def time_vspace_flatten():
val = {'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0],
'k5': np.array([4., 5., 6.]),
'k6': np.array([[7., 8.], [9., 10.]])}
vspace_flatten(val)
def test_flatten_dict():
val = {'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0]}
vect, unflatten = flatten(val)
val_recovered = unflatten(vect)
vect_2, _ = flatten(val_recovered)
assert np.all(vect == vect_2)
def test_flatten_dict():
val = {'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0]}
vect, unflatten = flatten(val)
val_recovered = unflatten(vect)
vect_2, _ = flatten(val_recovered)
assert np.all(vect == vect_2)
def time_vspace_flatten():
val = {
'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0],
'k5': np.array([4., 5., 6.]),
'k6': np.array([[7., 8.], [9., 10.]])
}
vspace_flatten(val)
def time_flatten():
val = {'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0],
'k5': np.array([4., 5., 6.]),
'k6': np.array([[7., 8.], [9., 10.]])}
vect, unflatten = flatten(val)
val_recovered = unflatten(vect)
vect_2, _ = flatten(val_recovered)
def time_flatten():
val = {
'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0],
'k5': np.array([4., 5., 6.]),
'k6': np.array([[7., 8.], [9., 10.]])
}
vect, unflatten = flatten(val)
val_recovered = unflatten(vect)
vect_2, _ = flatten(val_recovered)
def time_grad_flatten():
val = {'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0],
'k5': np.array([4., 5., 6.]),
'k6': np.array([[7., 8.], [9., 10.]])}
vect, unflatten = flatten(val)
def fun(vec):
v = unflatten(vec)
return np.sum(v['k5']) + np.sum(v['k6'])
grad(fun)(vect)
def time_grad_flatten():
val = {
'k': npr.random((4, 4)),
'k2': npr.random((3, 3)),
'k3': 3.0,
'k4': [1.0, 4.0, 7.0, 9.0],
'k5': np.array([4., 5., 6.]),
'k6': np.array([[7., 8.], [9., 10.]])
}
vect, unflatten = flatten(val)
def fun(vec):
v = unflatten(vec)
return np.sum(v['k5']) + np.sum(v['k6'])
grad(fun)(vect)
def test():
b = npr.random([8])
def f(b):
return fn(a, b)
deriv = autograd.grad(f)
print f(b)
print deriv(b)
def test_norm_pdf(): combo_check(stats.norm.pdf, [0,1,2], [R(4)], [R(4)], [R(4)**2 + 1.1]) def test_norm_cdf(): combo_check(stats.norm.cdf, [0,1,2], [R(4)], [R(4)], [R(4)**2 + 1.1]) def test_norm_logpdf(): combo_check(stats.norm.logpdf, [0,1,2], [R(4)], [R(4)], [R(4)**2 + 1.1]) def test_norm_logcdf(): combo_check(stats.norm.logcdf, [0,1,2], [R(4)], [R(4)], [R(4)**2 + 1.1]) def test_norm_pdf_broadcast(): combo_check(stats.norm.pdf, [0,1,2], [R(4,3)], [R(1,3)], [R(4,1)**2 + 1.1]) def test_norm_cdf_broadcast(): combo_check(stats.norm.cdf, [0,1,2], [R(4,3)], [R(1,3)], [R(4,1)**2 + 1.1]) def test_norm_logpdf_broadcast(): combo_check(stats.norm.logpdf, [0,1,2], [R(4,3)], [R(1,3)], [R(4,1)**2 + 1.1]) def test_norm_logcdf_broadcast(): combo_check(stats.norm.logcdf, [0,1,2], [R(4,3)], [R(1,3)], [R(4,1)**2 + 1.1]) def make_psd(mat): return np.dot(mat.T, mat) + np.eye(mat.shape[0]) def test_mvn_pdf(): combo_check(mvn.logpdf, [0, 1, 2], [R(4)], [R(4)], [make_psd(R(4, 4))]) def test_mvn_logpdf(): combo_check(mvn.logpdf, [0, 1, 