def model():
latent = named.Object("latent")
latent.list = named.List()
loc = latent.list.add().loc.param_(torch.zeros(1))
latent.dict = named.Dict()
foo = latent.dict["foo"].foo.sample_(dist.Normal(loc, torch.ones(1)))
latent.object.bar.sample_(dist.Normal(loc, torch.ones(1)), obs=foo)
def model(data, k):
latent = named.Object("latent")
# Create parameters for a Gaussian mixture model.
latent.probs.param_(torch.ones(k) / k, constraint=constraints.simplex)
latent.locs.param_(torch.zeros(k))
latent.scales.param_(torch.ones(k), constraint=constraints.positive)
# Observe all the data. We pass a local latent in to the local_model.
latent.local = named.List()
for x in data:
local_model(latent.local.add(), latent.probs, latent.locs, latent.scales, obs=x)
def guide_recurse(data, latent):
latent.post_loc.param_(torch.tensor(0.0))
latent.post_scale.param_(torch.tensor(1.0), constraint=constraints.positive)
latent.z.sample_(dist.Normal(latent.post_loc, latent.post_scale))
if torch.is_tensor(data):
pass
elif isinstance(data, list):
latent.list = named.List()
for datum in data:
guide_recurse(datum, latent.list.add())
elif isinstance(data, dict):
latent.dict = named.Dict()
for key, value in data.items():
guide_recurse(value, latent.dict[key])
else:
raise TypeError("Unsupported type {}".format(type(data)))
def test_eval_str():
state = named.Object("state")
state.x = 0
state.ys = named.List()
state.ys.add().foo = 1
state.zs = named.Dict()
state.zs[42].bar = 2
assert state is eval(str(state))
assert state.x is eval(str(state.x))
assert state.ys is eval(str(state.ys))
assert state.ys[0] is eval(str(state.ys[0]))
assert state.ys[0].foo is eval(str(state.ys[0].foo))
assert state.zs is eval(str(state.zs))
assert state.zs[42] is eval(str(state.zs[42]))
assert state.zs[42].bar is eval(str(state.zs[42].bar))
def model_recurse(data, latent):
if torch.is_tensor(data):
latent.x.sample_(dist.Normal(latent.z, 1.0), obs=data)
elif isinstance(data, list):
latent.prior_scale.param_(torch.tensor(1.0), constraint=constraints.positive)
latent.list = named.List()
for data_i in data:
latent_i = latent.list.add()
latent_i.z.sample_(dist.Normal(latent.z, latent.prior_scale))
model_recurse(data_i, latent_i)
elif isinstance(data, dict):
latent.prior_scale.param_(torch.tensor(1.0), constraint=constraints.positive)
latent.dict = named.Dict()
for key, value in data.items():
latent.dict[key].z.sample_(dist.Normal(latent.z, latent.prior_scale))
model_recurse(value, latent.dict[key])
else:
raise TypeError("Unsupported type {}".format(type(data)))
def guide(data, k):
latent = named.Object("latent")
latent.local = named.List()
for x in data:
# We pass a local latent in to the local_guide.
local_guide(latent.local.add(), k)
def model():
latent = named.List("latent")
loc = latent.add().param_(torch.zeros(1))
foo = latent.add().sample_(dist.Normal(loc, torch.ones(1)))
latent.add().sample_(dist.Normal(loc, torch.ones(1)), obs=foo)
latent.add().z.sample_(dist.Normal(loc, torch.ones(1)))
