def model():
mu_latent = pyro.sample("mu_latent", normal, self.mu0,
torch.pow(self.lam0, -0.5))
for i, x in enumerate(self.data):
pyro.observe("obs_%d" % i, dist.normal, x, mu_latent,
torch.pow(self.lam, -0.5))
return mu_latent
def forward(self, prefix, alpha, temperature=0.8, condition=False, generate=0):
assert len(prefix) > 0
hidden = self.rnn.init_hidden()
out = self.char_tensor(prefix[0])
generated_chars = prefix[0]
for i in range(1, len(prefix) + generate):
out, hidden = self.rnn(out, hidden)
ps = softmax(out.mul(alpha.expand(out.size())))
dist = Categorical(ps, one_hot=False)
name = 'char_{0}'.format(i)
if i < len(prefix):
# Use character provided in prefix
char = prefix[i]
if condition:
char_index = self.all_chars.index(char)
observe(name, dist, Tensor([char_index]))
else:
# Sample a character
char_index = sample(name, dist).data[0][0] # FIXME
char = self.all_chars[char_index]
generated_chars += char
out = self.char_tensor(char)
return generated_chars
def model(self, data):
a_tilda = Variable(torch.ones(self.data_size[0]))
b_tilda = Variable(torch.ones(self.data_size[0]))
eps = pyro.sample('eps', dist.gamma, a_tilda, a_tilda / b_tilda)
eps = torch.cat([eps] * self.K, 1)
a = Variable(torch.ones(self.data_size[0], self.K))
theta = pyro.sample('theta', dist.gamma, a, eps)
c_tilda = Variable(torch.ones(self.data_size[1]))
d_tilda = Variable(torch.ones(self.data_size[1]))
eta = pyro.sample('eta', dist.gamma, c_tilda, c_tilda / d_tilda)
eta = torch.cat([eta] * self.K, 1)
c = Variable(torch.ones(self.data_size[1], self.K))
beta = pyro.sample('beta', dist.gamma, c, eta)
zeta = pyro.sample('zeta', dist.poisson,
torch.matmul(theta, torch.t(beta)))
for i in range(self.data_size[0]):
for j in range(self.data_size[1]):
if data[i, j] == 0:
continue
pyro.observe("obs_{}{}".format(i, j), dist.poisson, data[i, j],
zeta[i, j])
def model():
mu = pyro.sample(
"mu", Normal(Variable(torch.zeros(1)),
Variable(torch.ones(1))))
xd = Normal(mu, Variable(torch.ones(1)), batch_size=50)
pyro.observe("xs", xd, self.data)
return mu
def model():
mu_latent = pyro.sample(
"mu_latent", DiagNormal(self.mu0, torch.pow(self.tau0, -0.5)))
x_dist = LogNormal(mu_latent, torch.pow(self.tau, -0.5))
pyro.observe("obs0", x_dist, self.data[0])
pyro.observe("obs1", x_dist, self.data[1])
return mu_latent
def model():
lambda_latent = pyro.sample("lambda_latent",
Gamma(self.alpha0, self.beta0))
x_dist = Exponential(lambda_latent)
pyro.observe("obs0", x_dist, self.data[0])
pyro.observe("obs1", x_dist, self.data[1])
return lambda_latent
def model_given_c(data, cll):
decoder_c = pyro.module("decoder_c", pt_decode_c)
decoder_z = pyro.module("decoder_z", pt_decode_z)
z_mu, z_sigma = decoder_c.forward(cll)
z = pyro.sample("latent_z", DiagNormal(z_mu, z_sigma))
img_mu = decoder_z.forward(z)
pyro.observe("obs", Bernoulli(img_mu), data.view(-1, 784))
def model(self, data):
decoder = pyro.module('decoder', self.vae_decoder)
z_mean, z_std = ng_zeros([data.size(0),
20]), ng_ones([data.size(0), 20])
z = pyro.sample('latent', Normal(z_mean, z_std))
img = decoder.forward(z)
pyro.observe('obs', Bernoulli(img), data.view(-1, 784))
def model(self):
# register PyTorch module `decoder` with Pyro
pyro.module("decoder", self.decoder)
