatives using the layer functions
scale = 2
scale_func = partial(be.multiply, be.float_scalar(scale))
vis_derivs_scaled = rbm.layers[0].derivatives(vdata, [hid_mean_scaled],
[rbm.connections[0].W(trans=True)], weighting_function=scale_func)
hid_derivs_scaled = rbm.layers[1].derivatives(hid_mean, [vdata_scaled],
[rbm.connections[0].W()], weighting_function=scale_func)
weight_derivs_scaled = rbm.connections[0].weights.derivatives(vdata, hid_mean_scaled,
weighting_function=scale_func)
assert be.allclose(d_visible_loc, vis_derivs[0].loc), \
"derivative of visible loc wrong in onehot-onehot rbm"
assert be.allclose(d_hidden_loc, hid_derivs[0].loc), \
"derivative of hidden loc wrong in onehot-onehot rbm"
assert be.allclose(d_W, weight_derivs[0].matrix), \
"derivative of weights wrong in onehot-onehot rbm"
assert be.allclose(scale * d_visible_loc, vis_derivs_scaled[0].loc), \
"weighted derivative of visible loc wrong in onehot-onehot rbm"
assert be.allclose(scale * d_hidden_loc, hid_derivs_scaled[0].loc), \
"weighted derivative of hidden loc wrong in onehot-onehot rbm"
assert be.allclose(scale * d_W, weight_derivs_scaled[0].matrix), \
"weighted derivative of weights wrong in onehot-onehot rbm"
def test_gaussian_derivatives():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.GaussianLayer(num_visible_units)
hid_layer = layers.GaussianLayer(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units,))
b = be.randn((num_hidden_units,))
log_var_a = 0.1 * be.randn((num_visible_units,))
log_var_b = 0.1 * be.randn((num_hidden_units,))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.layers[0].params.log_var[:] = log_var_a
rbm.layers[1].params.log_var[:] = log_var_b
rbm.connections[0].weights.params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
visible_var = be.exp(log_var_a)
vdata_scaled = vdata / visible_var
# compute the mean of the hidden layer
hid_mean = rbm.layers[1].conditional_mean(
[vdata_scaled], [rbm.connections[0].W()])
hidden_var = be.exp(log_var_b)
hid_mean_scaled = rbm.layers[1].rescale(hid_mean)
# compute the derivatives
d_vis_loc = be.mean((a-vdata)/visible_var, axis=0)
d_vis_logvar = -0.5 * be.mean(be.square(be.subtract(a, vdata)), axis=0)
d_vis_logvar += be.batch_quadratic(hid_mean_scaled, be.transpose(W), vdata,
axis=0) / len(vdata)
d_vis_logvar /= visible_var
d_hid_loc = be.mean((b-hid_mean)/hidden_var, axis=0)
d_hid_logvar = -0.5 * be.mean(be.square(hid_mean - b), axis=0)
d_hid_logvar += be.batch_quadratic(vdata_scaled, W, hid_mean,
axis=0) / len(hid_mean)
d_hid_logvar /= hidden_var
d_W = -be.batch_outer(vdata_scaled, hid_mean_scaled) / len(vdata_scaled)
# compute the derivatives using the layer functions
vis_derivs = rbm.layers[0].derivatives(vdata, [hid_mean_scaled],
[rbm.connections[0].W(trans=True)])
hid_derivs = rbm.layers[1].derivatives(hid_mean, [vdata_scaled],
[rbm.connections[0].W()])
weight_derivs = rbm.connections[0].weights.derivatives(vdata_scaled, hid_mean_scaled)
# compute simple weighted derivatives using the layer functions
scale = 2
scale_func = partial(be.multiply, be.float_scalar(scale))
vis_derivs_scaled = rbm.layers[0].derivatives(vdata, [hid_mean_scaled],
[rbm.connections[0].W(trans=True)], weighting_function=scale_func)
