def __init__(self,
D_in,
D_out,
T,
timestep,
kernel,
M=10,
mean_f=zero_mean,
mean_g=zero_mean,
sigma_f_prior=dist.Uniform(1., 2.),
alpha_f_prior=dist.Uniform(0.25, 0.75),
sigma_g_prior=dist.Uniform(1., 2.),
alpha_g_prior=dist.Uniform(0.25, 0.75)):
super(FlowGP, self).__init__(D_in, [], kernel, M)
self.T = T
self.M = M
self.dt = torch.tensor(timestep).float()
self.D_in = D_in
self.D_out = D_out
self.f = SGPR(D_in=D_in,
M=M,
kernel=kernel(D_in,
sigma_prior=sigma_f_prior,
alpha_prior=alpha_f_prior),
D_out=D_in,
mean=mean_f)
self.g = SGPR(D_in=D_in,
M=M,
kernel=kernel(D_out,
sigma_prior=sigma_g_prior,
alpha_prior=alpha_g_prior),
D_out=D_out,
mean=mean_g)
self.checkpoints = []
def mh_update(samples, sample_labels, width_proposal, prior, l_k, inputs, labels,
width_inc=1.02, width_dec=0.5, update_steps=250, sess=tf.Session()):
#Take reference to the original code from the paper, which provides a efficient method for MH process
import torch
import torch.distributions as dist
sample_size = samples.shape[0]
x = torch.tensor(samples).view(-1, 28*28)
acc_ratio = torch.zeros(sample_size)
for i in range(update_steps):
g_bottom = dist.Uniform(low=torch.max(x - width_proposal.unsqueeze(-1), prior.low), high=torch.min(x + width_proposal.unsqueeze(-1), prior.high))
x_new = g_bottom.sample()
loss_run_time, logits_run_time = sess.run([loss, logits], feed_dict={inputs: x_new.view(-1, 28, 28).numpy(),
labels: sample_labels})
s_xn = compute_property_function(logits_run_time, sample_labels)
g_top = dist.Uniform(low=torch.max(x_new - width_proposal.unsqueeze(-1), prior.low), high=torch.min(x_new + width_proposal.unsqueeze(-1), prior.high))
lg_alpha = (prior.log_prob(x_new) - prior.log_prob(x)+ g_top.log_prob(x) - g_bottom.log_prob(x_new)).sum(dim=1)
acceptance = torch.min(lg_alpha, torch.zeros_like(lg_alpha))
log_u = torch.log(torch.rand_like(acceptance))
acc_idx = torch.tensor((log_u <= acceptance).numpy() & (s_xn >= l_k))
acc_ratio += acc_idx.float()
x = torch.where(acc_idx.unsqueeze(-1), x_new, x)
width_proposal = torch.where(acc_ratio > 0.124, width_proposal*width_inc, width_proposal)
width_proposal = torch.where(acc_ratio < 0.124, width_proposal*width_dec, width_proposal)
return x.view(-1, 28, 28).numpy(), width_proposal
def __init__(self,
D_out,
sigma_prior=dist.Uniform(1., 2.),
alpha_prior=dist.Uniform(1., 2.)):
super(SquaredExp, self).__init__()
self.sigma = nn.Parameter(sigma_prior.sample(sample_shape=(1, 1)))
self.alpha = nn.Parameter(alpha_prior.sample(sample_shape=(1, 1)))
def get_random_rectangles_params(params_shape, images_height, images_width,
erase_scale_range, aspect_ratio_range):
images_area = images_height * images_width
target_areas = tdist.Uniform(
erase_scale_range[0],
erase_scale_range[1]).sample(params_shape) * images_area
if aspect_ratio_range[0] < 1. and aspect_ratio_range[1] > 1.:
aspect_ratios1 = tdist.Uniform(aspect_ratio_range[0],
1).sample(params_shape)
aspect_ratios2 = tdist.Uniform(
1, aspect_ratio_range[1]).sample(params_shape)
rand_idxs = torch.round(torch.rand(params_shape)).bool()
aspect_ratios = torch.where(rand_idxs, aspect_ratios1, aspect_ratios2)
else:
aspect_ratios = tdist.Uniform(
aspect_ratio_range[0], aspect_ratio_range[1]).sample(params_shape)
# based on target areas and aspect ratios, rectangle params are computed
heights = torch.min(
torch.max(torch.round((target_areas * aspect_ratios)**(1 / 2)),
torch.tensor(1.)), torch.tensor(float(images_height))).int()
widths = torch.min(
torch.max(torch.round((target_areas / aspect_ratios)**(1 / 2)),
