def gen_default_priors(self, data, K, L,
# sig_prior=Gamma(10, 10),
sig_prior=LogNormal(0, 0.01),
alpha_prior=Gamma(1, 1),
mu0_prior=None,
mu1_prior=None,
W_prior=None,
eta0_prior=None,
eta1_prior=None):
if L is None:
L = [5, 5]
self.__cache_model_constants__(data, K, L)
if mu0_prior is None:
mu0_prior = Gamma(1, 1)
if mu1_prior is None:
mu1_prior = Gamma(1, 1)
if W_prior is None:
W_prior = Dirichlet(torch.ones(self.K) / self.K)
if eta0_prior is None:
eta0_prior = Dirichlet(torch.ones(self.L[0]) / self.L[0])
if eta1_prior is None:
eta1_prior = Dirichlet(torch.ones(self.L[1]) / self.L[1])
self.priors = {'mu0': mu0_prior, 'mu1': mu1_prior, 'sig': sig_prior,
'eta0': eta0_prior, 'eta1': eta1_prior, 'W': W_prior,
'alpha': alpha_prior}
def test_dirichlet_shape(self):
alpha = Variable(torch.exp(torch.randn(2, 3)), requires_grad=True)
alpha_1d = Variable(torch.exp(torch.randn(4)), requires_grad=True)
self.assertEqual(Dirichlet(alpha).sample().size(), (2, 3))
self.assertEqual(Dirichlet(alpha).sample((5, )).size(), (5, 2, 3))
self.assertEqual(Dirichlet(alpha_1d).sample().size(), (4, ))
self.assertEqual(Dirichlet(alpha_1d).sample((1, )).size(), (1, 4))
def log_prior(Z, variable_types):
if 'pi_unconstrained' in variable_types.keys():
pi = softmax(Z['pi_unconstrained'][0], dim=-1)
logp = Dirichlet(
torch.ones_like(pi)).log_prob(pi) + log_det_jacobian_softmax(
pi, dim=-1)
else:
logp = 0
for i, (key, z) in enumerate(Z.items()):
z = z[0]
if key != 'pi_unconstrained':
if variable_types[key] == 'Categorical':
alpha = softmax(z, dim=-1, additional=-50.)
logp += torch.sum(Dirichlet(torch.ones_like(alpha)).log_prob(alpha) \
+ log_det_jacobian_softmax(alpha, dim=-1), dim=-1)
#elif variable_types[key] == 'Bernoulli':
# theta = torch.sigmoid(z)
# logp += torch.sum(Beta(torch.ones_like(theta), torch.ones_like(theta)).log_prob(theta)\
# + log_det_jacobian_sigmoid(theta), dim=-1)
elif variable_types[key] == 'Bernoulli':
logp += TransBeta.log_prob(z).sum()
elif variable_types[key] == 'Beta':
alpha, beta = torch.exp(z)
logp += torch.sum(Gamma(1.0, 1.0).log_prob(alpha) +
torch.log(alpha),
dim=-1)
logp += torch.sum(Gamma(1.0, 1.0).log_prob(beta) +
torch.log(beta),
dim=-1)
return torch.mean(logp)
def log_prior(Z, variable_types):
"""
Z : A dictionary containing the draws from variational posterior for each parameter.
variable_types : a dictionary that contains distribution name assigned to each parameter.
"""
## We proceed similarly as in log-likelihood computation, however since the elements of
## Z are in expanded form and the prior is not data dependent we compute the contribution
## of only the first element of element of Z.
pi = softmax(Z['pi_unconstrained'][0], dim=-1)
logp = Dirichlet(
torch.ones_like(pi)).log_prob(pi) + log_det_jacobian_softmax(pi,
dim=-1)
for i, (key, z) in enumerate(Z.items()):
if key != 'pi_unconstrained':
z = z[0]
if variable_types[key] == 'Categorical':
alpha = softmax(z, dim=-1, additional=-50.)
