def __init__(self, x, y, mask, n_inputs, n_outputs, s_hiddens, s_act,
t_hiddens, t_act):
"""
Constructor of the backward computation graph.
:param x: theano array, the conditional input
:param y: theano array, the output
:param mask: theano array, a mask indicating which outputs are unchanged
:param n_inputs: int, number of conditional inputs
:param n_outputs: int, number of outputs
:param s_hiddens: list of hidden widths for the scale net
:param s_act: string, activation function for the scale net
:param t_hiddens: list of hidden widths for the translate net
:param t_act: string, activation function for the translate net
"""
# save input arguments
self.mask = mask
self.n_inputs = n_inputs
self.n_outputs = n_outputs
self.s_hiddens = s_hiddens
self.s_act = s_act
self.t_hiddens = t_hiddens
self.t_act = t_act
# masked output
my = mask * y
# scale function
self.s_net = nn.FeedforwardNet(n_inputs + n_outputs,
tt.concatenate([x, my], axis=1))
for h in s_hiddens:
self.s_net.addLayer(h, s_act)
self.s_net.addLayer(n_outputs, 'linear')
# translate function
self.t_net = nn.FeedforwardNet(n_inputs + n_outputs,
tt.concatenate([x, my], axis=1))
for h in t_hiddens:
self.t_net.addLayer(h, t_act)
self.t_net.addLayer(n_outputs, 'linear')
# output
s = self.s_net.output
t = self.t_net.output
self.u = my + (1.0 - mask) * tt.exp(-s) * (y - t)
# log det dx/dy
self.logdet_dudx = -tt.sum((1.0 - mask) * s, axis=1)
# parameters
self.parms = self.s_net.parms + self.t_net.parms
# theano evaluation function, will be compiled when first needed
self.eval_forward_f = None
def two_sample_test_classifier(x0, x1, rng=np.random):
"""
Classifier-based two sample test. Given two datasets, trains a binary classifier to discriminate between them, and
reports how well it does.
:param x0: first dataset
:param x1: second dataset
:param rng: random generator to use
:return: discrimination accuracy
"""
import ml.models.neural_nets as nn
import ml.trainers as trainers
import ml.loss_functions as lf
# create dataset
x0 = np.asarray(x0)
x1 = np.asarray(x1)
n_x0, n_dims = x0.shape
n_x1 = x1.shape[0]
n_data = n_x0 + n_x1
assert n_dims == x1.shape[1], 'inconsistent sizes'
xs = np.vstack([x0, x1])
ys = np.hstack([np.zeros(n_x0), np.ones(n_x1)])
# split in training / validation sets
n_val = int(n_data * 0.1)
xs_val, ys_val = xs[:n_val], ys[:n_val]
xs_trn, ys_trn = xs[n_val:], ys[n_val:]
# create classifier
classifier = nn.FeedforwardNet(n_dims)
classifier.addLayer(n_dims * 10, 'relu', rng=rng)
classifier.addLayer(n_dims * 10, 'relu', rng=rng)
classifier.addLayer(1, 'logistic', rng=rng)
# train classifier
trn_target, trn_loss = lf.CrossEntropy(classifier.output)
val_target, val_loss = lf.CrossEntropy(classifier.output)
trainer = trainers.SGD(
model=classifier,
trn_data=[xs_trn, ys_trn],
trn_loss=trn_loss,
trn_target=trn_target,
val_data=[xs_val, ys_val],
val_loss=val_loss,
val_target=val_target
)
trainer.train(
minibatch=100,
patience=20,
monitor_every=1,
logger=None
)
# measure accuracy
pred = classifier.eval(xs)[:, 0] > 0.5
acc = np.mean(pred == ys)
return acc
def eval_forward(self, x, u):
"""
Evaluates the layer forward, i.e. from input x and random numbers u to output y.
:param x: numpy array
:param u: numpy array
:return: numpy array
"""
if self.eval_forward_f is None:
# conditional input
tt_x = tt.matrix('x')
# masked random numbers
tt_u = tt.matrix('u')
mu = self.mask * tt_u
# scale net
s_net = nn.FeedforwardNet(self.n_inputs + self.n_outputs,
tt.concatenate([tt_x, mu], axis=1))
for h in self.s_hiddens:
s_net.addLayer(h, self.s_act)
s_net.addLayer(self.n_outputs, 'linear')
util.copy_model_parms(self.s_net, s_net)
s = s_net.output
# translate net
t_net = nn.FeedforwardNet(self.n_inputs + self.n_outputs,
tt.concatenate([tt_x, mu], axis=1))
for h in self.t_hiddens:
t_net.addLayer(h, self.t_act)
t_net.addLayer(self.n_outputs, 'linear')
util.copy_model_parms(self.t_net, t_net)
t = t_net.output
# transform (x,u) -> y
y = mu + (1.0 - self.mask) * (tt_u * tt.exp(s) + t)
# compile theano function
self.eval_forward_f = theano.function(inputs=[tt_x, tt_u],
outputs=y)
return self.eval_forward_f(x.astype(dtype), u.astype(dtype))
def create_net(rng=np.random):
"""
Creates a network with logistic output.
"""
n_inputs = sim.Prior().n_dims
net = nn.FeedforwardNet(n_inputs)
net.addLayer(100, 'relu', rng=rng)
net.addLayer(100, 'relu', rng=rng)
net.addLayer(1, 'logistic', rng=rng)
return net
def eval_forward(self, u):
"""
Evaluates the layer forward, i.e. from random numbers u to output x.