torch.tensor(1.)), torch.tensor(float(images_width))).int()
xs = (torch.rand(params_shape) * (images_width - widths + 1).float()).int()
ys = (torch.rand(params_shape) *
(images_height - heights + 1).float()).int()
return widths, heights, xs, ys
def sample(self, sample_shape=torch.Size()):
X = self.p.sample(sample_shape)
# short-circuit into C++
if (type(p), type(q)) in lookup.keys():
# print("short-circuiting into C++")
func = lookup[(type(p), type(q))]
return (X,
func(self.p.loc, self.p.scale, self.q.loc,
self.q.scale, X))
Y = X.clone()
W = dist.Uniform(0, self.p.log_prob(X).exp()).sample()
threshold = self.q.log_prob(Y).exp()
msk = (W <= threshold) # which samples have succeeded
# basically, ignore the values where (msk == 1), because they're finalized
# this is inefficient because:
# 1 - this creates many extra samples - because threads don't exit
# 2 - this computes the log prob many more times
# we can speed this up using a custom implementation of a coupled kernel
while not msk.all():
Yp = self.q.sample()
threshold = self.p.log_prob(Yp).exp()
W = dist.Uniform(0, self.q.log_prob(Yp).exp()).sample()
add = ((1 - msk) & (W > threshold))
Y = Y * (1 - add).type(Y.type()) + Yp * add.type(Y.type())
msk += add
return (X, Y)
def __init__(self, D, K, low = 0.1, high = 0.2, device="cuda:0", train_b=False):
super().__init__()
self.prior = td.Uniform(torch.zeros(K, device=device), torch.ones(K, device=device))
pW = td.Uniform(low, high)
W = pW.sample([K,D]).to(device)
self._W = nn.Parameter(torch.log(W.exp() - 1.0))
self.b = nn.Parameter(torch.zeros(D), requires_grad=train_b)
def __init__(self, noise_strength: float):
""" """
# Setup the distributions
self.amp_dist = distributions.Uniform(low=0.2, high=1)
self.freq_dist = distributions.Uniform(low=0.1, high=0.25)
self.phase_dist = distributions.Uniform(low=0, high=math.pi)
self.noise_dist = distributions.Uniform(low=-noise_strength,
high=noise_strength)
def get_simulator_and_prior(task):
if task == "nonlinear-gaussian":
simulator = simulators.NonlinearGaussianSimulator()
prior = distributions.Uniform(
low=-3 * torch.ones(simulator.parameter_dim),
high=3 * torch.ones(simulator.parameter_dim),
)
elif task == "nonlinear-gaussian-gaussian":
simulator = simulators.NonlinearGaussianSimulator()
prior = distributions.MultivariateNormal(
loc=torch.zeros(5), covariance_matrix=torch.eye(5)
)
elif task == "two-moons":
simulator = simulators.TwoMoonsSimulator()
a = 2
prior = distributions.Uniform(
low=-a * torch.ones(simulator.parameter_dim),
high=a * torch.ones(simulator.parameter_dim),
)
elif task == "linear-gaussian":
dim, std = 20, 0.5
simulator = simulators.LinearGaussianSimulator(dim=dim, std=std)
prior = distributions.MultivariateNormal(
loc=torch.zeros(dim), covariance_matrix=torch.eye(dim)
)
elif task == "lotka-volterra":
simulator = simulators.LotkaVolterraSimulator(
summarize_observations=True, gaussian_prior=False
)
prior = distributions.Uniform(
low=-5 * torch.ones(simulator.parameter_dim),
high=2 * torch.ones(simulator.parameter_dim),
)
elif task == "lotka-volterra-gaussian":
simulator = simulators.LotkaVolterraSimulator(
summarize_observations=True, gaussian_prior=True
)
prior = distributions.MultivariateNormal(
loc=torch.zeros(4), covariance_matrix=2 * torch.eye(4)
)
elif task == "mg1":
simulator = simulators.MG1Simulator(summarize_observations=True)
prior = distributions_.MG1Uniform(
low=torch.zeros(3), high=torch.Tensor([10.0, 10.0, 1.0 / 3.0])
)
else:
raise ValueError(f"'{task}' simulator choice not understood.")