2], [R(4)], [R(4)], [make_psd(R(4, 4))]) def test_mvn_entropy():combo_check(mvn.entropy,[0, 1], [R(4)], [make_psd(R(4, 4))]) alpha = npr.random(4)**2 + 1.2 x = stats.dirichlet.rvs(alpha, size=1)[0,:] # Need to normalize input so that x's sum to one even when we perturb them to compute numeric gradient. def normalize(x): return x / sum(x) def normalized_dirichlet_pdf( x, alpha): return stats.dirichlet.pdf( normalize(x), alpha) def normalized_dirichlet_logpdf(x, alpha): return stats.dirichlet.logpdf(normalize(x), alpha) def test_dirichlet_pdf_x(): combo_check(normalized_dirichlet_pdf, [0], [x], [alpha]) def test_dirichlet_pdf_alpha(): combo_check(stats.dirichlet.pdf, [1], [x], [alpha]) def test_dirichlet_logpdf_x(): combo_check(normalized_dirichlet_logpdf, [0], [x], [alpha]) def test_dirichlet_logpdf_alpha(): combo_check(stats.dirichlet.logpdf, [1], [x], [alpha]) ### Misc ### R = npr.randn def test_logsumexp1(): combo_check(autograd.scipy.misc.logsumexp, [0], [1.1, R(4), R(3,4)], axis=[None, 0], keepdims=[True, False])
combo_check(mvn.logpdf, [0, 1, 2], [R(4)], [R(4)], [make_psd(R(4, 4))])
def test_mvn_entropy():
combo_check(mvn.entropy, [0, 1], [R(4)], [make_psd(R(4, 4))])
def test_mvn_pdf_broadcast():
combo_check(mvn.logpdf, [0, 1, 2], [R(5, 4)], [R(4)], [make_psd(R(4, 4))])
def test_mvn_logpdf_broadcast():
combo_check(mvn.logpdf, [0, 1, 2], [R(5, 4)], [R(4)], [make_psd(R(4, 4))])
alpha = npr.random(4) ** 2 + 1.2
x = stats.dirichlet.rvs(alpha, size=1)[0, :]
# Need to normalize input so that x's sum to one even when we perturb them to compute numeric gradient.
def normalize(x):
return x / sum(x)
def normalized_dirichlet_pdf(x, alpha):
return stats.dirichlet.pdf(normalize(x), alpha)
def normalized_dirichlet_logpdf(x, alpha):
return stats.dirichlet.logpdf(normalize(x), alpha)
# -------------- LOADING DATASET ------------------------
# load the images
npr.seed(0)
_, train_images, train_labels, test_images, test_labels = load_mnist()
rand_idx = np.arange(train_images.shape[0])
npr.shuffle(rand_idx)
train_images = train_images[rand_idx]
train_labels = train_labels[rand_idx]
# UNIFORM CLASS SAMPLE CODE
# uniformly sample each class
cls_labels = train_labels.argmax(axis=1)
cls_images = [train_images[cls_labels == i] for i in range(10)]
rand_cls = np.int32(npr.random(30) / 0.1)
rand_idx = [npr.randint(cls_images[cls].shape[0]) for cls in rand_cls]
train_images = np.vstack(
[cls_images[rand_cls[i]][rand_idx[i]] for i in range(30)])
train_labels = np.vstack([
train_labels[cls_labels == rand_cls[i]][rand_idx[i]] for i in range(30)
])
# binarize
train_images = np.round(train_images)
test_images = np.round(test_images)
# -------------- LOADING DATASET ------------------------
print('LOADED DATASET')
import autograd
import autograd.numpy as np
import autograd.numpy.random as npr
a = npr.random([2, 2, 2, 2])
def fn(a, b):
bs = np.reshape(b, [2, 2, 2])
c = np.einsum("apqb,cqd->acpbd", a, bs)
return np.sum(c)
def test():
b = npr.random([8])
def f(b):
return fn(a, b)
deriv = autograd.grad(f)
print f(b)
print deriv(b)