hid_derivs_scaled = rbm.layers[1].derivatives(hid_mean, [vdata_scaled],
[rbm.connections[0].W()], weighting_function=scale_func)
weight_derivs_scaled = rbm.connections[0].weights.derivatives(vdata_scaled, hid_mean_scaled,
weighting_function=scale_func)
assert be.allclose(d_vis_loc, vis_derivs[0].loc), \
"derivative of visible loc wrong in gaussian-gaussian rbm"
assert be.allclose(d_hid_loc, hid_derivs[0].loc), \
"derivative of hidden loc wrong in gaussian-gaussian rbm"
assert be.allclose(d_vis_logvar, vis_derivs[0].log_var, rtol=1e-05, atol=1e-01), \
"derivative of visible log_var wrong in gaussian-gaussian rbm"
assert be.allclose(d_hid_logvar, hid_derivs[0].log_var, rtol=1e-05, atol=1e-01), \
"derivative of hidden log_var wrong in gaussian-gaussian rbm"
assert be.allclose(d_W, weight_derivs[0].matrix), \
"derivative of weights wrong in gaussian-gaussian rbm"
assert be.allclose(scale * d_vis_loc, vis_derivs_scaled[0].loc), \
"weighted derivative of visible loc wrong in gaussian-gaussian rbm"
assert be.allclose(scale * d_hid_loc, hid_derivs_scaled[0].loc), \
"weighted derivative of hidden loc wrong in gaussian-gaussian rbm"
assert be.allclose(scale * d_vis_logvar, vis_derivs_scaled[0].log_var, rtol=1e-05, atol=1e-01), \
"weighted derivative of visible log_var wrong in gaussian-gaussian rbm"
assert be.allclose(scale * d_hid_logvar, hid_derivs_scaled[0].log_var, rtol=1e-05, atol=1e-01), \
"weighted derivative of hidden log_var wrong in gaussian-gaussian rbm"
assert be.allclose(scale * d_W, weight_derivs_scaled[0].matrix), \
"weighted derivative of weights wrong in gaussian-gaussian rbm"
def test_rbm(paysage_path=None):
num_hidden_units = 50
batch_size = 50
num_epochs = 1
learning_rate = schedules.PowerLawDecay(initial=0.01, coefficient=0.1)
mc_steps = 1
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'examples', 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'examples', 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
import pandas
samples = pre.binarize_color(
be.float_tensor(
pandas.read_hdf(shuffled_filepath,
key='train/images').values[:10000]))
samples_train, samples_validate = batch.split_tensor(samples, 0.95)
data = batch.Batch({
'train':
batch.InMemoryTable(samples_train, batch_size),
'validate':
batch.InMemoryTable(samples_validate, batch_size)
})
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = ProgressMonitor()
untrained_performance = perf.epoch_update(data,
rbm,
store=True,
show=False)
# set up the optimizer and the fit method
opt = optimizers.RMSProp(stepsize=learning_rate)
cd = fit.SGD(rbm, data)
# fit the model
print('training with contrastive divergence')
cd.train(opt, num_epochs, method=fit.pcd, mcsteps=mc_steps)
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = cd.monitor.memory[-1]
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_independent():
"""
Test sampling from an rbm with two layers connected by a weight matrix that
contains all zeros, so that the layers are independent.
Note:
This test compares values estimated by *sampling* to values computed
analytically. It can fail for small batch_size, or strict tolerances,
even if everything is working propery.
