mean = torch.matmul(design[..., :-2], w1.unsqueeze(-1)).squeeze(-1)
offset_mean = mean + random_effect_offset
base_dist = dist.Normal(offset_mean, torch.tensor(1.)).to_event(1)
transforms = [
AffineTransform(loc=torch.tensor(0.), scale=random_effect_k),
SigmoidTransform()
]
response_dist = dist.TransformedDistribution(base_dist, transforms)
y = pyro.sample("y", response_dist)
return y
@pytest.mark.parametrize(
"model1,model2,design",
[(zero_mean_unit_obs_sd_lm(torch.tensor(
[10., 10.]))[0], lm_2p_10_10_1, torch.tensor([[1., -1.]])),
(lm_2p_10_10_1, zero_mean_unit_obs_sd_lm(torch.tensor(
[10., 10.]))[0], torch.tensor([[100., -100.]])),
(group_linear_model(torch.tensor(0.), torch.tensor(
[10.]), torch.tensor(0.), torch.tensor([10.]), torch.tensor(1.)),
lm_2p_10_10_1_w12, torch.tensor([[-1.5, 0.5], [1.5, 0.]])),
(known_covariance_linear_model(torch.tensor(
[1., -1.]), torch.tensor([10., 10.]), torch.tensor(1.)),
nz_lm_2p_10_10_1, torch.tensor([[-1., 0.5], [2.5, -2.]])),
(normal_inverse_gamma_linear_model(torch.tensor(
[1., -1.]), torch.tensor(.1), torch.tensor(2.), torch.tensor(2.)),
normal_inv_gamma_2_2_10_10, torch.tensor([[1., -0.5], [1.5, 2.]])),
(logistic_regression_model(torch.tensor([1., -1.]), torch.tensor(10.)),
lr_10_10, torch.tensor([[6., -1.5], [.5, 0.]])),
(sigmoid_model(torch.tensor([1., -1.]), torch.tensor([10., 10.]),
torch.tensor(0.), torch.tensor([1., 1.]), torch.tensor(1.),
and expensive-to-compute functions in pyro.
[1] ["Bayesian Regression"](http://pyro.ai/examples/bayesian_regression.html)
[2] Long Ouyang, Michael Henry Tessler, Daniel Ly, Noah Goodman (2016),
"Practical optimal experiment design with probabilistic programs",
(https://arxiv.org/abs/1608.05046)
[3] ["Bayesian Optimization"](http://pyro.ai/examples/bo.html)
"""
# Set up regression model dimensions
N = 100 # number of participants
p = 2 # number of features
prior_sds = torch.tensor([10., 2.5])
# Model and guide using known obs_sd
model, guide = zero_mean_unit_obs_sd_lm(prior_sds)
def estimated_ape(ns, num_vi_steps):
designs = [
group_assignment_matrix(torch.tensor([n1, N - n1])) for n1 in ns
]
X = torch.stack(designs)
est_ape = vi_eig(
model,
X,
observation_labels="y",
target_labels="w",
vi_parameters={
"guide": guide,
"optim": optim.Adam({"lr": 0.05}),
base_dist = dist.Normal(offset_mean, torch.tensor(1.0)).to_event(1)
transforms = [
AffineTransform(loc=torch.tensor(0.0), scale=random_effect_k),
SigmoidTransform(),
]
response_dist = dist.TransformedDistribution(base_dist, transforms)
y = pyro.sample("y", response_dist)
return y
@pytest.mark.parametrize(
"model1,model2,design",
[
(
zero_mean_unit_obs_sd_lm(torch.tensor([10.0, 10.0]))[0],
lm_2p_10_10_1,
torch.tensor([[1.0, -1.0]]),
),
(
lm_2p_10_10_1,
zero_mean_unit_obs_sd_lm(torch.tensor([10.0, 10.0]))[0],
torch.tensor([[100.0, -100.0]]),
),
(
group_linear_model(
torch.tensor(0.0),
torch.tensor([10.0]),
torch.tensor(0.0),
torch.tensor([10.0]),
torch.tensor(1.0),