# Setup hyperparameters for prior p(z)
z_mu = ng_zeros([self.n_samples, self.n_latent])
z_sigma = ng_ones([self.n_samples, self.n_latent])
# sample from prior
z = pyro.sample("latent", dist.normal, z_mu, z_sigma)
# decode the latent code z
z_adj = self.decoder(z)
# Subsampling
if self.subsampling:
with pyro.iarange("data",
self.n_subsample,
subsample=self.sample()) as ind:
pyro.observe('obs', dist.bernoulli,
self.adj_labels.view(1, -1)[0][ind],
z_adj.view(1, -1)[0][ind])
# Reweighting
else:
with pyro.iarange("data"):
pyro.observe('obs',
weighted_bernoulli,
self.adj_labels.view(1, -1),
z_adj.view(1, -1),
weight=self.pos_weight)
def model():
mu_latent = pyro.sample("mu_latent", dist.normal,
self.mu0, torch.pow(self.tau0, -0.5))
sigma = torch.pow(self.tau, -0.5)
pyro.observe("obs0", dist.lognormal, self.data[0], mu_latent, sigma)
pyro.observe("obs1", dist.lognormal, self.data[1], mu_latent, sigma)
return mu_latent
def guide_given_c(data, cll):
encoder_x = pyro.module("encoder_x", pt_encode_x)
encoder_z = pyro.module("encoder_z", pt_encode_z)
z_mu, z_sigma = encoder_x.forward(data)
z = pyro.sample("latent_z", DiagNormal(z_mu, z_sigma))
alpha_cat = encoder_z.forward(z)
pyro.observe("latent_class", Categorical(alpha_cat), cll)
def model_xz(data, foo):
decoder_xz = pyro.module("decoder_xz", pt_decode_xz)
z_mu, z_sigma = Variable(torch.zeros([data.size(0), 20])), Variable(
torch.ones([data.size(0), 20]))
z = pyro.sample("latent", DiagNormal(z_mu, z_sigma))
img_mu = decoder_xz.forward(z)
pyro.observe("obs", Bernoulli(img_mu), data.view(-1, 784))
return z
def model():
mu_latent = pyro.sample("mu_latent", dist.normal, self.mu0,
torch.pow(self.tau0, -0.5))
bijector = AffineExp(torch.pow(self.tau, -0.5), mu_latent)
x_dist = TransformedDistribution(dist.normal, bijector)
pyro.observe("obs0", x_dist, self.data[0], ng_zeros(1), ng_ones(1))
pyro.observe("obs1", x_dist, self.data[1], ng_zeros(1), ng_ones(1))
return mu_latent
def local_model(i, datum):
beta = Variable(torch.ones(1, 10)) * 0.1
cll = pyro.sample("class_of_datum_" + str(i), Categorical(beta))
mean_param = Variable(torch.zeros(1, 784), requires_grad=True)
# do MLE for class means
mu = pyro.param("mean_of_class_" + str(cll[0]), mean_param)
pyro.observe("obs_" + str(i), Bernoulli(mu), datum)
return cll
def obs_inner(i, _i, _x):
for k in range(n_superfluous_top):
pyro.sample("z_%d_%d" % (i, k),
dist.Normal(ng_zeros(4 - i, 1), ng_ones(4 - i, 1), reparameterized=False))
pyro.observe("obs_%d" % i, dist.normal, _x, mu_latent, torch.pow(self.lam, -0.5))
for k in range(n_superfluous_top, n_superfluous_top + n_superfluous_bottom):
pyro.sample("z_%d_%d" % (i, k),
dist.Normal(ng_zeros(4 - i, 1), ng_ones(4 - i, 1), reparameterized=False))
def model():
mu_latent = pyro.sample(
"mu_latent",
dist.Normal(self.mu0, torch.pow(self.lam0, -0.5), reparameterized=reparameterized))
for i, x in enumerate(self.data):
pyro.observe("obs_%d" % i, dist.normal, x, mu_latent,
torch.pow(self.lam, -0.5))
return mu_latent
def model():
mu_latent = pyro.sample("mu_latent", dist.normal,
self.mu0, torch.pow(self.tau0, -0.5))
bijector = AffineExp(torch.pow(self.tau, -0.5), mu_latent)
x_dist = TransformedDistribution(dist.normal, bijector)
pyro.observe("obs0", x_dist, self.data[0], ng_zeros(1), ng_ones(1))
