def test_exp_matrix_colwise():
inpt = T.matrix()
norm = exp(inpt, axis=0)
f = theano.function([inpt], norm, mode='FAST_COMPILE')
res = f(test_matrix)
correct = roughly(res, [2.85361711, 24.90040964, 97.4061874, 1.])
assert correct, 'exp norm colwise not working'
def test_exp_matrix_rowwise():
inpt = T.matrix()
norm = exp(inpt, axis=1)
f = theano.function([inpt], norm, mode='FAST_COMPILE')
res = f(test_matrix)
correct = roughly(res, [114.68499678, 11.47521737])
assert correct, 'exp norm rowwise not working'
def test_l1_matrix_colwise():
inpt = T.matrix()
norm = l1(inpt, axis=0)
f = theano.function([inpt], norm, mode='FAST_COMPILE')
res = f(test_matrix)
correct = roughly(res, [3., 4.2, 6.5, 100.2])
assert correct, 'l1 norm colwise not working'
def test_squared():
X, Y = T.matrix(), T.matrix()
X.tag.test_value = test_X
Y.tag.test_value = test_Y
dist = squared(X, Y).sum()
f = theano.function([X, Y], dist, mode='FAST_COMPILE')
res = f(test_X, test_Y)
assert roughly(res, 0.0097), 'squared loss not working'
def test_model_function():
pars = ParameterSet(weights=(2, 3))
inpt = T.matrix()
output = T.dot(inpt, pars.weights)
model = Model()
model.exprs = {'inpt': inpt, 'output': output}
model.parameters = pars
f = model.function(['inpt'], 'output')
fx = model.function(['inpt'], 'output', explicit_pars=True)
np.random.seed(1010)
test_inpt = np.random.random((10, 2))
correct = roughly(f(test_inpt), fx(pars.data, test_inpt))
assert correct, 'implicit pars and explicit pars have different output'
f1 = model.function(['inpt'], ['output'])
f2 = model.function([inpt], ['output'])
f3 = model.function([inpt], [output])
f4 = model.function(['inpt'], [output])
assert roughly(f1(test_inpt), f2(test_inpt)), "f1 and f2 don't agree"
assert roughly(f1(test_inpt), f3(test_inpt)), "f1 and f3 don't agree"
assert roughly(f1(test_inpt), f4(test_inpt)), "f1 and f4 don't agree"
assert roughly(f2(test_inpt), f3(test_inpt)), "f2 and f3 don't agree"
assert roughly(f2(test_inpt), f4(test_inpt)), "f2 and f4 don't agree"
assert roughly(f3(test_inpt), f4(test_inpt)), "f3 and f4 don't agree"
def test_neg_cross_entropy_rowwise():
X, Y = T.matrix(), T.matrix()
X.tag.test_value = test_X > 0.2
Y.tag.test_value = test_Y
dist = nce(X, Y).sum(axis=1)
f = theano.function([X, Y], dist, mode='FAST_COMPILE')
res = f(test_X, test_Y)
correct = roughly(res, [0.85798192, 0.74838975])
assert correct, 'nce loss rowwise not working'
def test_neg_cross_entropy_colwise():
X, Y = T.matrix(), T.matrix()
X.tag.test_value = test_X > 0.2
Y.tag.test_value = test_Y
dist = nce(X, Y).sum(axis=0)
f = theano.function([X, Y], dist, mode='FAST_COMPILE')
res = f(test_X, test_Y)
correct = roughly(res, [[0.49795728, 0.48932523, 0.61908916]])
assert correct, 'nce loss colwise not working'
def test_l2_matrix_colwise():
inpt = T.matrix()
norm = l2(inpt, axis=0)
f = theano.function([inpt], norm, mode='FAST_COMPILE')
res = f(test_matrix)
correct = roughly(
res,
[5., 1.12400000e+01, 2.42500000e+01, 1.00400400e+04])
assert correct, 'l2 norm colwise not working'
def test_distance_matrix():
X = T.matrix()
D = distance_matrix(X)
f = theano.function([X], D, mode='FAST_COMPILE')
x = np.array([[1], [2], [3]])
res = f(x)
print res
correct = roughly(res, np.array([[0, 1, 4], [1, 0, 1], [4, 1, 0]]))
assert correct, 'distance matrix not working right'
def test_neg_cross_entropy():
X, Y = T.matrix(), T.matrix()
X.tag.test_value = test_X > 0.2
Y.tag.test_value = test_Y
dist = nce(X, Y).sum()
f = theano.function([X, Y], dist, mode='FAST_COMPILE')