logp += torch.sum(Dirichlet(torch.ones_like(alpha)).log_prob(alpha) \
+ log_det_jacobian_softmax(alpha, dim=-1), dim=-1)
elif variable_types[key] == 'Bernoulli':
theta = torch.sigmoid(z)
logp += torch.sum(Beta(torch.ones_like(theta), torch.ones_like(theta)).log_prob(theta)\
+ log_det_jacobian_sigmoid(theta), dim=-1)
elif variable_types[key] == 'Beta':
alpha, beta = torch.exp(z)
logp += torch.sum(Gamma(1.0, 1.0).log_prob(alpha) +
torch.log(alpha),
dim=-1)
logp += torch.sum(Gamma(1.0, 1.0).log_prob(beta) +
torch.log(beta),
dim=-1)
return logp
def loss(
self,
tensors,
inference_outputs,
generative_outputs,
n_obs: int = 1.0,
):
# generative_outputs is a dict of the return value from `generative(...)`
# assume that `n_obs` is the number of training data points
p_x_c = generative_outputs["p_x_c"]
gamma = generative_outputs["gamma"]
# compute Q
# take mean of number of cells and multiply by n_obs (instead of summing n)
q_per_cell = torch.sum(gamma * -p_x_c, 1)
# third term is log prob of prior terms in Q
theta_log = F.log_softmax(self.theta_logit, dim=-1)
theta_log_prior = Dirichlet(self.dirichlet_concentration)
theta_log_prob = -theta_log_prior.log_prob(
torch.exp(theta_log) + THETA_LOWER_BOUND)
prior_log_prob = theta_log_prob
delta_log_prior = Normal(self.delta_log_mean,
self.delta_log_log_scale.exp().sqrt())
delta_log_prob = torch.masked_select(
delta_log_prior.log_prob(self.delta_log), (self.rho > 0))
prior_log_prob += -torch.sum(delta_log_prob)
loss = (torch.mean(q_per_cell) * n_obs + prior_log_prob) / n_obs
return LossRecorder(loss, q_per_cell, torch.zeros_like(q_per_cell),
prior_log_prob)
def test_dirichlet_shape(self):
dist = Dirichlet(torch.Tensor([[0.6, 0.3], [1.6, 1.3], [2.6, 2.3]]))
self.assertEqual(dist._batch_shape, torch.Size((3,)))
self.assertEqual(dist._event_shape, torch.Size((2,)))
self.assertEqual(dist.sample().size(), torch.Size((3, 2)))
self.assertEqual(dist.sample((5, 4)).size(), torch.Size((5, 4, 3, 2)))
self.assertEqual(dist.log_prob(self.tensor_sample_1).size(), torch.Size((3,)))
self.assertRaises(ValueError, dist.log_prob, self.tensor_sample_2)
def select_action(self, obs):
concentration, value = self.forward(obs)
m = Dirichlet(concentration)
action = m.sample()
self.saved_actions.append(SavedAction(m.log_prob(action), value))
return list(action.cpu().numpy())
def test_forward(self):
source_dirichlet = Dirichlet(torch.ones(10))
batch_size = 12
observed_data = source_dirichlet.sample((batch_size, ))
# Check that forward function returns
_, membership_probs = self.dmm(observed_data)
# Ensure membership probabilities sum to one
for prob in membership_probs.sum(dim=1):
self.assertAlmostEqual(prob.item(), 1, places=5)
def test_dirichlet_log_prob(self):
num_samples = 10
alpha = torch.exp(torch.randn(5))
dist = Dirichlet(alpha)
x = dist.sample((num_samples,))
actual_log_prob = dist.log_prob(x)
for i in range(num_samples):
expected_log_prob = scipy.stats.dirichlet.logpdf(x[i].numpy(), alpha.numpy())
self.assertAlmostEqual(actual_log_prob[i], expected_log_prob, places=3)
def sample_partitions(B, N, K, alpha=1.0, rand_K=True, device='cpu'):
pi = Dirichlet(alpha * torch.ones(K)).sample([B]).to(device)
if rand_K:
to_use = (torch.rand(B, K) < 0.5).float().to(device)
to_use[..., 0] = 1
pi = pi * to_use
pi = pi / pi.sum(1, keepdim=True)
labels = Categorical(probs=pi).sample([N]).to(device)
labels = labels.transpose(0, 1).contiguous()
return labels
def test_backward(self):
source_dirichlet = Dirichlet(torch.ones(10))
batch_size = 12
observed_data = source_dirichlet.sample((batch_size, ))
# Obtain loss
nll, _ = self.dmm(observed_data)
loss = nll.sum()
# Check that gradient is non-zero for params
loss.backward()
for param in self.dmm.parameters():
self.assertIsNotNone(param.grad)
def sample_labels(B, N, K_low, K_high, alpha=1.0):
pi = Dirichlet(alpha * torch.ones(K_high)).sample([B])
K = torch.randint(K_low, K_high + 1, size=(B, ))
to_use = torch.zeros(B, K_high).int()
for i, k in enumerate(K):
to_use[i, :k] = 1
pi = pi * to_use
pi = pi / pi.sum(1, keepdim=True)
labels = Categorical(probs=pi).sample([N])
labels = labels.transpose(0, 1).contiguous()
return labels
def forward(self, state):
concentration = self._get_concentration(state)
if self.training:
# PyTorch can't backwards pass _sample_dirichlet
action = Dirichlet(concentration).rsample()
else:
# ONNX can't export Dirichlet()
action = torch._sample_dirichlet(concentration)
log_prob = Dirichlet(concentration).log_prob(action)
return rlt.ActorOutput(action=action,
log_prob=log_prob.unsqueeze(dim=1))
def sample_l1_sphere(device, shape):
'''Sample uniformly from the unit l1 sphere, i.e. the cross polytope.