return simulator, prior
def __init__(self, is_test):
super().__init__()
#self.flip_var_order = flip_var_order
#if is_test:
#self.pX = D.Uniform(torch.tensor([0.0]), torch.tensor([1.0]))
#else:
mix = D.Categorical(torch.ones(2, ))
comp = D.Uniform(torch.tensor([0.0, 0.35]), torch.tensor([0.45, 1.0]))
self.pX = D.MixtureSameFamily(mix, comp)
self.pY1 = D.Uniform(torch.tensor([0.0]), torch.tensor([1.0]))
self.pY2 = lambda X: D.Normal(torch.sin(10 * X), 0.05)
def uniform_perturbation(sample, x_min, x_max, n=1, sigma=1, seed=0, image_size=28):
import torch
import torch.distributions as dist
sample = torch.tensor(sample).view(-1, image_size*image_size)
if isinstance(x_min, (int, float, complex)) and isinstance(x_max, (int, float, complex)):
prior = dist.Uniform(low=torch.max(sample-sigma, torch.tensor([x_min])), high=torch.min(sample+sigma, torch.tensor([x_max])))
elif isinstance(x_min, torch.Tensor) and isinstance(x_max, torch.Tensor):
prior = dist.Uniform(low=torch.max(sample-sigma, x_min), high=torch.min(sample+sigma, x_max))
else:
raise ValueError('Type of x_min and x_max {0} is not supported'.format(type(x_min)))
x = prior.sample(torch.Size([n])).view(-1, image_size, image_size)
return x.numpy(), prior
def __init__(self, D_in, layers_sizes, kernel, M = 10,
sigma_prior = dist.Uniform(1., 2),
alpha_prior = dist.Uniform(0.25, 0.75)):
super(DGP, self).__init__()
self.layers_sizes = layers_sizes
self.layers = nn.ModuleList([])
for D_out in layers_sizes:
self.layers.append(SGPR(D_in, M = M,
kernel = kernel(D_out, sigma_prior = sigma_prior,
alpha_prior = alpha_prior),
D_out = D_out,
mean = lambda X: identity_mean))
D_in = D_out
def sample_z(args):
# generate samples from the prior
z_cat = OneHotCategorical(
logits=torch.zeros(args.batch_size, args.cat_dim)).sample()
z_noise = dist.Uniform(-1, 1).sample(
torch.Size((args.batch_size, args.noise_dim)))
z_cont = dist.Uniform(-1, 1).sample(
torch.Size((args.batch_size, args.cont_dim)))
# concatenate the incompressible noise, discrete latest, and continuous latents
z = torch.cat([z_noise, z_cat, z_cont], dim=1)
return z.to(args.device), z_cat.to(args.device), z_noise.to(
args.device), z_cont.to(args.device)
def test_():