pyro.observe("obs1", x_dist, self.data[1], ng_zeros(1), ng_ones(1))
return mu_latent
def pyro_model(self, x):
pyro.module("output_embedding", self.output_embedding)
mu = variable(self, torch.zeros(x.size(0), self.num_latent_factors))
sigma = variable(self, torch.ones(x.size(0), self.num_latent_factors))
z = pyro.sample("latent", dist.normal, mu, sigma)
out_prob = (self.sigmoid(self.decode(z)) + fudge) * (1 - 2 * fudge)
pyro.observe("obs", dist.bernoulli, x, out_prob)
return z
def guide_observed2(data, cll):
encoder_c = pyro.module("encoder_c", pt_encode_c)
alpha = encoder_c.forward(data)
pyro.observe("latent_class", Categorical(alpha), cll)
encoder = pyro.module("encoder_o", pt_encode_o)
z_mu, z_sigma = encoder.forward(data, cll)
z = pyro.sample("latent_z", DiagNormal(z_mu, z_sigma))
def sub_model(datum):
mu_latent = Variable(torch.ones(nr_samples, dim_z)) * 0.5
sigma_latent = Variable(torch.ones(mu_latent.size()))
z = pyro.sample("embedding_of_datum_" + str(i),
DiagNormal(mu_latent, sigma_latent))
mean_beta = z.mm(weight)
beta = sigmoid(mean_beta)
pyro.observe("obs_" + str(i), Bernoulli(beta), datum)
def model(self, input_variable, target_variable, step):
# register PyTorch module `decoder` with Pyro
pyro.module("decoder_dense", self.decoder_dense)
pyro.module("decoder_rnn", self.decoder_rnn)
# setup hyperparameters for prior p(z)
# the type_as ensures we get CUDA Tensors if x is on gpu
z_mu = ng_zeros([self.num_layers, self.z_dim], type_as=target_variable.data)
z_sigma = ng_ones([self.num_layers, self.z_dim], type_as=target_variable.data)
# sample from prior
# (value will be sampled by guide when computing the ELBO)
z = pyro.sample("latent", dist.normal, z_mu, z_sigma)
# init vars
target_length = target_variable.shape[0]
decoder_input = dataset.to_onehot([[self.dataset.SOS_index]])
decoder_input = decoder_input.cuda() if USE_CUDA else decoder_input
decoder_outputs = np.ones((target_length))
decoder_hidden = self.decoder_dense(z)
# # Teacher forcing
for di in range(target_length):
decoder_output, decoder_hidden = self.decoder_rnn(
decoder_input, decoder_hidden)
decoder_input = target_variable[di]
if self.use_cuda:
decoder_outputs[di] = np.argmax(decoder_output.cpu().data.numpy())
else:
decoder_outputs[di] = np.argmax(decoder_output.data.numpy())
pyro.observe("obs_{}".format(di), dist.bernoulli, target_variable[di], decoder_output[0])
# ----------------------------------------------------------------
# prepare offer
if self.use_cuda:
offer = np.argmax(input_variable.cpu().data.numpy(), axis=1).astype(int)
else:
offer = np.argmax(input_variable.data.numpy(), axis=1).astype(int)
# prepare answer
if self.use_cuda:
answer = np.argmax(target_variable.cpu().data.numpy(), axis=1).astype(int)
else:
answer = np.argmax(target_variable.data.numpy(), axis=1).astype(int)
# prepare rnn
rnn_response = list(map(int, decoder_outputs))
# print output
if step % 10 == 0:
print("---------------------------")
print("Offer: ", dataset.to_phrase(offer))
print("Answer:", self.dataset.to_phrase(answer))
print("RNN:", self.dataset.to_phrase(rnn_response))
def model_latent(data):
decoder_c = pyro.module("decoder_c", pt_decode_c)
decoder_z = pyro.module("decoder_z", pt_decode_z)
alpha = Variable(torch.ones([data.size(0), 10])) / 10.