Inputs:
device: 'cpu' | 'cuda' | other torch devices
shape: a pair (batchsize, dim)
Outputs:
matrix of shape `shape` such that each row is a sample.
'''
batchsize, dim = shape
dirdist = Dirichlet(concentration=torch.ones(dim, device=device))
noises = dirdist.sample([batchsize])
signs = torch.sign(torch.rand_like(noises) - 0.5)
return noises * signs
def test_dirichlet_sample(self):
self._set_rng_seed()
alpha = torch.exp(torch.randn(3))
self._check_sampler_sampler(Dirichlet(alpha),
scipy.stats.dirichlet(alpha.numpy()),
'Dirichlet(alpha={})'.format(list(alpha)),
multivariate=True)
def test_TwoFactorFractionalSIR(self):
dist = Dirichlet(torch.tensor([10000., 1., 1.]))
sir = m.TwoFactorSIR((0.1, 0.05, 0.05, 0.05), dist, dt=1e-1)
x = sir.sample_path(1000)
self.assertEqual(x.shape, torch.Size([1000, 3]))
def _AMN_optimization_ENS_mixture(AMN_net,
expert_net,
optimizer,
state_batch,
feature_regression=False,
tau=0.1,
beta=0.01,
GAMMA=0.99,
training=True):
"""
Apply the standard procedure to deep Q network.
"""
if not training:
return None
AMN_policy = 0
alpha = Dirichlet(torch.ones(AMN_net.get_num_ensembles())).sample()
for i in range(AMN_net.get_num_ensembles()):
AMN_q_value = AMN_net(state_batch, ens_num=i, last_layer=False)
AMN_policy += alpha[i] * to_policy(AMN_q_value)
loss = 0
expert_q_value = expert_net(state_batch, last_layer=False)
expert_policy = to_policy(expert_q_value).detach()
loss -= torch.sum(expert_policy * torch.log(AMN_policy + 1e-8))
optimizer.zero_grad()
loss.backward()
optimizer.step()
return loss.detach()
def gen_default_priors(self,
K,
L,
sig_prior=LogNormal(0, 1),
alpha_prior=Gamma(1., 1.),
mu0_prior=None,
mu1_prior=None,
W_prior=None,
eta0_prior=None,
eta1_prior=None):
if L is None:
L = [5, 5]
self.L = L
if K is None:
K = 30
self.K = K
# FIXME: these should be an ordered prior of TruncatedNormals
if mu0_prior is None:
mu0_prior = Gamma(1, 1)
if mu1_prior is None:
mu1_prior = Gamma(1, 1)
if W_prior is None:
W_prior = Dirichlet(torch.ones(self.K) / self.K)
if eta0_prior is None:
eta0_prior = Dirichlet(torch.ones(self.L[0]) / self.L[0])
if eta1_prior is None:
eta1_prior = Dirichlet(torch.ones(self.L[1]) / self.L[1])
self.priors = {
'mu0': mu0_prior,
'mu1': mu1_prior,
'sig': sig_prior,
'H': Normal(0, 1),
'eta0': eta0_prior,
'eta1': eta1_prior,
'W': W_prior,
'alpha': alpha_prior
}
def kl_divergence(model_concentrations, target_concentrations, mode='reverse'):
"""
Input: Model concentrations, target concentrations parameters.