# if torch.cuda.is_available():
# device = torch.device("cuda")
# torch.set_default_tensor_type("torch.cuda.FloatTensor")
# else:
# input("CUDA not available, do you wish to continue?")
# device = torch.device("cpu")
# torch.set_default_tensor_type("torch.FloatTensor")
loc = torch.Tensor([0, 0])
covariance_matrix = torch.Tensor([[1, 0.99], [0.99, 1]])
likelihood = distributions.MultivariateNormal(
loc=loc, covariance_matrix=covariance_matrix)
bound = 1.5
low, high = -bound * torch.ones(2), bound * torch.ones(2)
prior = distributions.Uniform(low=low, high=high)
# def potential_function(inputs_dict):
# parameters = next(iter(inputs_dict.values()))
# return -(likelihood.log_prob(parameters) + prior.log_prob(parameters).sum())
prior = distributions.Uniform(low=-5 * torch.ones(4),
high=2 * torch.ones(4))
from nsf import distributions as distributions_
likelihood = distributions_.LotkaVolterraOscillating()
potential_function = PotentialFunction(likelihood, prior)
# kernel = Slice(potential_function=potential_function)
from pyro.infer.mcmc import HMC, NUTS
# kernel = HMC(potential_fn=potential_function)
kernel = NUTS(potential_fn=potential_function)
num_chains = 3
sampler = MCMC(
kernel=kernel,
num_samples=10000 // num_chains,
warmup_steps=200,
initial_params={"": torch.zeros(num_chains, 4)},
num_chains=num_chains,
)
sampler.run()
samples = next(iter(sampler.get_samples().values()))
utils.plot_hist_marginals(utils.tensor2numpy(samples),
ground_truth=utils.tensor2numpy(loc),
lims=[-6, 3])
# plt.show()
plt.savefig("/home/conor/Dropbox/phd/projects/lfi/out/mcmc.pdf")
plt.close()
def dequantize(x, constraint=0.9, inverse=False):
if inverse:
x = 2 / (torch.exp(-x) + 1) - 1
x /= constraint
x = (x + 1) / 2
return x, 0
else:
B, C, H, W = x.shape
noise = distributions.Uniform(0., 1.).sample(x.shape)
x = (x * 255 + noise) / 256
return x
x = x * 2 - 1
x *= constraint
x = (x + 1) / 2
logit_x = x.log() - (1 - x).log()
pre_logit_scale = torch.Tensor([constraint]).log() \
- torch.Tensor([1 - constraint]).log()
log_det_J = F.softplus(logit_x) + F.softplus(-logit_x) \
- F.softplus(-pre_logit_scale)
return logit_x, log_det_J.view(B, -1).sum(1, keepdim=True)
def brute_force(model, domain, count_iterations=100):
count_above = int(0)
count_total = int(0)
if CUDA:
low_params = domain[:,0].cuda()
high_params = domain[:,1].cuda()
else:
low_params = domain[:,0]
high_params = domain[:,1]
prior = dist.Uniform(low=low_params, high=high_params)
count_particles = int(1000000)
start = time.time()
max_val = -math.inf
for i in range(count_iterations):
x = prior.sample(torch.Size([count_particles]))
s_x = model(x).squeeze(-1)
count_above += int((s_x >= 0).float().sum().item())
count_total += count_particles
max_val = max(s_x.max().item(), max_val)
if i % 1000 == 0:
time_per_iter = (time.time() - start) / (i+1)
time_left = (count_iterations - i) * time_per_iter / 60
print(f'{i/count_iterations*100}% done ({round(time_left,3)} mins left)')
print(f'{count_above} adversarial examples observed from {count_total} samples')
if count_above > 0:
return math.log(count_above) - math.log(count_total), max_val, count_above, count_total
else:
return -math.inf, max_val, count_above, count_total
def simulate(
self,
input_obj: torch.Tensor,
num_sim: int,
seed: Optional[int] = None,
) -> np.ndarray:
if seed is not None:
torch.manual_seed(seed)
# Get the predicted location and scale parameters
predictions = self.forward(input_obj)
locations, scales = predictions[:, 0], predictions[:, 1]
# Build a Folded Logistic Di stribution
# X ~ Uniform(0, 1)
# f = a + b * logit(X)
# Y ~ f(X) ~ Logistic(a, b)
# Z ~ |Y| ~ FoldedLogistic(a, b)
base_distribution = dists.Uniform(0, 1)