cll = pyro.sample('latent_class', Categorical(alpha))
z_mu, z_sigma = decoder_c.forward(cll)
z = pyro.sample("latent_z", DiagNormal(z_mu, z_sigma))
img_mu = decoder_z.forward(z)
pyro.observe("obs", Bernoulli(img_mu), data.view(-1, 784))
def model():
latent1 = pyro.sample(
"latent1",
Normal(Variable(torch.zeros(2)), Variable(torch.ones(2))))
latent2 = pyro.sample("latent2",
Normal(latent1, 5 * Variable(torch.ones(2))))
x_dist = Normal(latent2, Variable(torch.ones(2)))
pyro.observe("obs", x_dist, Variable(torch.ones(2)))
return latent1
def model():
latent1 = pyro.sample("latent1",
Normal(Variable(torch.zeros(2)),
Variable(torch.ones(2))))
latent2 = pyro.sample("latent2",
Normal(latent1,
5 * Variable(torch.ones(2))))
x_dist = Normal(latent2, Variable(torch.ones(2)))
pyro.observe("obs", x_dist, Variable(torch.ones(2)))
return latent1
def local_model(i, datum):
beta = Variable(torch.ones(1)) * 0.5
c = pyro.sample("class_of_datum_" + str(i), Bernoulli(beta))
mean_param = Variable(torch.zeros(784), requires_grad=True)
# do MLE for class means
m = pyro.param("mean_of_class_" + str(c[0]), mean_param)
sigma = Variable(torch.ones(m.size()))
pyro.observe("obs_" + str(i), DiagNormal(m, sigma), datum)
return c
def model():
mu_latent = pyro.sample(
"mu_latent", DiagNormal(self.mu0, torch.pow(self.tau0, -0.5)))
unit_normal = dist.DiagNormal(Variable(torch.zeros(1, 1)),
Variable(torch.ones(1, 1)))
bijector = AffineExp(torch.pow(self.tau, -0.5), mu_latent)
x_dist = TransformedDistribution(unit_normal, bijector)
# x_dist = LogNormal(mu_latent, torch.pow(self.tau,-0.5))
pyro.observe("obs0", x_dist, self.data[0])
pyro.observe("obs1", x_dist, self.data[1])
return mu_latent
def model():
mu_latent_prime = pyro.sample(
"mu_latent_prime",
dist.Normal(self.mu0, torch.pow(self.lam0, -0.5), reparameterized=repa1))
mu_latent = pyro.sample(
"mu_latent",
dist.Normal(mu_latent_prime, torch.pow(self.lam0, -0.5), reparameterized=repa2))
for i, x in enumerate(self.data):
pyro.observe("obs_%d" % i, dist.normal, x, mu_latent,
torch.pow(self.lam, -0.5))
return mu_latent
def model():
alpha_p_log = pyro.param(
"alpha_p_log", Variable(
self.alpha_p_log_0.clone(), requires_grad=True), tags="model")
beta_p_log = pyro.param(
"beta_p_log", Variable(
self.beta_p_log_0.clone(), requires_grad=True), tags="model")
alpha_p, beta_p = torch.exp(alpha_p_log), torch.exp(beta_p_log)
lambda_latent = pyro.sample("lambda_latent", dist.gamma, alpha_p, beta_p)
pyro.observe("obs", dist.poisson, self.data, lambda_latent)
return lambda_latent
def gmm_batch_model(data):
p = pyro.param("p", Variable(torch.Tensor([0.3]), requires_grad=True))
p = torch.cat([p, 1 - p])
sigma = pyro.param("sigma", Variable(torch.Tensor([1.0]), requires_grad=True))
mus = Variable(torch.Tensor([-1, 1]))
with pyro.iarange("data", len(data)) as batch:
n = len(batch)
z = pyro.sample("z", dist.Categorical(p.unsqueeze(0).expand(n, 2)))
assert z.size() == (n, 2)
mu = torch.mv(z, mus)
pyro.observe("x", dist.Normal(mu, sigma.expand(n)), data[batch])
def gmm_model(data, verbose=False):
p = pyro.param("p", Variable(torch.Tensor([0.3]), requires_grad=True))
sigma = pyro.param("sigma", Variable(torch.Tensor([1.0]), requires_grad=True))
mus = Variable(torch.Tensor([-1, 1]))
for i in pyro.irange("data", len(data)):
z = pyro.sample("z_{}".format(i), dist.Bernoulli(p))
assert z.size() == (1,)
z = z.long().data[0]
if verbose:
print("M{} z_{} = {}".format(" " * i, i, z))
pyro.observe("x_{}".format(i), dist.Normal(mus[z], sigma), data[i])
def gmm_batch_model(data):
p = pyro.param("p", Variable(torch.Tensor([0.3]), requires_grad=True))
p = torch.cat([p, 1 - p])
sigma = pyro.param("sigma",
Variable(torch.Tensor([1.0]), requires_grad=True))
mus = Variable(torch.Tensor([-1, 1]))
