Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

allow for batched inference with observational SIR model, add test for batched inference #566

Open
wants to merge 3 commits into
base: master
Choose a base branch
from
Open

allow for batched inference with observational SIR model, add test for batched inference #566

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
029e157
added test and revised the notebook
rfl-urbaniak Aug 26, 2024
601ee23
format, lint
rfl-urbaniak Aug 26, 2024
02aeefa
linting
rfl-urbaniak Aug 27, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
162 changes: 68 additions & 94 deletions docs/source/dynamical_intro.ipynb
View file
Open in desktop

Large diffs are not rendered by default.

398 changes: 398 additions & 0 deletions tests/dynamical/test_batched_inference.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,398 @@
from typing import Dict

import pyro
import pyro.distributions as dist
import torch
from pyro.infer.autoguide import AutoMultivariateNormal

from chirho.dynamical.handlers import StaticBatchObservation
from chirho.dynamical.handlers.solver import TorchDiffEq
from chirho.dynamical.ops import State, simulate
from chirho.observational.handlers import condition

seed = 123

num_steps = 10
num_samples = 10


class CannibalisticDynamics(pyro.nn.PyroModule):
def __init__(self, b, ml, ma, cea, cel, cpa):
super().__init__()
self.b = b
self.ml = ml
self.ma = ma
self.cea = cea # adult cannibalism of eggs
self.cel = cel # larval cannibalism of eggs
self.cpa = cpa # adult cannibalism of pupae

def forward(self, X: State[torch.Tensor]):
dX = dict()
dX["L"] = self.b * X["A"] * torch.exp(-self.cel * X["L"] - self.cea * X["A"])
dX["P"] = (1 - self.ml) * X["L"]
dX["A"] = X["P"] * torch.exp(-self.cpa * X["A"]) + (1 - self.ma) * X["A"]
return dX


def beetle_observational(X: State[torch.Tensor]) -> None:
event_dim = 1 if X["L"].shape and X["L"].shape[-1] > 1 else 0

# test will fail without the plating here
n = X["L"].shape[-2] if len(X["L"].shape) >= 2 else 1
with pyro.plate("data", n, dim=-2):
pyro.sample("L_obs", dist.Poisson(X["L"]).to_event(event_dim))
pyro.sample("P_obs", dist.Poisson(X["P"]).to_event(event_dim))
pyro.sample("A_obs", dist.Poisson(X["A"]).to_event(event_dim))


init_state = dict(L=torch.tensor(70.0), P=torch.tensor(35.0), A=torch.tensor(64.0))
start_time = torch.tensor(0.0)
end_time = torch.tensor(18.05)
step_size = torch.tensor(0.1)
obs_step_size = torch.tensor(1.0)
obs_logging_times = torch.arange(start_time, end_time, obs_step_size)


def bayesian_cannibalistic():
b = pyro.sample("b", dist.Uniform(0, 40))
ml = pyro.sample("ml", dist.Uniform(0, 1))
ma = pyro.sample("ma", dist.Uniform(0, 1))
cea = pyro.sample("cea", dist.Uniform(0, 0.1))
cel = pyro.sample("cel", dist.Uniform(0, 0.1))
cpa = pyro.sample("cpa", dist.Uniform(0, 0.1))
dynamics = CannibalisticDynamics(b, ml, ma, cea, cel, cpa)
return dynamics