transforms = [
dists.transforms.SigmoidTransform().inv,
dists.transforms.AffineTransform(loc=locations, scale=scales),
dists.transforms.AbsTransform(),
]
folded_logistic_dists = (
dists.transformed_distribution.TransformedDistribution(
base_distribution, transforms))
# Sample from the distributions
samples = folded_logistic_dists.sample((num_sim, ))
return samples.numpy()
def getGPmodel(self):
pyro.set_rng_seed(self.seed)
pyro.clear_param_store()
args = self.getInitXY()
if self.dim1(args) == True:
X4Kernel = torch.tensor(self.getFlotList(args[0]))
dim = 1
else:
X4Kernel = torch.tensor([self.getFlotList(args[0])])
dim = X4Kernel.shape[1]
y4Kernel = torch.tensor(self.getFlotList(args[1]))
#X4Kernel = torch.tensor([[2.0, 40], [6.0, 40], [6.0, 80], [8.0, 90]])
#y4Kernel = torch.tensor([1.0, 2.0, 3.0 ,4.0])
#kernel = gp.kernels.RBF(input_dim=dim)
kernel = gp.kernels.Matern52(input_dim=dim)
kernel.set_prior("variance",
dist.Uniform(torch.tensor(0.5), torch.tensor(1.5)))
kernel.set_prior("lengthscale",
dist.Gamma(torch.tensor(7.5), torch.tensor(1.)))
gpmodel = gp.models.GPRegression(X4Kernel,
y4Kernel,
kernel,
noise=torch.tensor(self.noise),
jitter=1.0e-4) #1.0e-4
return gpmodel
def __init__(self, n_hidden=100, bottom_width=7, in_ch=128):
super(Generator, self).__init__()
self.n_hidden = n_hidden
self.in_ch = in_ch
self.bottom_width = bottom_width
self.uniform = distributions.Uniform(-1, 1)
# register parameters
self.l0 = nn.Linear(in_features=self.n_hidden,
out_features=self.bottom_width*self.bottom_width*self.in_ch,
bias=False)
self.dc1 = nn.ConvTranspose2d(in_channels=self.in_ch,
out_channels=self.in_ch//2,
kernel_size=4,
stride=2,
padding=1,
bias=False) # (N, 3, 14, 14)
self.dc2 = nn.ConvTranspose2d(in_channels=self.in_ch//2,
out_channels=1,
kernel_size=4,
stride=2,
padding=1,
bias=False) # (N, 1, 24, 24)
self.bn0 = nn.BatchNorm1d(
num_features=self.bottom_width*self.bottom_width*self.in_ch)
self.bn1 = nn.BatchNorm2d(num_features=self.in_ch//2)
def logistic_distribution(loc: Tensor, scale: Tensor):
base_distribution = td.Uniform(loc.new_zeros(1), scale.new_zeros(1))
transforms = [
td.SigmoidTransform().inv,
td.AffineTransform(loc=loc, scale=scale)
]
return td.TransformedDistribution(base_distribution, transforms)
def run_pgdn(net, data, get_feature, logger, ad_save_path, params):
""" Computes PGD on negative loss for given data. """
dist = dists.Uniform(-params.epsilon, params.epsilon)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
net.to(device)
net.eval()
for i, (x, y) in enumerate(data):
print('pgd sample {}/{}'.format(i + 1, len(data)))
file = data.dataset.filenames[i]
file = file if '.wav' in file else file + '.wav'
ad_ex, db, converged = _pgdn(net, params, x.to(device), y.to(device),
get_feature, dist)
orig_pred = net.predict(get_feature(x.to(device))).item()
true_pred = y.item()
if converged >= 0:
print('\nsave current file: {} with distortion: {}db (r:{}/o:{})'.
format(file, db, true_pred, orig_pred))
new_pred = net.predict(get_feature(ad_ex)).item()
save_adv_example(ad_ex.cpu(), os.path.join(ad_save_path, file))
logger.append(
[file, converged, db, True, true_pred, orig_pred, new_pred])
else:
print('\ncould not find robust adv. example for file {}'.format(
file))
logger.append(
[file, converged, 0, False, true_pred, orig_pred, orig_pred])
def sample_gamma(self, shape, scale):
augment = 10
# get Gamma(shape + factor, 1)
with torch.no_grad():
sample = distributions.Gamma(shape + augment, 1).sample()
eps = torch.sqrt(9. * (shape + augment) - 3.) * ((
(sample / (shape + augment - (1. / 3.)))**(1. / 3.)) - 1.)
z = (shape + augment -
(1. / 3.)) * ((1. +
(eps / torch.sqrt(9. *
(shape + augment) - 3.)))**3.)