with pyro.iarange("data", len(data)) as batch:
n = len(batch)
z = pyro.sample("z", dist.Categorical(p.unsqueeze(0).expand(n, 2)))
assert z.size() == (n, 2)
mu = torch.mv(z, mus)
pyro.observe("x", dist.Normal(mu, sigma.expand(n)), data[batch])
def gmm_model(data, verbose=False):
p = pyro.param("p", Variable(torch.Tensor([0.3]), requires_grad=True))
sigma = pyro.param("sigma",
Variable(torch.Tensor([1.0]), requires_grad=True))
mus = Variable(torch.Tensor([-1, 1]))
for i in pyro.irange("data", len(data)):
z = pyro.sample("z_{}".format(i), dist.Bernoulli(p))
assert z.size() == (1, )
z = z.long().data[0]
if verbose:
print("M{} z_{} = {}".format(" " * i, i, z))
pyro.observe("x_{}".format(i), dist.Normal(mus[z], sigma), data[i])
def model(*args, **kwargs):
next_mean = self.mu0
for k in range(1, self.N + 1):
latent_dist = dist.Normal(next_mean, torch.pow(self.lambdas[k - 1], -0.5))
mu_latent = pyro.sample("mu_latent_%d" % k, latent_dist)
next_mean = mu_latent
mu_N = next_mean
for i, x in enumerate(self.data):
pyro.observe("obs_%d" % i, dist.normal, x, mu_N,
torch.pow(self.lambdas[self.N], -0.5))
return mu_N
def model():
alpha_p_log = pyro.param(
"alpha_p_log", Variable(self.alpha_p_log_0,
requires_grad=True))
beta_p_log = pyro.param(
"beta_p_log", Variable(self.beta_p_log_0, requires_grad=True))
alpha_p, beta_p = torch.exp(alpha_p_log), torch.exp(beta_p_log)
lambda_latent = pyro.sample("lambda_latent",
Gamma(alpha_p, beta_p))
x_dist = Poisson(lambda_latent)
pyro.observe("obs", x_dist, self.data)
return lambda_latent
def model(*args, **kwargs):
next_mean = self.mu0
for k in range(1, self.N + 1):
latent_dist = dist.Normal(next_mean, torch.pow(self.lambdas[k - 1], -0.5))
mu_latent = pyro.sample("mu_latent_%d" % k, latent_dist)
next_mean = mu_latent
mu_N = next_mean
for i, x in enumerate(self.data):
pyro.observe("obs_%d" % i, dist.normal, x, mu_N,
torch.pow(self.lambdas[self.N], -0.5))
return mu_N
def model(self, x):
# register PyTorch module `decoder` with Pyro
pyro.module("decoder", self.decoder)
# setup hyperparameters for prior p(z)
# the type_as ensures we get cuda Tensors if x is on gpu
z_mu = ng_zeros([x.size(0), self.z_dim], type_as=x.data)
z_sigma = ng_ones([x.size(0), self.z_dim], type_as=x.data)
# sample from prior (value will be sampled by guide when computing the ELBO)
z = pyro.sample("latent", dist.normal, z_mu, z_sigma)
# decode the latent code z
mu_img = self.decoder.forward(z)
# score against actual images
pyro.observe("obs", dist.bernoulli, x.view(-1, 784), mu_img)
def model(self, x):
# register PyTorch module `decoder` with Pyro
pyro.module("decoder", self.decoder)
# setup hyperparameters for prior p(z)
# the type_as ensures we get cuda Tensors if x is on gpu
z_mu = ng_zeros([x.size(0), self.z_dim], type_as=x.data)
z_sigma = ng_ones([x.size(0), self.z_dim], type_as=x.data)
# sample from prior (value will be sampled by guide when computing the ELBO)
z = pyro.sample("latent", dist.normal, z_mu, z_sigma)
# decode the latent code z
mu_img = self.decoder.forward(z)
# score against actual images
pyro.observe("obs", dist.bernoulli, x.view(-1, 784), mu_img)
def model():
p_latent = pyro.sample("p_latent", dist.beta, self.alpha0, self.beta0)
pyro.map_data("aaa",
self.data, lambda i, x: pyro.observe(
"obs_{}".format(i), dist.bernoulli, x, p_latent),
batch_size=self.batch_size)
return p_latent
def model_classify(self, xs, ys=None):
"""
this model is used to add an auxiliary (supervised) loss as described in the
NIPS 2014 paper by Kingma et al titled
"Semi-Supervised Learning with Deep Generative Models"
"""
# register all pytorch (sub)modules with pyro
pyro.module("ss_vae", self)
# inform Pyro that the variables in the batch of xs, ys are conditionally independent
with pyro.iarange("independent"):