combined_desharnais = {
"L_obs": torch.tensor(
[
[
70.0,
263.0,
75.0,
125.0,
203.0,
57.0,
182.0,
27.0,
265.0,
32.0,
309.0,
8.0,
360.0,
24.0,
357.0,
13.0,
373.0,
14.0,
404.0,
],
[
70.0,
198.0,
75.0,
111.0,
226.0,
31.0,
246.0,
48.0,
302.0,
35.0,
213.0,
109.0,
178.0,
171.0,
59.0,
299.0,
9.0,
419.0,
3.0,
],
[
70.0,
176.0,
87.0,
96.0,
180.0,
13.0,
222.0,
125.0,
146.0,
101.0,
124.0,
156.0,
69.0,
164.0,
80.0,
187.0,
69.0,
293.0,
42.0,
],
[
70.0,
249.0,
28.0,
181.0,
173.0,
76.0,
254.0,
29.0,
286.0,
8.0,
411.0,
28.0,
308.0,
52.0,
213.0,
114.0,
92.0,
217.0,
75.0,
],
]
),
"P_obs": torch.tensor(
[
[
35.0,
4.0,
109.0,
28.0,
77.0,
71.0,
36.0,
136.0,
35.0,
76.0,
28.0,
252.0,
5.0,
236.0,
20.0,
176.0,
7.0,
189.0,
12.0,
],
[
35.0,
4.0,
77.0,
18.0,
40.0,
67.0,
11.0,
127.0,
8.0,
154.0,
20.0,
156.0,
48.0,
141.0,
73.0,
75.0,
114.0,
54.0,
157.0,
],
[
35.0,
12.0,
71.0,
44.0,
31.0,
72.0,
7.0,
132.0,
59.0,
125.0,
49.0,
82.0,
87.0,
38.0,
99.0,
47.0,
107.0,
38.0,
121.0,
],
[
35.0,
12.0,
100.0,
18.0,
61.0,
47.0,
36.0,
119.0,
27.0,
62.0,
5.0,
232.0,
6.0,
193.0,
12.0,
130.0,
52.0,
81.0,
73.0,
],
]
),
"A_obs": torch.tensor(
[
[
64.0,
78.0,
78.0,
77.0,
61.0,
85.0,
102.0,
104.0,
120.0,
122.0,
132.0,
120.0,
113.0,
97.0,
136.0,
122.0,
117.0,
105.0,
120.0,
],
[
64.0,
88.0,
84.0,
80.0,
69.0,
77.0,
98.0,
88.0,
100.0,
90.0,
120.0,
107.0,
115.0,
119.0,
121.0,
127.0,
117.0,
121.0,
113.0,
],
[
64.0,
77.0,
79.0,
76.0,
61.0,
46.0,
72.0,
69.0,
105.0,
106.0,
99.0,
96.0,
98.0,
95.0,
94.0,
108.0,
98.0,
106.0,
88.0,
],
[
64.0,
86.0,
80.0,
88.0,
80.0,
75.0,
76.0,
77.0,
110.0,
106.0,
108.0,
99.0,
120.0,
93.0,
132.0,
115.0,
134.0,
117.0,
134.0,
],
]
),
}


def get_initial_states(
states: Dict[str, Dict[str, torch.Tensor]]
) -> Dict[str, torch.Tensor]:
initial_states = {}
for key in states:
initial_states[key[0]] = states[key][:, 0]
return initial_states


desharnais_initial_states = get_initial_states(combined_desharnais)


def conditioned_cannibalistic(data):
dynamics = bayesian_cannibalistic()
obs = condition(data=data)(beetle_observational)

with TorchDiffEq(), StaticBatchObservation(obs_logging_times, observation=obs):
simulate(dynamics, desharnais_initial_states, start_time, obs_logging_times[-1])


with pyro.poutine.trace() as tr:
conditioned_cannibalistic(combined_desharnais)


def run_svi_inference(
model,
num_steps=num_steps,
verbose=True,
lr=0.03,
vi_family=AutoMultivariateNormal,
guide=None,
obs_n=1,
**model_kwargs
):
if guide is None:
guide = vi_family(model)
elbo = pyro.infer.Trace_ELBO()(model, guide)
elbo(**model_kwargs)
adam = torch.optim.Adam(elbo.parameters(), lr=lr)
for step in range(1, num_steps + 1):
adam.zero_grad()
loss = elbo(**model_kwargs)
loss.backward()
adam.step()
if (step % 100 == 0) or (step == 1) & verbose:
print(
"[iteration %04d] loss: %.4f" % (step, loss),
"avg loss: ",
round(loss.item() / obs_n),
)

print("inference_complete")
return guide


def test_batched_inference():
pyro.clear_param_store()

run_svi_inference(
conditioned_cannibalistic,
num_steps=num_steps,
obs_n=len(obs_logging_times),
data=combined_desharnais,
)
Loading