def __init__(self, model, data=None):
""" The constructor initializes the CD trainer with a given model and data.
:param model: The model to sample from.
:type model: Valid model class.
:param data: Data for initialization, only has effect if the centered gradient is used.
:type data: numpy array [num. samples x input dim]
"""
# Store model variable
self.model = model
# Create the Gibbs-sampler
self.sampler = sampler.GibbsSampler(model)
# Count the number of parameters
parameters = self.model.get_parameters()
self.num_parameters = len(parameters)
self.hidden_offsets = 0.5 * numx.ones((1, self.model.output_dim))
if data is not None:
if self.model.input_dim != data.shape[1]:
raise ex.ValueError("Data dimension and model input dimension have to be equal!")
self.visible_offsets = data.mean(axis=0).reshape(1, data.shape[1])
else:
self.visible_offsets = 0.5 * numx.ones((1, self.model.input_dim))
# Storage variables for the gradients
self.parameter_updates = []
for i in range(self.num_parameters):
self.parameter_updates.append(numx.zeros((
parameters[i].shape[0],
parameters[i].shape[1]),
dtype=model.dtype))
def test_Gibbs_sampler(self):
sys.stdout.write('RBM Sampler -> Performing GibbsSampler test ... ')
sys.stdout.flush()
numx.random.seed(42)
sampler = Sampler.GibbsSampler(self.bbrbm)
probCD1, probCD2, probCS1, probCS2, probCS3, probCS4, sumProbs = self.execute_sampler(sampler, self.num_samples)
assert numx.all(numx.abs(1.0 / 4.0 - probCD1) < self.epsilon)
assert numx.all(numx.abs(1.0 / 4.0 - probCD2) < self.epsilon)
assert numx.all(numx.abs(1.0 / 8.0 - probCS1) < self.epsilon)
assert numx.all(numx.abs(1.0 / 8.0 - probCS2) < self.epsilon)
assert numx.all(numx.abs(1.0 / 8.0 - probCS3) < self.epsilon)
assert numx.all(numx.abs(1.0 / 8.0 - probCS4) < self.epsilon)
assert numx.all(numx.abs(1.0 - sumProbs) < self.epsilon)
print('successfully passed!')
sys.stdout.flush()
def __init__(self, model, batch_size):
# Set batch size
self.batch_size = batch_size
# Store model
self.model = model
self.rbm = RBM_MODEL.BinaryBinaryRBM(
number_visibles=model.input_dim + model.hidden2_dim,
number_hiddens=model.hidden1_dim,
data=None,
initial_weights=numx.vstack((model.W1, model.W2.T)),
initial_visible_bias=numx.hstack((model.b1, model.b3)),
initial_hidden_bias=model.b2,
initial_visible_offsets=numx.hstack((model.o1, model.o3)),
initial_hidden_offsets=model.o2)
self.sampler = RBM_SAMPLER.GibbsSampler(self.rbm)