# this here is the extra Term to yield an auxiliary loss that we do gradient descend on
# similar to the NIPS 14 paper (Kingma et al).
if ys is not None:
alpha = self.encoder_y.forward(xs)
pyro.observe("y_aux", dist.categorical, ys, alpha, log_pdf_mask=self.aux_loss_multiplier)
def model():
mu_latent = pyro.sample("mu_latent", dist.normal,
self.mu0, torch.pow(self.lam0, -0.5))
pyro.map_data("aaa", self.data, lambda i,
x: pyro.observe(
"obs_%d" % i, dist.normal,
x, mu_latent, torch.pow(self.lam, -0.5)),
batch_size=self.batch_size)
return mu_latent
def model(data):
# Create unit normal priors over the parameters
mu = Variable(torch.zeros(p, 1)).type_as(data)
sigma = Variable(torch.ones(p, 1)).type_as(data)
bias_mu = Variable(torch.zeros(1)).type_as(data)
bias_sigma = Variable(torch.ones(1)).type_as(data)
w_prior, b_prior = Normal(mu, sigma), Normal(bias_mu, bias_sigma)
priors = {'linear.weight': w_prior, 'linear.bias': b_prior}
# lift module parameters to random variables sampled from the priors
lifted_module = pyro.random_module("module", regression_model, priors)
# sample a regressor (which also samples w and b)
lifted_reg_model = lifted_module()
with pyro.iarange("map", N, subsample=data):
x_data = data[:, :-1]
y_data = data[:, -1]
# run the regressor forward conditioned on inputs
prediction_mean = lifted_reg_model(x_data).squeeze()
pyro.observe("obs", Normal(prediction_mean, Variable(torch.ones(data.size(0))).type_as(data)), y_data.squeeze())
def model(self, mini_batch, mini_batch_reversed, mini_batch_mask,
mini_batch_seq_lengths, annealing_factor=1.0):
# this is the number of time steps we need to process in the mini-batch
T_max = mini_batch.size(1)
# register all PyTorch (sub)modules with pyro
# this needs to happen in both the model and guide
pyro.module("dmm", self)
# set z_prev = z_0 to setup the recursive conditioning in p(z_t | z_{t-1})
z_prev = self.z_0.expand(mini_batch.size(0), self.z_0.size(0))
# sample the latents z and observed x's one time step at a time
for t in range(1, T_max + 1):
# the next three lines of code sample z_t ~ p(z_t | z_{t-1})
# note that (both here and elsewhere) log_pdf_mask takes care of both
# (i) KL annealing; and
# (ii) raggedness in the observed data (i.e. different sequences
# in the mini-batch have different lengths)
# first compute the parameters of the diagonal gaussian distribution p(z_t | z_{t-1})
z_mu, z_sigma = self.trans(z_prev)
# then sample z_t according to dist.Normal(z_mu, z_sigma)
z_t = pyro.sample("z_%d" % t,
dist.normal,
z_mu,
z_sigma,
log_pdf_mask=annealing_factor * mini_batch_mask[:, t - 1:t])
# compute the probabilities that parameterize the bernoulli likelihood
emission_probs_t = self.emitter(z_t)
# the next statement instructs pyro to observe x_t according to the
# bernoulli distribution p(x_t|z_t)
pyro.observe("obs_x_%d" % t, dist.bernoulli, mini_batch[:, t - 1, :],
emission_probs_t,
log_pdf_mask=mini_batch_mask[:, t - 1:t])
# the latent sampled at this time step will be conditioned upon
# in the next time step so keep track of it
z_prev = z_t
def model(*args, **kwargs):
top_latent_dist = dist.Normal(self.mu0, torch.pow(self.lambdas[0], -0.5))
previous_names = ["mu_latent_1"]
top_latent = pyro.sample(previous_names[0], top_latent_dist)
previous_latents_and_names = list(zip([top_latent], previous_names))
# for sampling model variables in different sequential orders
def permute(x, n):
if model_permutation:
return [x[self.model_permutations[n - 1][i]] for i in range(len(x))]
return x
def unpermute(x, n):
if model_permutation:
return [x[self.model_unpermutations[n - 1][i]] for i in range(len(x))]
return x
for n in range(2, self.N + 1):
new_latents_and_names = []
for prev_latent, prev_name in permute(previous_latents_and_names, n - 1):
latent_dist = dist.Normal(prev_latent, torch.pow(self.lambdas[n - 1], -0.5))
couple = []
for LR in ['L', 'R']:
new_name = prev_name + LR
mu_latent_LR = pyro.sample(new_name, latent_dist)
couple.append([mu_latent_LR, new_name])
new_latents_and_names.append(couple)
_previous_latents_and_names = unpermute(new_latents_and_names, n - 1)
previous_latents_and_names = []
for x in _previous_latents_and_names:
previous_latents_and_names.append(x[0])
previous_latents_and_names.append(x[1])
for i, data_i in enumerate(self.data):
for k, x in enumerate(data_i):
pyro.observe("obs_%s_%d" % (previous_latents_and_names[i][1], k),
dist.normal, x, previous_latents_and_names[i][0],
torch.pow(self.lambdas[-1], -0.5))
return top_latent
def model(self, xs, ys=None):
"""
The model corresponds to the following generative process:
p(z) = normal(0,I) # handwriting style (latent)
p(y|x) = categorical(I/10.) # which digit (semi-supervised)
p(x|y,z) = bernoulli(mu(y,z)) # an image
mu is given by a neural network `decoder`
:param xs: a batch of scaled vectors of pixels from an image
:param ys: (optional) a batch of the class labels i.e.