"""
num_visible_units = 20
num_hidden_units = 10
batch_size = 1000
steps = 100
mean_tol = 0.2
corr_tol = 0.2
# set a seed for the random number generator
be.set_seed()
layer_types = [
layers.BernoulliLayer,
layers.GaussianLayer]
for layer_type in layer_types:
# set up some layer and model objects
vis_layer = layer_type(num_visible_units)
hid_layer = layer_type(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.rand((num_visible_units,))
b = be.rand((num_hidden_units,))
W = be.zeros((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.connections[0].weights.params.matrix[:] = W
if layer_type == layers.GaussianLayer:
log_var_a = be.randn((num_visible_units,))
log_var_b = be.randn((num_hidden_units,))
rbm.layers[0].params.log_var[:] = log_var_a
rbm.layers[1].params.log_var[:] = log_var_b
# initialize a state
state = State.from_model(batch_size, rbm)
# run a markov chain to update the state
state = rbm.markov_chain(steps, state)
# compute the mean
state_for_moments = State.from_model(1, rbm)
sample_mean = [be.mean(state[i], axis=0) for i in range(state.len)]
model_mean = [rbm.layers[i].conditional_mean(
rbm._connected_rescaled_units(i, state_for_moments),
rbm._connected_weights(i)) for i in range(rbm.num_layers)]
# check that the means are roughly equal
for i in range(rbm.num_layers):
ave = sample_mean[i]
close = be.allclose(ave, model_mean[i][0], rtol=mean_tol, atol=mean_tol)
assert close, "{0} {1}: sample mean does not match model mean".format(layer_type, i)
# check the cross correlation between the layers
crosscov = be.cov(state[0], state[1])
norm = be.outer(be.std(state[0], axis=0), be.std(state[1], axis=0))
crosscorr = be.divide(norm, crosscov)
assert be.tmax(be.tabs(crosscorr)) < corr_tol, "{} cross correlation too large".format(layer_type)
def test_tap_machine(paysage_path=None):
num_hidden_units = 10
batch_size = 50
num_epochs = 1
learning_rate = 0.01
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
data = batch.Batch(shuffled_filepath,
'train/images',
batch_size,
transform=batch.binarize_color,
train_fraction=0.1)
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = tap_machine.TAP_rbm([vis_layer, hid_layer],
tolerance_EMF=1e-2,
max_iters_EMF=50)
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = fit.ProgressMonitor(data, metrics=['ReconstructionError'])
untrained_performance = perf.check_progress(rbm)
# set up the optimizer and the fit method
opt = optimizers.Gradient(stepsize=learning_rate,
scheduler=optimizers.PowerLawDecay(0.1),
tolerance=1e-3,
ascent=True)
solver = fit.SGD(rbm, data, opt, num_epochs, method=fit.tap, monitor=perf)
# fit the model
print('training with stochastic gradient ascent')
solver.train()
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = perf.check_progress(rbm)
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_rbm(paysage_path=None):
num_hidden_units = 50
batch_size = 50
num_epochs = 1
learning_rate = schedules.PowerLawDecay(initial=0.01, coefficient=0.1)
mc_steps = 1
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
data = batch.HDFBatch(shuffled_filepath,
'train/images',
batch_size,
transform=pre.binarize_color,
train_fraction=0.99)
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = model.Model([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = fit.ProgressMonitor(data, metrics=['ReconstructionError'])
untrained_performance = perf.check_progress(rbm)
# set up the optimizer and the fit method
opt = optimizers.RMSProp(stepsize=learning_rate)
sampler = fit.DrivenSequentialMC.from_batch(rbm, data)
cd = fit.SGD(rbm,
data,
opt,
num_epochs,
sampler,
method=fit.pcd,
mcsteps=mc_steps,
monitor=perf)
# fit the model
print('training with contrastive divergence')
cd.train()
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = perf.check_progress(rbm)
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
def test_rbm(paysage_path=None):
"""TODO : this is just a placeholder, need to clean up & simplifiy setup. Also
need to figure how to deal with consistent random seeding throughout the
codebase to obtain deterministic checkable results.
"""
num_hidden_units = 50
batch_size = 50
num_epochs = 1
learning_rate = 0.01
mc_steps = 1
if not paysage_path:
paysage_path = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
filepath = os.path.join(paysage_path, 'mnist', 'mnist.h5')
if not os.path.exists(filepath):
raise IOError(
"{} does not exist. run mnist/download_mnist.py to fetch from the web"
.format(filepath))
shuffled_filepath = os.path.join(paysage_path, 'mnist',