# reduce factor
with torch.no_grad():
expand_shape = shape.unsqueeze(-1).repeat(1, 1, augment)
factor_range = torch.arange(
1, augment + 1, dtype=torch.float,
device=self.device).expand_as(expand_shape)
u = distributions.Uniform(
torch.zeros(factor_range.size(), device=self.device),
torch.ones(factor_range.size(), device=self.device)).sample()
u_prod = torch.prod(
u**(1. / (expand_shape + factor_range - 1. + 1e-12)), -1)
z = z * u_prod * scale
return z
def random(state, algorithm, body):
'''Random action sampling that returns the same data format as default(), but without forward pass. Uses gym.space.sample()'''
state = try_preprocess(
state, algorithm, body,
append=True) # for consistency with init_action_pd inner logic
if body.action_type == 'discrete':
action_pd = distributions.Categorical(logits=torch.ones(
body.action_space.high, device=algorithm.net.device))
elif body.action_type == 'continuous':
# Possibly this should this have a 'device' set
action_pd = distributions.Uniform(
low=torch.tensor(body.action_space.low).float(),
high=torch.tensor(body.action_space.high).float())
elif body.action_type == 'multi_discrete':
action_pd = distributions.Categorical(
logits=torch.ones(body.action_space.high.size,
body.action_space.high[0],
device=algorithm.net.device))
elif body.action_type == 'multi_continuous':
raise NotImplementedError
elif body.action_type == 'multi_binary':
raise NotImplementedError
else:
raise NotImplementedError
sample = body.action_space.sample()
action = torch.tensor(sample, device=algorithm.net.device)
return action, action_pd
def MHKernel(X, q, pi):
X = X.clone()
Xp = q(X).sample()
U = dist.Uniform(0, 1).sample(X.size())
msk = (U.log() <= pi.log_prob(Xp) + q(Xp).log_prob(X) -
(pi.log_prob(X) + q(X).log_prob(Xp)))
X = (1 - msk).float() * X + msk.float() * Xp
return X
def random(state, algorithm, body):
'''Random action sampling that returns the same data format as default(), but without forward pass. Uses gym.space.sample()'''
action_pd = distributions.Uniform(
low=torch.from_numpy(np.array(body.action_space.low)).float(),
high=torch.from_numpy(np.array(body.action_space.high)).float())
sample = body.action_space.sample()
action = torch.tensor(sample)
return action, action_pd
def __init__(self, loc, scale, **kwargs):
loc, scale = map(torch.as_tensor, (loc, scale))
base_distribution = ptd.Uniform(torch.zeros_like(loc),
torch.ones_like(loc), **kwargs)
transforms = [
ptd.SigmoidTransform().inv,
ptd.AffineTransform(loc=loc, scale=scale),
]
super().__init__(base_distribution, transforms)
def unif_sample(sample_shape=[1], low=0, high=1):
"""Generate a sample from U(low, high).
:param sample_shape: The shape S
:param low: Lower bound on uniform
:param high: Upper bound on uniform
:return: A matrix of shape S containg samples from the uniform.
