def makeResponseLikelihood(self, *argv, **kwargs):
"""Make a prediction and predict its probability from a multivariate
normal distribution
Arguments:
* argv: an undetermined number of tensors containg the weights
and biases
* realVals: the actual values for the predicted quantities
* sd: standard deviation for output distribution, uses
current hyper parameter value if nothing is given
* hyperStates: A list containing all the hyper paramters
* predict: the function used to make a prediction from the
current neural net
* dtype: the datatype of the network
Returns:
* result: the log probabilities of the real vals given the
predicted values
"""
if (kwargs["sd"] is None):
sd = kwargs["hyperStates"][-1]
else:
sd = kwargs["sd"]
current = kwargs["predict"](True, argv[0])
current = tf.transpose(current)
sigma = tf.ones_like(current) * sd
realVals = tf.reshape(kwargs["realVals"], current.shape)
result = multivariateLogProb(sigma, current, realVals, kwargs["dtype"])
return (result)
def calculateHyperProbs(self, hypers, slopes):
"""Calculates the log probability of a set of weights and biases given
new distribtuions as well as the probability of the new distribution
means and SDs given their distribtuions.
Arguments:
* hypers: a list containg 4 new possible hyper parameters
* weightBias: a list with the current weight and bias matrices
Returns:
* prob: log probability of weights and biases given the new hypers
and the probability of the new hyper parameters given their priors
"""
mean = hypers[0]
sd = hypers[1]
slopes = tf.square(slopes[0])
prob = tf.reduce_sum(
multivariateLogProb(sd, mean, slopes, dtype=self.dtype))
# Calculate probability of new hypers
val = self.meanHyper.log_prob([mean])
prob += tf.reduce_sum(input_tensor=val)
# Calculate probability of weights and biases given new hypers
val = self.sdHyper.log_prob([sd])
prob += tf.reduce_sum(input_tensor=val)
return (prob)
def calcultateLogProb(self, *argv, **kwargs):
"""Make a prediction and predict its probability from a multivariate
normal distribution
rguments:
* argv: an undetermined number of tensors containg the weights
and biases
* realVals: the actual values for the predicted quantities
* hypers: A list containing all the hyper paramters
* predict: the function used to make a prediction from the
current neural net
* dtype: the datatype of the network
* n: Use every n networks
Returns:
* result: the log probabilities of the real vals given the
predicted values
"""
sd = []
for x in range(len(kwargs["hypers"])):
sd.append(kwargs["hypers"][x][-1])
current = kwargs["predict"](argv[0], n=kwargs["n"])
for x in range(len(current)):
current[x] = tf.transpose(current[x])
realVals = tf.reshape(kwargs["realVals"], current[0].shape)
result = []
for x in range(len(current)):
result.append(
multivariateLogProb(
tf.ones_like(current[0]) * sd[x], current[x], realVals,
kwargs["dtype"]))
return (result)
def calculateProbs(self, slopes):
"""Calculates the log probability of the slopes given
their distributions in this layer.
Arguments:
* weightsBias: list with new possible weight and bias tensors
Returns:
* prob: log prob of weights and biases given their distributions
"""
prob = tf.reduce_sum(
multivariateLogProb(self.hypers[1],
self.hypers[0],
slopes,
dtype=self.dtype))
return (prob)