def test_em_nested_in_svi(assignment_grad):
args = make_args()
args.assignment_grad = assignment_grad
detections = generate_data(args)
pyro.clear_param_store()
pyro.param('noise_scale',
torch.tensor(args.init_noise_scale),
constraint=constraints.positive)
pyro.param('objects_loc', torch.randn(args.max_num_objects, 1))
# Learn object_loc via EM and noise_scale via SVI.
optim = Adam({'lr': 0.1})
elbo = TraceEnum_ELBO(max_plate_nesting=2)
newton = Newton(trust_radii={'objects_loc': 1.0})
svi = SVI(poutine.block(model, hide=['objects_loc']),
poutine.block(guide, hide=['objects_loc']), optim, elbo)
for svi_step in range(50):
for em_step in range(2):
objects_loc = pyro.param('objects_loc').detach_().requires_grad_()
assert pyro.param('objects_loc').grad_fn is None
loss = elbo.differentiable_loss(model, guide, detections,
args) # E-step
updated = newton.get_step(loss,
{'objects_loc': objects_loc}) # M-step
assert updated['objects_loc'].grad_fn is not None
pyro.get_param_store()['objects_loc'] = updated['objects_loc']
assert pyro.param('objects_loc').grad_fn is not None
loss = svi.step(detections, args)
logger.debug(
'step {: >2d}, loss = {:0.6f}, noise_scale = {:0.6f}'.format(
svi_step, loss,
pyro.param('noise_scale').item()))
def test_svi_multi():
args = make_args()
args.assignment_grad = True
detections = generate_data(args)
pyro.clear_param_store()
pyro.param('noise_scale',
torch.tensor(args.init_noise_scale),
constraint=constraints.positive)
pyro.param('objects_loc', torch.randn(args.max_num_objects, 1))
# Learn object_loc via Newton and noise_scale via Adam.
elbo = TraceEnum_ELBO(max_plate_nesting=2)
adam = Adam({'lr': 0.1})
newton = Newton(trust_radii={'objects_loc': 1.0})
optim = MixedMultiOptimizer([(['noise_scale'], adam),
(['objects_loc'], newton)])
for svi_step in range(50):
with poutine.trace(param_only=True) as param_capture:
loss = elbo.differentiable_loss(model, guide, detections, args)
params = {
name: pyro.param(name).unconstrained()
for name in param_capture.trace.nodes.keys()
}
optim.step(loss, params)
logger.debug(
'step {: >2d}, loss = {:0.6f}, noise_scale = {:0.6f}'.format(
svi_step, loss.item(),
pyro.param('noise_scale').item()))
def test_em(assignment_grad):
args = make_args()
args.assignment_grad = assignment_grad
detections = generate_data(args)
pyro.clear_param_store()
pyro.param('noise_scale', torch.tensor(args.init_noise_scale),
constraint=constraints.positive)
pyro.param('objects_loc', torch.randn(args.max_num_objects, 1))
# Learn object_loc via EM algorithm.
elbo = TraceEnum_ELBO(max_plate_nesting=2)
newton = Newton(trust_radii={'objects_loc': 1.0})
for step in range(10):
# Detach previous iterations.
objects_loc = pyro.param('objects_loc').detach_().requires_grad_()
loss = elbo.differentiable_loss(model, guide, detections, args) # E-step
newton.step(loss, {'objects_loc': objects_loc}) # M-step
logger.debug('step {}, loss = {}'.format(step, loss.item()))
import pytest
import torch
from torch.distributions import constraints
import pyro
import pyro.distributions as dist
import pyro.optim
import pyro.poutine as poutine
from pyro.optim.multi import MixedMultiOptimizer, Newton, PyroMultiOptimizer, TorchMultiOptimizer
from tests.common import assert_equal
FACTORIES = [
lambda: PyroMultiOptimizer(pyro.optim.Adam({'lr': 0.05})),
lambda: TorchMultiOptimizer(torch.optim.Adam, {'lr': 0.05}),
lambda: Newton(trust_radii={'z': 0.2}),
lambda: MixedMultiOptimizer(
[(['y'], PyroMultiOptimizer(pyro.optim.Adam({'lr': 0.05}))),
(['x', 'z'], Newton())]),
lambda: MixedMultiOptimizer([(['y'], pyro.optim.Adam({'lr': 0.05})),
(['x', 'z'], Newton())]),
]
@pytest.mark.parametrize('factory', FACTORIES)
def test_optimizers(factory):
optim = factory()
def model(loc, cov):
x = pyro.param("x", torch.randn(2))
y = pyro.param("y", torch.randn(3, 2))
import pyro
import pyro.distributions as dist
import pyro.optim
import pyro.poutine as poutine
from pyro.optim.multi import (
MixedMultiOptimizer,
Newton,
PyroMultiOptimizer,
TorchMultiOptimizer,
)
from tests.common import assert_equal
FACTORIES = [
lambda: PyroMultiOptimizer(pyro.optim.Adam({"lr": 0.05})),
lambda: TorchMultiOptimizer(torch.optim.Adam, {"lr": 0.05}),
lambda: Newton(trust_radii={"z": 0.2}),
lambda: MixedMultiOptimizer([
(["y"], PyroMultiOptimizer(pyro.optim.Adam({"lr": 0.05}))),
(["x", "z"], Newton()),
]),
lambda: MixedMultiOptimizer([(["y"], pyro.optim.Adam({"lr": 0.05})),
(["x", "z"], Newton())]),
]
@pytest.mark.parametrize("factory", FACTORIES)
def test_optimizers(factory):
optim = factory()
def model(loc, cov):
x = pyro.param("x", torch.randn(2))