the digit corresponding to the image(s)
:return: None
"""
# register this pytorch module and all of its sub-modules with pyro
pyro.module("ss_vae", self)
batch_size = xs.size(0)
with pyro.iarange("independent"):
# sample the handwriting style from the constant prior distribution
prior_mu = Variable(torch.zeros([batch_size, self.z_dim]))
prior_sigma = Variable(torch.ones([batch_size, self.z_dim]))
zs = pyro.sample("z", dist.normal, prior_mu, prior_sigma)
# if the label y (which digit to write) is supervised, sample from the
# constant prior, otherwise, observe the value (i.e. score it against the constant prior)
alpha_prior = Variable(torch.ones([batch_size, self.output_size]) / (1.0 * self.output_size))
if ys is None:
ys = pyro.sample("y", dist.categorical, alpha_prior)
else:
pyro.observe("y", dist.categorical, ys, alpha_prior)
# finally, score the image (x) using the handwriting style (z) and
# the class label y (which digit to write) against the
# parametrized distribution p(x|y,z) = bernoulli(decoder(y,z))
# where `decoder` is a neural network
mu = self.decoder.forward([zs, ys])
pyro.observe("x", dist.bernoulli, xs, mu)
def model():
p_latent = pyro.param("p1", Variable(torch.Tensor([[0.7], [0.3]])))
p_obs = pyro.param("p2", Variable(torch.Tensor([[0.9], [0.1]])))
latents = [Variable(torch.ones(1, 1))]
observes = []
for t in range(self.model_steps):
latents.append(
pyro.sample("latent_{}".format(str(t)),
Bernoulli(torch.index_select(p_latent, 0, latents[-1].view(-1).long()))))
observes.append(
pyro.observe("observe_{}".format(str(t)),
Bernoulli(torch.index_select(p_obs, 0, latents[-1].view(-1).long())),
self.data[t]))
return torch.sum(torch.cat(latents))
def model():
ps = pyro.param("ps", Variable(torch.Tensor([[0.8], [0.3]])))
mu = pyro.param("mu", Variable(torch.Tensor([[-0.1], [0.9]])))
sigma = Variable(torch.ones(1, 1))
latents = [Variable(torch.ones(1))]
observes = []
for t in range(3):
latents.append(
pyro.sample("latent_{}".format(str(t)),
Bernoulli(ps[latents[-1][0].long().data])))
observes.append(
pyro.observe("observe_{}".format(str(t)),
Normal(mu[latents[-1][0].long().data], sigma),
pyro.ones(1)))
return latents
def model():
p_latent = pyro.sample("p_latent", dist.beta, self.alpha0, self.beta0)
for i, x in enumerate(self.data):
pyro.observe("obs_{}".format(i), dist.bernoulli, x,
torch.pow(torch.pow(p_latent, 2.0), 0.5))
return p_latent
def model():
mu = pyro.sample("mu", Normal(Variable(torch.zeros(1)),
Variable(torch.ones(1))))
xd = Normal(mu, Variable(torch.ones(1)), batch_size=50)
pyro.observe("xs", xd, self.data)
return mu
def model():
lambda_latent = pyro.sample("lambda_latent", dist.gamma, self.alpha0, self.beta0)
for i, x in enumerate(self.data):
pyro.observe("obs_{}".format(i), dist.poisson, x, lambda_latent)
return lambda_latent
def model():
lambda_latent = pyro.sample("lambda_latent", dist.gamma, self.alpha0, self.beta0)
pyro.observe("obs0", dist.exponential, self.data[0], lambda_latent)
pyro.observe("obs1", dist.exponential, self.data[1], lambda_latent)
return lambda_latent
def model_step(self, t, n, prev, batch, z_pres_prior_p=default_z_pres_prior_p):