:param u: numpy array
:return: numpy array
"""
if self.eval_forward_f is None:
# masked random numbers
tt_u = tt.matrix('u')
mu = self.mask * tt_u
# scale net
s_net = nn.FeedforwardNet(self.n_inputs, mu)
for h in self.s_hiddens:
s_net.addLayer(h, self.s_act)
s_net.addLayer(self.n_inputs, 'linear')
util.copy_model_parms(self.s_net, s_net)
s = s_net.output
# translate net
t_net = nn.FeedforwardNet(self.n_inputs, mu)
for h in self.t_hiddens:
t_net.addLayer(h, self.t_act)
t_net.addLayer(self.n_inputs, 'linear')
util.copy_model_parms(self.t_net, t_net)
t = t_net.output
# transform u -> x
x = mu + (1.0 - self.mask) * (tt_u * tt.exp(s) + t)
# compile theano function
self.eval_forward_f = theano.function(
inputs=[tt_u],
outputs=x
)
return self.eval_forward_f(u.astype(dtype))
def __init__(self,
n_inputs,
n_hiddens,
act_fun,
n_outputs,
n_components,
rng=np.random):
"""
Constructs an mdn with a given architecture. Note that the mdn has full precision matrices.
:param n_inputs: dimensionality of the input
:param n_hiddens: list with number of hidden units in the net
:param act_fun: activation function type to use in the net
:param n_outputs: dimensionality of the output
:param n_components: number of mixture components
:return: None
"""
# check if inputs are of the right type
assert util.math.isposint(
n_inputs), 'Number of inputs must be a positive integer.'
assert util.math.isposint(
n_outputs), 'Number of outputs must be a positive integer.'
assert util.math.isposint(
n_components), 'Number of components must be a positive integer.'
assert isinstance(
n_hiddens, list
), 'Number of hidden units must be a list of positive integers.'
for h in n_hiddens:
assert util.math.isposint(
h
), 'Number of hidden units must be a list of positive integers.'
assert act_fun in ['logistic', 'tanh', 'linear', 'relu',
'softplus'], 'Unsupported activation function.'
# construct the net
self.net = nn.FeedforwardNet(n_inputs)
for h in n_hiddens:
self.net.addLayer(h, act_fun, rng)
self.input = self.net.input
# mixing coefficients
self.Wa = theano.shared(
(rng.randn(self.net.n_outputs, n_components) /
np.sqrt(self.net.n_outputs + 1)).astype(dtype),
name='Wa',
borrow=True)
self.ba = theano.shared(rng.randn(n_components).astype(dtype),
name='ba',
borrow=True)
self.a = tt.nnet.softmax(tt.dot(self.net.hs[-1], self.Wa) + self.ba)
# mixture means
# the mean of each component is calculated separately. consider vectorizing this
self.Wms = [
theano.shared((rng.randn(self.net.n_outputs, n_outputs) /
np.sqrt(self.net.n_outputs + 1)).astype(dtype),
name='Wm' + str(i),
borrow=True) for i in xrange(n_components)
]
self.bms = [
theano.shared(rng.randn(n_outputs).astype(dtype),
name='bm' + str(i),
borrow=True) for i in xrange(n_components)
]
self.ms = [
tt.dot(self.net.hs[-1], Wm) + bm
for Wm, bm in izip(self.Wms, self.bms)
]
# mixture precisions
# note that U here is an upper triangular matrix such that U'*U is the precision
self.WUs = [
theano.shared((rng.randn(self.net.n_outputs, n_outputs**2) /
np.sqrt(self.net.n_outputs + 1)).astype(dtype),
name='WU' + str(i),
borrow=True) for i in xrange(n_components)
]
self.bUs = [
theano.shared(rng.randn(n_outputs**2).astype(dtype),
name='bU' + str(i),
borrow=True) for i in xrange(n_components)
]
aUs = [
tt.reshape(
tt.dot(self.net.hs[-1], WU) + bU, [-1, n_outputs, n_outputs])
for WU, bU in izip(self.WUs, self.bUs)
]
triu_mask = np.triu(np.ones([n_outputs, n_outputs], dtype=dtype), 1)
diag_mask = np.eye(n_outputs, dtype=dtype)
self.Us = [
triu_mask * aU + diag_mask * tt.exp(diag_mask * aU) for aU in aUs
]
ldetUs = [tt.sum(tt.sum(diag_mask * aU, axis=2), axis=1) for aU in aUs]
# log probabilities
self.y = tt.matrix('y')
L_comps = [
-0.5 * tt.sum(tt.sum(
(self.y - m).dimshuffle([0, 'x', 1]) * U, axis=2)**2,
axis=1) + ldetU
for m, U, ldetU in izip(self.ms, self.Us, ldetUs)
]
self.L = tt.log(
tt.sum(tt.exp(tt.stack(L_comps, axis=1) + tt.log(self.a)),
axis=1)) - (0.5 * n_outputs * np.log(2 * np.pi))
self.trn_loss = -tt.mean(self.L)
# all parameters in one container
self.parms = self.net.parms + [
self.Wa, self.ba
] + self.Wms + self.bms + self.WUs + self.bUs
# save these for later
self.n_inputs = self.net.n_inputs
self.n_outputs = n_outputs
self.n_components = n_components
self.act_fun = act_fun
# theano evaluation functions, will be compiled when first needed
self.eval_comps_f = None
self.eval_lprobs_f = None
self.eval_grad_f = None
self.eval_score_f = None