res = f(test_X, test_Y)
correct = roughly(res, 1.6063716678910529)
assert correct, 'nce loss not working'
def test_squared_colwise():
X, Y = T.matrix(), T.matrix()
X.tag.test_value = test_X
Y.tag.test_value = test_Y
dist = squared(X, Y).sum(axis=0)
f = theano.function([X, Y], dist, mode='FAST_COMPILE')
res = f(test_X, test_Y)
correct = roughly(res, [0.0018, 0.005, 0.0029])
assert correct, 'squared loss colwise not working'
def test_absolute_colwise():
X, Y = T.matrix(), T.matrix()
X.tag.test_value = test_X
Y.tag.test_value = test_Y
dist = absolute(X, Y).sum(axis=0)
f = theano.function([X, Y], dist, mode='FAST_COMPILE')
res = f(test_X, test_Y)
correct = roughly(res, [0.06, 0.08, 0.07])
assert correct, 'absolute loss colwise not working'
def test_tanh():
inpt = T.matrix()
expr = tanh(inpt)
f = theano.function([inpt], expr)
result = f(test_matrix)
desired = np.array([
[-0.96402758, 0.9966824, 0.99975321, -1.],
[0.76159416, -0.76159416, 0.96402758, 0.]])
correct = roughly(result, desired)
assert correct, 'tanh not working'
def test_tanhplus():
inpt = T.matrix()
expr = tanhplus(inpt)
f = theano.function([inpt], expr)
result = f(test_matrix)
desired = np.array([
[-2.96402758, 4.1966824 , 5.49975321, -101.2],
[ 1.76159416,-1.76159416, 2.96402758, 0.]])
correct = roughly(result, desired)
assert correct, 'tanh plus not working'
def test_sigmoid():
inpt = T.matrix()
expr = sigmoid(inpt)
f = theano.function([inpt], expr)
result = f(test_matrix)
desired = np.array([
[1.19202922e-01, 9.60834277e-01, 9.89013057e-01, 3.04574061e-44],
[7.31058579e-01, 2.68941421e-01, 8.80797078e-01, 5.00000000e-01]])
correct = roughly(result, desired)
assert correct, 'sigmoid not working'
def test_rectifier():
inpt = T.matrix()
expr = rectifier(inpt)
f = theano.function([inpt], expr)
result = f(test_matrix)
desired = np.array([
[0., 3.2, 4.5, 0.],
[1., 0., 2., 0.]])
correct = roughly(result, desired)
assert correct, 'relu not working'
def test_soft():
inpt = T.matrix()
expr = softplus(inpt)
f = theano.function([inpt], expr)
result = f(test_matrix)
desired = np.array([
[0.12692801, 3.23995333, 4.51104774, 0.],
[1.31326169, 0.31326169, 2.12692801, 0.69314718]])
correct = roughly(result, desired)
assert correct, 'soft relu not working'
def test_log_product_of_t():
inpt = T.matrix()
expr = logproduct_of_t(inpt)
f = theano.function([inpt], expr)
result = f(test_matrix)
desired = np.array([
[1.60943791, 2.41947884, 3.0563569 , 9.21443597],
[0.69314718, 0.69314718, 1.60943791, 0.]])
correct = roughly(result, desired)
assert correct, 'pot not working'
def test_nominal_neg_cross_entropy():
Xc, Y = T.ivector(), T.matrix()
Xc.tag.test_value = (test_X > 0.2).argmax(axis=1).astype('uint8')
Y.tag.test_value = test_Y
dist = nnce(Xc, Y).sum()
f = theano.function([Xc, Y], dist, mode='FAST_COMPILE')
res = f(test_Xc, test_Y)
desired = -np.log(test_Y)[np.arange(test_Xc.shape[0]), test_Xc].sum()
correct = roughly(res, desired)
print res
print desired
assert correct, 'nnce loss not working'
def test_theano_function_with_nested_exprs():
def expr_generator(a, b):
ra = [T.pow(a[i], i) for i in range(len(a))]
return ra, T.exp(b)
a = [T.scalar('a%d' % i) for i in range(5)]
b = T.scalar('b')
f = breze.arch.util.theano_function_with_nested_exprs(
[a, b], expr_generator(a, b))
va = [2 for _ in a]
vb = 3
resa, resb = f(va, vb)
print "va: ", va
print "vb: ", vb
print "resa: ", resa
print "resb: ", resb
for i in range(len(va)):
assert roughly(resa[i], math.pow(va[i], i))
assert roughly(resb, math.exp(vb))
def test_lds_values():