"""
d = dist.Uniform(low, high)
return d.rsample(sample_shape)
def sample_relation_type(self, prev_parts):
"""
Sample a relation type from the prior for the current stroke,
conditioned on the previous strokes
Parameters
----------
prev_parts : list of StrokeType
previous part types
Returns
-------
r : RelationType
relation type sample
"""
for p in prev_parts:
assert isinstance(p, StrokeType)
nprev = len(prev_parts)
stroke_ix = nprev
# first sample the relation category
if nprev == 0:
category = 'unihist'
else:
indx = self.rel_mixdist.sample()
category = self.__relation_categories[indx]
# now sample the category-specific type-level parameters
if category == 'unihist':
data_id = torch.tensor([stroke_ix])
gpos = self.Spatial.sample(data_id)
# convert (1,2) tensor to (2,) tensor
gpos = torch.squeeze(gpos)
r = RelationIndependent(category, gpos, self.Spatial.xlim,
self.Spatial.ylim)
elif category in ['start', 'end', 'mid']:
# sample random stroke uniformly from previous strokes. this is the
# stroke we will attach to
probs = torch.ones(nprev)
attach_ix = dist.Categorical(probs=probs).sample()
if category == 'mid':
# sample random sub-stroke uniformly from the selected stroke
nsub = prev_parts[attach_ix].nsub
probs = torch.ones(nsub)
attach_subix = dist.Categorical(probs=probs).sample()
# sample random type-level spline coordinate
_, lb, ub = bspline_gen_s(self.ncpt, 1)
eval_spot = dist.Uniform(lb, ub).sample()
r = RelationAttachAlong(category, attach_ix, attach_subix,
eval_spot, self.ncpt)
else:
r = RelationAttach(category, attach_ix)
else:
raise TypeError('invalid relation')
return r
def __init__(self):
mean = torch.log(torch.Tensor([0.01, 0.5, 1, 0.01]))
sigma = 0.5
covariance = sigma**2 * torch.eye(4)
self._gaussian = distributions.MultivariateNormal(
loc=mean, covariance_matrix=covariance)
self._uniform = distributions.Uniform(low=-5 * torch.ones(4),
high=2 * torch.ones(4))
self._log_normalizer = -torch.log(
torch.erf((2 - mean) / sigma) -
torch.erf((-5 - mean) / sigma)).sum()
def logistic_distribution(loc, log_scale):
scale = torch.exp(log_scale) + 1e-5
base_distribution = distributions.Uniform(torch.zeros_like(loc),
torch.ones_like(loc))
transforms = [
LogisticTransform(),
distributions.AffineTransform(loc=loc, scale=scale)
]
logistic = distributions.TransformedDistribution(base_distribution,
transforms)
return logistic
def score_relation_type(self, prev_parts, r):
"""
Compute the log-probability of the relation type of the current stroke
under the prior
Parameters
----------
prev_parts : list of StrokeType
previous stroke types
r : RelationType
relation type to score
Returns
-------
ll : tensor
scalar; log-probability of the relation type
"""
assert isinstance(r, RelationType)
for p in prev_parts:
assert isinstance(p, StrokeType)
nprev = len(prev_parts)
stroke_ix = nprev
# first score the relation category
if nprev == 0:
ll = 0.
else:
ix = self.__relation_categories.index(r.category)
ix = torch.tensor(ix, dtype=torch.long)
ll = self.rel_mixdist.log_prob(ix)
# now score the category-specific type-level parameters
if r.category == 'unihist':
data_id = torch.tensor([stroke_ix])
# convert (2,) tensor to (1,2) tensor
gpos = r.gpos.view(1, 2)
# score the type-level location
ll = ll + torch.squeeze(self.Spatial.score(gpos, data_id))
elif r.category in ['start', 'end', 'mid']:
# score the stroke attachment index
probs = torch.ones(nprev)
ll = ll + dist.Categorical(probs=probs).log_prob(r.attach_ix)
if r.category == 'mid':
# score the sub-stroke attachment index
nsub = prev_parts[r.attach_ix].nsub
probs = torch.ones(nsub)
ll = ll + dist.Categorical(probs=probs).log_prob(
r.attach_subix)
# score the type-level spline coordinate
_, lb, ub = bspline_gen_s(self.ncpt, 1)
ll = ll + dist.Uniform(lb, ub).log_prob(r.eval_spot)
else:
raise TypeError('invalid relation')
return ll