'shuffled_mnist.h5')
# shuffle the data
if not os.path.exists(shuffled_filepath):
shuffler = batch.DataShuffler(filepath, shuffled_filepath, complevel=0)
shuffler.shuffle()
# set a seed for the random number generator
be.set_seed()
# set up the reader to get minibatches
data = batch.Batch(shuffled_filepath,
'train/images',
batch_size,
transform=batch.binarize_color,
train_fraction=0.99)
# set up the model and initialize the parameters
vis_layer = layers.BernoulliLayer(data.ncols)
hid_layer = layers.BernoulliLayer(num_hidden_units)
rbm = hidden.Model([vis_layer, hid_layer])
rbm.initialize(data)
# obtain initial estimate of the reconstruction error
perf = fit.ProgressMonitor(0, data, metrics=[M.ReconstructionError()])
untrained_performance = perf.check_progress(rbm, 0)
# set up the optimizer and the fit method
opt = optimizers.RMSProp(rbm,
stepsize=learning_rate,
scheduler=optimizers.PowerLawDecay(0.1))
sampler = fit.DrivenSequentialMC.from_batch(rbm, data, method='stochastic')
cd = fit.PCD(rbm,
data,
opt,
sampler,
num_epochs,
mcsteps=mc_steps,
skip=200,
metrics=[M.ReconstructionError()])
# fit the model
print('training with contrastive divergence')
cd.train()
# obtain an estimate of the reconstruction error after 1 epoch
trained_performance = perf.check_progress(rbm, 0)
assert (trained_performance['ReconstructionError'] <
untrained_performance['ReconstructionError']), \
"Reconstruction error did not decrease"
# close the HDF5 store
data.close()
from paysage import batch
from paysage import preprocess as pre
from paysage import layers
from paysage.models import model
from paysage import fit
from paysage import optimizers
from paysage import backends as be
from paysage import schedules
from paysage import penalties as pen
be.set_seed(137) # for determinism
import example_util as util
def run(paysage_path=None, num_epochs=10, show_plot=False):
num_hidden_units = 256
batch_size = 100
learning_rate = schedules.PowerLawDecay(initial=0.01, coefficient=0.1)
mc_steps = 1
(_, _, shuffled_filepath) = \
util.default_paths(paysage_path)
# set up the reader to get minibatches
data = batch.HDFBatch(shuffled_filepath,
'train/images',
batch_size,
transform=pre.binarize_color,
train_fraction=0.95)
def test_gaussian_conditional_params():
num_visible_units = 100
num_hidden_units = 50
batch_size = 25
# set a seed for the random number generator
be.set_seed()
# set up some layer and model objects
vis_layer = layers.GaussianLayer(num_visible_units)
hid_layer = layers.GaussianLayer(num_hidden_units)
rbm = BoltzmannMachine([vis_layer, hid_layer])
# randomly set the intrinsic model parameters
a = be.randn((num_visible_units,))
b = be.randn((num_hidden_units,))
log_var_a = 0.1 * be.randn((num_visible_units,))
log_var_b = 0.1 * be.randn((num_hidden_units,))
W = be.randn((num_visible_units, num_hidden_units))
rbm.layers[0].params.loc[:] = a
rbm.layers[1].params.loc[:] = b
rbm.layers[0].params.log_var[:] = log_var_a
rbm.layers[1].params.log_var[:] = log_var_b
rbm.connections[0].weights.params.matrix[:] = W
# generate a random batch of data
vdata = rbm.layers[0].random((batch_size, num_visible_units))
hdata = rbm.layers[1].random((batch_size, num_hidden_units))
# compute the variance
visible_var = be.exp(log_var_a)
hidden_var = be.exp(log_var_b)
# rescale the data
vdata_scaled = vdata / visible_var
hdata_scaled = hdata / hidden_var
# test rescale
assert be.allclose(vdata_scaled, rbm.layers[0].rescale(vdata)),\
"visible rescale wrong in gaussian-gaussian rbm"
assert be.allclose(hdata_scaled, rbm.layers[1].rescale(hdata)),\
"hidden rescale wrong in gaussian-gaussian rbm"
# compute the mean
hidden_mean = be.dot(vdata_scaled, W) # (batch_size, num_hidden_units)
hidden_mean += b
visible_mean = be.dot(hdata_scaled, be.transpose(W)) # (batch_size, num_hidden_units)
visible_mean += a
# update the conditional parameters using the layer functions
vis_mean_func, vis_var_func = rbm.layers[0].conditional_params(
[hdata_scaled], [rbm.connections[0].W(trans=True)])
hid_mean_func, hid_var_func = rbm.layers[1].conditional_params(
[vdata_scaled], [rbm.connections[0].W()])
assert be.allclose(visible_var, vis_var_func),\
"visible variance wrong in gaussian-gaussian rbm"
assert be.allclose(hidden_var, hid_var_func),\
"hidden variance wrong in gaussian-gaussian rbm"
assert be.allclose(visible_mean, vis_mean_func),\
"visible mean wrong in gaussian-gaussian rbm"
assert be.allclose(hidden_mean, hid_mean_func),\
"hidden mean wrong in gaussian-gaussian rbm"