# The following test values were generated with David Barber's LDS
# implementation coming along with his book 'Bayesian Reasoning and
# machine learning'.
l = LinearDynamicalSystem(2, 3)
l.parameters['transition'] = np.arange(1, 10).reshape((3, 3)).T
l.parameters['emission'] = np.arange(1, 7).reshape((3, 2))
l.parameters['visible_noise_mean'] = np.array((-1, 1))
cov_block = np.arange(1, 5).reshape((2, 2))
l.parameters['visible_noise_cov'] = (
np.dot(cov_block, cov_block.T) * 0.2 + np.eye(2) * 100)
l.parameters['hidden_noise_mean'] = np.array((1, 2, 3))
cov_block = np.arange(1, 10).reshape((3, 3))
l.parameters['hidden_noise_cov'] = (
np.dot(cov_block, cov_block.T) + np.eye(3) * 10)
l.parameters['hidden_mean_initial'] = -np.array((.1, .2, .3))
l.parameters['hidden_cov_initial'] = (
np.dot(cov_block, cov_block.T) * 0.2 + np.eye(3) * 10)
f_forward = l.function(
['inpt'],
['filtered_means', 'filtered_covs', 'log_likelihood'],
mode='FAST_COMPILE')
f_backward = l.function(
['filtered_means', 'filtered_covs'],
['smoothed_means', 'smoothed_covs'],
mode='FAST_COMPILE')
X = np.array([
[1, 1.5],
[2, 2.5],
[3, 3.5],
[4, 4.5]])
X.shape = 4, 1, 2
f, F, ll = f_forward(X)
f_desired = np.array([
[[-0.0395, 0.0649, 0.1694]],
[[0.5087, 0.2911, 0.0734]],
[[0.5593, 0.3754, 0.1914]],
[[0.8622, 0.5272, 0.1921]]])
F_desired = np.zeros((4, 1, 3, 3))
F_desired[0, 0] = [
[9.4992, -1.0806, -1.6604],
[-1.0806, 7.5570, -3.8054],
[-1.6604, -3.8054, 4.0497],
]
F_desired[1, 0] = [
[16.0466, 0.6834, -4.6797],
[0.6834, 8.0361, -4.6112],
[-4.6797, -4.6112, 5.4574],
]
F_desired[2, 0] = [
[18.8713, 1.4414, -5.9885],
[1.4414, 8.2395, -4.9624],
[-5.9885, -4.9624, 6.0637],
]
F_desired[3, 0] = [
[19.7744, 1.6837, -6.4070],
[1.6837, 8.3045, -5.0747],
[-6.4070, -5.0747, 6.2577],
]
assert roughly(f, f_desired, 1E-4), 'filtered means not correct'
assert roughly(F, F_desired, 1E-4), 'filtered covs not correct'
assert roughly(ll, [-38.3636], 1E-4), 'log likelihood calculated wrong: %f' % ll
s, S = f_backward(f, F)
s_desired = np.array([
[-0.2173, 0.1706, -0.5076, 0.8622],
[-0.0597, 0.0528, -0.0275, 0.5272],
[0.0978, -0.0651, 0.4526, 0.1921]]).T
S_desired = np.zeros((4, 1, 3, 3))
S_desired[0, 0] = np.array([[5.5498, -2.5166, -0.583],
[-2.5166, 6.9683, -3.5468],
[-0.583, -3.5468, 3.4894]])
S_desired[1, 0] = np.array([[7.3348, -2.1811, -1.697],
[-2.1811, 7.0324, -3.7542],
[-1.697, -3.7542, 4.1886]])
S_desired[2, 0] = np.array([[9.8391, -1.7404, -3.3199],
[-1.7404, 7.11, -4.0396],
[-3.3199, -4.0396, 5.2407]])
S_desired[3, 0] = np.array([[19.7744, 1.6837, -6.407],
[1.6837, 8.3045, -5.0747],
[-6.407, -5.0747, 6.2577]])
assert roughly(s[:, 0, :], s_desired, 1E-4), 'smooooothed means not correct'
assert roughly(S, S_desired, 1E-4), 'smooooothed covs not correct'