def test_basic_log_prob(self):
n_samples = self.n_samples
layer = self.layer
X, X_ = testing.fmatrix((n_samples, layer.n_X), name="X")
Y, Y_ = testing.fmatrix((n_samples, layer.n_Y), name="H")
log_prob = layer.log_prob(X, Y)
do_log_prob = theano.function([X, Y], log_prob, name="log_prob")
log_prob_ = do_log_prob(X_, Y_)
assert log_prob_.shape == (n_samples, )
assert not np.isnan(log_prob_).any()
def test_basic_log_prob(self):
n_samples = self.n_samples
layer = self.layer
X, X_ = testing.fmatrix((n_samples, layer.n_X), name="X")
Y, Y_ = testing.fmatrix((n_samples, layer.n_Y), name="H")
log_prob = layer.log_prob(X, Y)
do_log_prob = theano.function([X, Y], log_prob, name="log_prob")
log_prob_ = do_log_prob(X_, Y_)
assert log_prob_.shape == (n_samples,)
assert not np.isnan(log_prob_).any()
def test_log_likelihood(self):
batch_size = 20
stack = self.stack
X, X_ = testing.fmatrix((batch_size, self.n_vis), 'X')
n_samples, n_samples_ = testing.iscalar('n_samples')
n_samples_ = self.n_samples
log_PX, w, log_P, log_Q, KL, Hp, Hq = stack.log_likelihood(
X, n_samples=n_samples)
do_log_likelihood = theano.function([X, n_samples],
[log_PX, log_P, log_Q, w])
log_PX_, log_P_, log_Q_, w_ = do_log_likelihood(X_, n_samples_)
print "log_P.shape", log_P_.shape
print "log_Q.shape", log_Q_.shape
print "log_PX.shape", log_PX_.shape
print "w.shape", w_.shape
assert log_PX_.shape == (batch_size, )
assert log_P_.shape == (batch_size, n_samples_)
assert log_Q_.shape == (batch_size, n_samples_)
assert w_.shape == (batch_size, n_samples_)
n_layers = len(stack.p_layers)
assert len(KL) == n_layers
assert len(Hp) == n_layers
assert len(Hq) == n_layers
def test_log_likelihood(self):
batch_size = 20
stack = self.stack
X, X_ = testing.fmatrix((batch_size, self.n_vis), "X")
n_samples, n_samples_ = testing.iscalar("n_samples")
n_samples_ = self.n_samples
log_PX, w, log_P, log_Q, KL, Hp, Hq = stack.log_likelihood(X, n_samples=n_samples)
do_log_likelihood = theano.function([X, n_samples], [log_PX, log_P, log_Q, w])
log_PX_, log_P_, log_Q_, w_ = do_log_likelihood(X_, n_samples_)
print "log_P.shape", log_P_.shape
print "log_Q.shape", log_Q_.shape
print "log_PX.shape", log_PX_.shape
print "w.shape", w_.shape
assert log_PX_.shape == (batch_size,)
assert log_P_.shape == (batch_size, n_samples_)
assert log_Q_.shape == (batch_size, n_samples_)
assert w_.shape == (batch_size, n_samples_)
n_layers = len(stack.p_layers)
assert len(KL) == n_layers
assert len(Hp) == n_layers
assert len(Hq) == n_layers
def test_basic_sample(self):
n_samples = self.n_samples
layer = self.layer
Y, Y_ = testing.fmatrix((n_samples, layer.n_Y), name="Y")
X, log_prob = layer.sample(Y)
do_sample_p = theano.function([Y], [X, log_prob], name="sample")
X_, log_prob_ = do_sample_p(Y_)
assert X_.shape == (n_samples, layer.n_X)
assert log_prob_.shape == (n_samples, )
assert not np.isnan(log_prob_).any()
def test_basic_sample(self):
n_samples = self.n_samples
layer = self.layer
Y, Y_ = testing.fmatrix((n_samples, layer.n_Y), name="Y")
X, log_prob = layer.sample(Y)
do_sample_p = theano.function([Y], [X, log_prob], name="sample")
X_, log_prob_ = do_sample_p(Y_)
assert X_.shape == (n_samples, layer.n_X)
assert log_prob_.shape == (n_samples,)
assert not np.isnan(log_prob_).any()
def test_gradients(self):
batch_size = 20
stack = self.stack
n_layers = len(stack.p_layers)
X, X_ = testing.fmatrix((batch_size, self.n_vis), "X")
n_samples, n_samples_ = testing.iscalar("n_samples")
n_samples_ = self.n_samples
lr_p = np.ones(n_layers)
lr_q = np.ones(n_layers)
log_PX, gradients = stack.get_gradients(X, None, lr_p=lr_p, lr_q=lr_q, n_samples=n_samples_)
def test_sample_expected(self):
n_samples = self.n_samples
layer = self.layer
if getattr(layer, "sample_expected", None) is None:
raise unittest.SkipTest("sample_expected not implemented")
Y, Y_ = testing.fmatrix((n_samples, layer.n_Y), name="Y")
X, log_prob = layer.sample(Y)
do_sample_p = theano.function([Y], [X, log_prob], name="sample")
X_, log_prob_ = do_sample_p(Y_)
assert X_.shape == (n_samples, layer.n_X)
assert log_prob_.shape == (n_samples, )
assert not np.isnan(log_prob_).any()
def test_sample_expected(self):
n_samples = self.n_samples
layer = self.layer
if getattr(layer, "sample_expected", None) is None:
raise unittest.SkipTest("sample_expected not implemented")
Y, Y_ = testing.fmatrix((n_samples, layer.n_Y), name="Y")
X, log_prob = layer.sample(Y)
do_sample_p = theano.function([Y], [X, log_prob], name="sample")
X_, log_prob_ = do_sample_p(Y_)
assert X_.shape == (n_samples, layer.n_X)
assert log_prob_.shape == (n_samples,)
assert not np.isnan(log_prob_).any()
def test_gradients(self):
batch_size = 20
stack = self.stack
n_layers = len(stack.p_layers)
X, X_ = testing.fmatrix((batch_size, self.n_vis), 'X')
n_samples, n_samples_ = testing.iscalar('n_samples')
n_samples_ = self.n_samples
lr_p = np.ones(n_layers)
lr_q = np.ones(n_layers)
log_PX, gradients = stack.get_gradients(X,
None,
lr_p=lr_p,
lr_q=lr_q,
n_samples=n_samples_)