Output: Average of the KL between the two Dirichlet.
"""
assert torch.all(model_concentrations > 0)
assert torch.all(target_concentrations > 0)
target_dirichlet = Dirichlet(target_concentrations)
model_dirichlet = Dirichlet(model_concentrations)
kl_divergences = _kl_dirichlet_dirichlet(
p=target_dirichlet if mode == 'forward' else model_dirichlet,
q=model_dirichlet if mode == 'forward' else target_dirichlet)
assert_no_nan_no_inf(kl_divergences)
mean_kl = torch.mean(kl_divergences)
assert_no_nan_no_inf(mean_kl)
return mean_kl
def backward(ctx, grad_output):
grad_input=None
if not ctx.train:
raise RuntimeError('Running backward on shake when train is False')
if ctx.needs_input_grad[0]:
dist = Dirichlet(th.full((ctx.xsh[ctx.dim],),ctx.concentration))
beta = dist.sample(sample_shape=th.Size([ctx.xsh[ctx.batchdim]]))
beta = beta.to(th.device(ctx.dev))
sh = [1 for _ in range(len(ctx.xsh))]
sh[ctx.batchdim], sh[ctx.dim] = ctx.xsh[ctx.batchdim], ctx.xsh[ctx.dim]
beta = beta.view(*sh)
grad_output = grad_output.unsqueeze(ctx.dim).expand(*ctx.xsh)
grad_input = grad_output * beta
return grad_input, None, None, None, None, None
def test_TwoFactorSIRD(self):
dist = Dirichlet(torch.tensor([10000., 1., 1., 1.]))
sird = m.ThreeFactorSIRD((0.1, 0.05, 0.05, 0.01, 0.05, 0.05, 0.05),
dist,
dt=1e-1)
x = sird.sample_path(1000)
self.assertEqual(x.shape, torch.Size([1000, 4]))
def sample(self, B, N, K, return_gt=False):
device = 'cpu' if not torch.cuda.is_available() \
else torch.cuda.current_device()
pi = Dirichlet(torch.ones(K)).sample(torch.Size([B])).to(device)
labels = Categorical(probs=pi).sample(torch.Size([N])).to(device)
labels = labels.transpose(0, 1).contiguous()
X, params = self.mvn.sample(B, K, labels)
if return_gt:
return X, labels, pi, params
else:
return X
def construct_variational_parameters(T, N):
# T-1 params for the second part of variational Beta factors
kappa = Variable(Uniform(0, 2).rsample([T - 1]), requires_grad=True)
# T scale params for the variational Gamma factors
tau_0 = Uniform(0, 100).rsample([T])
# T rate params for the variational Gamma factors
tau_1 = LogNormal(0, 1).rsample([T])
tau = Variable(torch.stack((tau_0, tau_1)).T, requires_grad=True)
phi = Variable(
Dirichlet(1 / T * torch.ones(T)).rsample([N]),
requires_grad=True) # N,T params for the variational Cat factors
return kappa, tau, phi
def log_prior(theta_unconstrained, pi_unconstrained):
theta = torch.sigmoid(theta_unconstrained)
pi = softmax(pi_unconstrained, dim=-1)
""" Both tau and pi are transformed, so we need to add correction terms
to log densities. For tau this will be theta_unconstrained (log |d/dy exp(y)| = y)
and for pi this will be log_det_jacobian_softmax (for reasons too
complicated to be explained here).
