def build_net(x_train,
y_train,
num_train_steps=10000,
x_test=None,
y_test=None):
# Number of stats currently used to predict outcome- 23 per team + variable for side
inputs = 47
outputs = 1
if x_test is None:
x_test = x_train
if y_test is None:
y_test = y_train
# widths of fully-connected layers in NN
# Input data goes here (via feed_dict or equiv)
x = tf.placeholder(tf.float32, shape=[None, inputs])
layer_widths = [16, 16, 16, 16, 16, 16]
activations = [Nets.selu for _ in layer_widths] + [tf.identity]
layer_widths += [outputs]
net = Nets.SuperDenseNet(inputs, layer_widths, activations)
# Construct all parameters of NN, set to independant gaussian priors
params = [Nets.gauss_prior(shape) for shape in net.param_space()]
out = ed.models.Bernoulli(logits=net.apply(x, params))
# Variational 'posterior's for NN params
qparams = [Nets.gauss_var_post(w.shape) for w in params]
asd = tf.train.AdamOptimizer
# Map from random variables to their variational posterior objects
params_post = {params[i]: qparams[i] for i in range(len(params))}
# evaluate accuracy and likelihood of model over the dataset before training
print(
'accuracy, log_likelihood, crossentropy',
ed.evaluate(['accuracy', 'log_likelihood', 'crossentropy'],
data={
out: y_test,
x: x_test
}))
# Run variational inference, minimizing KL(q, p) using stochastic gradient descent over variational params
inference = ed.KLqp(params_post, data={out: y_train, x: x_train})
#inference.initialize(optimizer=YFOptimizer())
inference.run(n_samples=32, n_iter=num_train_steps)
# Get output object dependant on variational posteriors rather than priors
out_post = ed.copy(out, params_post)
# Re-evaluate metrics
print(
'accuracy, log_likelihood, crossentropy',
ed.evaluate(['accuracy', 'log_likelihood', 'crossentropy'],
data={
out_post: y_test,
x: x_test
}))
def __init__(self, teams, predictor, features=6, prior=None):
self.team_numbers = {}
self.team_names = []
self.features = features
self.predictor = predictor
self.param_space = list(predictor.param_space()) + [[teams, features]]
self.var_post = [
Nets.gauss_var_post(shape) for shape in self.param_space
]
if prior is None:
prior = [Nets.gauss_prior(shape) for shape in self.param_space]
self.prior = prior