def test_alloc_diag(self):
rng = numpy.random.RandomState(utt.fetch_seed())
x = theano.tensor.vector()
g = alloc_diag(x)
f = theano.function([x], g)
# test "normal" scenario (5x5 matrix) and special cases of 0x0 and 1x1
for shp in [5, 0, 1]:
m = rng.rand(shp).astype(self.floatX)
v = numpy.diag(m)
r = f(m)
# The right matrix is created
assert (r == v).all()
# Test we accept only vectors
xx = theano.tensor.matrix()
ok = False
try:
alloc_diag(xx)
except TypeError:
ok = True
assert ok
# Test infer_shape
f = theano.function([x], g.shape)
topo = f.maker.fgraph.toposort()
if config.mode != 'FAST_COMPILE':
assert sum([node.op.__class__ == AllocDiag for node in topo]) == 0
for shp in [5, 0, 1]:
m = rng.rand(shp).astype(self.floatX)
assert (f(m) == m.shape).all()
def test_alloc_diag():
rng = numpy.random.RandomState(utt.fetch_seed())
x = theano.tensor.vector()
g = alloc_diag(x)
f = theano.function([x], g)
# test "normal" scenario (5x5 matrix) and special cases of 0x0 and 1x1
for shp in [5, 0, 1]:
m = rng.rand(shp).astype(config.floatX)
v = numpy.diag(m)
r = f(m)
# The right diagonal is extracted
assert (r == v).all()
# Test we accept only vectors
xx = theano.tensor.matrix()
ok = False
try:
alloc_diag(xx)
except TypeError:
ok = True
assert ok
# Test infer_shape
f = theano.function([x], g.shape)
topo = f.maker.fgraph.toposort()
if config.mode != 'FAST_COMPILE':
assert sum([node.op.__class__ == AllocDiag for node in topo]) == 0
for shp in [5, 0, 1]:
m = rng.rand(shp).astype(config.floatX)
assert (f(m) == m.shape).all()
def test_W_jump(self):
" tests that W is where I think it should be "
stats = self.stats
cov_hs = stats.d['second_hs']
assert cov_hs.dtype == config.floatX
#mean_hsv[i,j] = E_D,Q h_i s_i v_j
mean_hsv = stats.d['mean_hsv']
regularized = cov_hs + alloc_diag(
T.ones_like(self.model.mu) * self.model.W_eps)
assert regularized.dtype == config.floatX
inv = matrix_inverse(regularized)
assert inv.dtype == config.floatX
new_W = T.dot(mean_hsv.T, inv)
f = function([], new_W)
Wv = f()
aWv = self.model.W.get_value()
diffs = Wv - aWv
max_diff = np.abs(diffs).max()
if max_diff > self.tol:
raise Exception("W deviates from its correct value by at most " +
str(max_diff))
def test_W_jump(self):
" tests that W is where I think it should be "
stats = self.stats
cov_hs = stats.d['second_hs']
assert cov_hs.dtype == config.floatX
#mean_hsv[i,j] = E_D,Q h_i s_i v_j
mean_hsv = stats.d['mean_hsv']
regularized = cov_hs + alloc_diag(T.ones_like(self.model.mu) * self.model.W_eps)
assert regularized.dtype == config.floatX
inv = matrix_inverse(regularized)
assert inv.dtype == config.floatX
new_W = T.dot(mean_hsv.T, inv)
f = function([], new_W)
Wv = f()
aWv = self.model.W.get_value()
diffs = Wv - aWv
max_diff = np.abs(diffs).max()
if max_diff > self.tol:
raise Exception("W deviates from its correct value by at most "+str(max_diff))
def test_grad_W(self):
"""tests that the gradient of the log probability with respect to W
matches my analytical derivation """
#self.model.set_param_values(self.new_params)
g = T.grad(self.prob,
self.model.W,
consider_constant=self.mf_obs.values())
B = self.model.B
W = self.model.W
mean_hsv = self.stats.d['mean_hsv']
mean_sq_hs = self.stats.d['mean_sq_hs']
mean_HS = self.mf_obs['H_hat'] * self.mf_obs['S_hat']
m = mean_HS.shape[0]
outer_prod = T.dot(mean_HS.T, mean_HS)
outer_prod.name = 'outer_prod<from_observations>'
outer = outer_prod / m
mask = T.identity_like(outer)
second_hs = (1. - mask) * outer + alloc_diag(mean_sq_hs)
term1 = (B * mean_hsv).T
term2 = -B.dimshuffle(0, 'x') * T.dot(W, second_hs)
analytical = term1 + term2
f = function([], (g, analytical))
gv, av = f()
assert gv.shape == av.shape
max_diff = np.abs(gv - av).max()
if max_diff > self.tol:
print "gv"
print gv
print "av"
print av
raise Exception(
"analytical gradient on W deviates from theano gradient on W by up to "
+ str(max_diff))
def test_grad_W(self):
"""tests that the gradient of the log probability with respect to W
matches my analytical derivation """
#self.model.set_param_values(self.new_params)
g = T.grad(self.prob, self.model.W, consider_constant = self.mf_obs.values())
B = self.model.B
W = self.model.W
mean_hsv = self.stats.d['mean_hsv']
mean_sq_hs = self.stats.d['mean_sq_hs']
mean_HS = self.mf_obs['H_hat'] * self.mf_obs['S_hat']
m = mean_HS.shape[0]
outer_prod = T.dot(mean_HS.T,mean_HS)
outer_prod.name = 'outer_prod<from_observations>'
outer = outer_prod/m
mask = T.identity_like(outer)
second_hs = (1.-mask) * outer + alloc_diag(mean_sq_hs)
term1 = (B * mean_hsv).T
term2 = - B.dimshuffle(0,'x') * T.dot(W, second_hs)
analytical = term1 + term2
f = function([],(g,analytical))
gv, av = f()
assert gv.shape == av.shape
max_diff = np.abs(gv-av).max()
if max_diff > self.tol:
print "gv"
print gv
print "av"
print av
raise Exception("analytical gradient on W deviates from theano gradient on W by up to "+str(max_diff))