"""
theta_logp = torch.sum(log_det_jacobian_sigmoid(theta_unconstrained))
pi_logp = torch.sum(
Dirichlet(torch.ones_like(pi)).log_prob(pi) +
log_det_jacobian_softmax(pi, dim=-1))
return pi_logp + theta_logp
def forward(ctx, x, dim, batchdim, concentration, train):
ctx.dim = dim
ctx.batchdim = batchdim
ctx.concentration = concentration
xsh, ctx.xsh = [x.shape]*2
ctx.dev = x.device
ctx.train = True
if train:
# Randomly sample from Dirichlet distribution
dist = Dirichlet(th.full((xsh[dim],), concentration))
alpha = dist.sample(sample_shape=th.Size([xsh[batchdim]]))
alpha = alpha.to(th.device(x.device))
sh = [1 for _ in range(len(xsh))]
sh[batchdim], sh[dim] = xsh[batchdim], xsh[dim]
alpha = alpha.view(*sh)
y = (x * alpha).sum(dim)
else:
y = x.mean(dim)
return y
def entropy(alpha, uncertainty_type, n_bins=10, plot=True):
entropy = []
if uncertainty_type == 'aleatoric':
p = torch.nn.functional.normalize(alpha, p=1, dim=-1)
entropy.append(
Categorical(p).entropy().squeeze().cpu().detach().numpy())
elif uncertainty_type == 'epistemic':
entropy.append(
Dirichlet(alpha).entropy().squeeze().cpu().detach().numpy())
if plot:
plt.hist(entropy, n_bins)
plt.show()
return entropy
def test_compose_affine(event_dims):
transforms = [AffineTransform(torch.zeros((1,) * e), 1, event_dim=e) for e in event_dims]
transform = ComposeTransform(transforms)
assert transform.codomain.event_dim == max(event_dims)
assert transform.domain.event_dim == max(event_dims)
base_dist = Normal(0, 1)
if transform.domain.event_dim:
base_dist = base_dist.expand((1,) * transform.domain.event_dim)
dist = TransformedDistribution(base_dist, transform.parts)
assert dist.support.event_dim == max(event_dims)
base_dist = Dirichlet(torch.ones(5))
if transform.domain.event_dim > 1:
base_dist = base_dist.expand((1,) * (transform.domain.event_dim - 1))
dist = TransformedDistribution(base_dist, transforms)
assert dist.support.event_dim == max(1, max(event_dims))
def main():
args = parse_arguments()
x, c, y, S, antes = load_data(args, 'support2')
print(antes[:5])
print(type(antes))
print(len(antes))
# create model
args['n_features'] = c.shape[-1]
model = create_model(args, antes, S.shape[0])
# use whole dataset for now
S = torch.tensor(S, dtype=torch.float)
c = torch.tensor(c, dtype=torch.float)
n_train = S.shape[0]
print(n_train)
optimizer = optim.SGD(model.parameters(), lr=0.1, momentum=0.9)
# optimizer = optim.Adam(model.parameters(), lr=0.001)
start_time = time.time()
for ep in range(50):
optimizer.zero_grad()
d_soft = model(c, S)
# print(d)
maxes, argmaxes = d_soft.max(dim=-1)
d = [antes[amax] for amax in argmaxes]
print(d)
n_classes = y.shape[-1]
B = torch.zeros((n_train, n_classes, len(d) + 1))
for i, xi in enumerate(x):
for j, lhs in enumerate(d):
if set(lhs).issubset(xi):
B[i, y[i, 0], j] = 1.
break
B[i, y[i, 0], -1] = 1 - B[i, y[i, 0], :-1].sum()
assert B.sum() == S.size()[0]
# get dirichlet prior
alpha = 1.
priors = alpha + B.sum(0)
thetas = torch.zeros((len(d) + 1, n_classes))
for i in range(len(d) + 1):
p_theta = Dirichlet(torch.tensor(priors[:, i]))
thetas[i] = p_theta.rsample()
# compute p(y | d)
log_py = 0
for i, yi in enumerate(y):
for j in range(len(d) + 1):
log_py += B[i, y[i, 0], j] * torch.log(thetas[j, y[i, 0]])
# compute p(d | input), as p(d_1 | input) p(d_2 | d_1, input) ...
log_pd = maxes.sum()
log_prob = -(log_py + log_pd)
elapsed = time.time() - start_time
print(
f"Epoch {ep}: log-prob: {log_prob:.2f}, log p(d|x): {log_pd:.2f}, ",
end='')
print(f"log p(y|d): {log_py:.2f} (Elapsed: {elapsed:.2f}s)")
log_prob.backward()
optimizer.step()
def __init__(self, *args, **kwargs):
super(ThreeGAN, self).__init__(*args, **kwargs)
if self.cls > 0:
raise NotImplementedError("ThreeGAN not implemented for cls > 0")
self.dirichlet = Dirichlet(torch.FloatTensor([1.0, 1.0, 1.0]))
def get_log_prob(self, state, action):
concentration = self._get_concentration(state)
log_prob = Dirichlet(concentration).log_prob(action)
return log_prob.unsqueeze(dim=1)