# Sample presence indicators.
if not self.fudge_z_pres:
z_pres_dist = Bernoulli(z_pres_prior_p(t) * prev.z_pres)
else:
z_pres_dist = Uniform(self.ng_zeros(n, 1), self.ng_ones(n, 1))
z_pres = pyro.sample('z_pres_{}'.format(t), z_pres_dist)
# If zero is sampled for a data point, then no more objects
# will be added to its output image. We can't
# straight-forwardly avoid generating further objects, so
# instead we zero out the log_pdf of future choices.
sample_mask = z_pres if self.use_masking else None
# Sample attention window position.
z_where = pyro.sample('z_where_{}'.format(t),
dist.normal,
self.z_where_mu_prior,
self.z_where_sigma_prior,
batch_size=n,
log_pdf_mask=sample_mask)
# Sample latent code for contents of the attention window.
z_what = pyro.sample('z_what_{}'.format(t),
dist.normal,
self.ng_zeros([self.z_what_size]),
self.ng_ones([self.z_what_size]),
batch_size=n,
log_pdf_mask=sample_mask)
# Map latent code to pixel space.
y_att = self.decode(z_what)
# Position/scale attention window within larger image.
y = window_to_image(z_where, self.window_size, self.x_size, y_att)
# Combine the image generated at this step with the image so far.
# (Note that there's no notion of occlusion here. Overlapping
# objects can create pixel intensities > 1.)
x = prev.x + (y * z_pres.view(-1, 1, 1))
if batch is not None:
# Add observations.
# Observations are made as soon as we are done generating
# objects for a data point. This ensures that future
# discrete choices are not included in the ELBO. i.e. The
# corresponding log(q/p) in the objectives will be zero
# since we mask out all future choices and make no further
# observations for data points that are complete.
if not self.use_masking:
observe_mask = 1.0
elif t == (self.num_steps - 1):
observe_mask = prev.z_pres
else:
observe_mask = prev.z_pres - z_pres
if self.use_masking or t == (self.num_steps - 1):
pyro.observe("obs_{}".format(t),
dist.normal,
batch.view(n, -1),
x.view(n, -1),
self.ng_ones(x.view(n, -1).size()) * 0.3,
log_pdf_mask=observe_mask)
return ModelState(x=x, z_pres=z_pres, z_where=z_where)
def obs_inner(i, _i, _x):
pyro.observe("obs_%d_%d" % (i, _i), dist.normal, _x, mu_latent,
torch.pow(self.lam, -0.5))
def model(self, data):
decoder = pyro.module('decoder', self.vae_decoder)
z_mean, z_std = ng_zeros([data.size(0), 20]), ng_ones([data.size(0), 20])
z = pyro.sample('latent', Normal(z_mean, z_std))
img = decoder.forward(z)
pyro.observe('obs', Bernoulli(img), data.view(-1, 784))
def model():
prior_dist = Normal(self.mu0, torch.pow(self.lam0, -0.5))
mu_latent = pyro.sample("mu_latent", prior_dist)
x_dist = Normal(mu_latent, torch.pow(self.lam, -0.5))
pyro.observe("obs", x_dist, self.data)
return mu_latent
def model(subsample_size):
with pyro.iarange("data", len(data), subsample_size) as ind:
x = data[ind]
z = pyro.sample("z", dist.Normal(ng_zeros(len(x)), ng_ones(len(x)),
reparameterized=reparameterized))
pyro.observe("x", dist.Normal(z, ng_ones(len(x)), reparameterized=reparameterized), x)
def model_obs_dup():
pyro.sample("mu_q", dist.normal, ng_zeros(1), ng_ones(1))
pyro.observe("mu_q", dist.normal, ng_zeros(1), ng_ones(1), ng_zeros(1))
def model():
lambda_latent = pyro.sample("lambda_latent", dist.gamma, self.alpha0, self.beta0)
pyro.map_data("aaa",
self.data, lambda i, x: pyro.observe(
"obs_{}".format(i), dist.poisson, x, lambda_latent), batch_size=3)
return lambda_latent
