-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
177 lines (157 loc) · 5.96 KB
/
eval.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
import logging, datetime, os, sys, traceback, re, argparse
import numpyro
import jax
from stannumpyro.dppl import NumPyroModel
from numpyro.infer import Trace_ELBO
from numpyro.optim import Adam
import numpyro.infer.autoguide as autoguide
from utils import (
compile_model,
get_posterior,
summary,
golds,
)
from cmdstanpy import CmdStanModel
import pandas as pd
import numpy
logger = logging.getLogger(__name__)
def run_advi(*, posterior, mode, num_steps, num_samples):
model = posterior.model
data = posterior.data.values()
stanfile = model.code_file_path("stan")
sm = CmdStanModel(stan_file=stanfile)
fit = sm.variational(
iter=num_steps, algorithm=mode, output_samples=num_samples, data=data
)
return fit
def run_svi(*, posterior, backend, mode, Autoguide, num_steps, num_samples):
"""
Compile and run the model.
Returns the summary Dataframe
"""
model = posterior.model
data = posterior.data.values()
stanfile = model.code_file_path("stan")
build_dir = f"_build_{backend}_{mode}"
numpyro_model = NumPyroModel(stanfile, recompile=False, build_dir=build_dir)
optim = Adam(step_size=0.0005)
loss = Trace_ELBO()
guide = Autoguide(numpyro_model.get_model())
svi = numpyro_model.svi(optim, loss, guide)
svi.run(jax.random.PRNGKey(0), data, num_steps=num_steps, num_samples=num_samples)
return svi
def compare(*, posterior, backend, mode, Autoguide, num_steps, num_samples, logfile):
"""
Compare gold standard with model.
"""
logger.info(f"Processing {posterior.name}")
sg = summary(posterior.reference_draws())
if backend == "stan":
fit = run_advi(
posterior=posterior, mode=mode, num_steps=num_steps, num_samples=num_samples
)
samples = {
k: numpy.array(fit.variational_sample[i])
for i, k in enumerate(fit.column_names)
}
sm = summary(samples)
sm = sm[~sm.index.str.endswith("__")]
sm = sm.rename(columns={"Mean": "mean", "StdDev": "std", "N_Eff": "n_eff"})
else:
svi = run_svi(
posterior=posterior,
backend=backend,
mode=mode,
Autoguide=Autoguide,
num_steps=num_steps,
num_samples=num_samples,
)
sm = svi.summary()
if not set(sg.index).issubset(set(sm.index)):
raise RuntimeError("Missing parameter")
# perf_cmdstan condition: err > 0.0001 and (err / stdev) > 0.3
sm = sm.loc[sg.index]
sm = sm[["mean", "std", "n_eff"]]
sm["err"] = abs(sm["mean"] - sg["mean"])
sm["rel_err"] = sm["err"] / sg["std"]
comp = sm[(sm["err"] > 0.0001) & (sm["rel_err"] > 0.3)].dropna()
if not comp.empty:
logger.error(f"Failed {posterior.name}")
print(f"{name},mismatch,{sm['rel_err'].max(skipna=False)},{sm['n_eff'].mean(skipna=False)}", file=logfile, flush=True)
else:
logger.info(f"Success {posterior.name}")
print(f"{name},success,{sm['rel_err'].max(skipna=False)},{sm['n_eff'].mean(skipna=False)}", file=logfile, flush=True)
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Run autoguide accuracy experiment on PosteriorDB models."
)
parser.add_argument(
"--backend",
help="inference backend (numpyro, or stan)",
required=True,
)
parser.add_argument(
"--mode",
help="compilation mode for NumPyro (generative, comprehensive, mixed), algo for Stan (fullrank or meanfield)",
required=True,
)
parser.add_argument(
"--test",
help="run test experiment (steps = 100, samples = 100)",
action="store_true",
)
parser.add_argument(
"--posteriors", nargs="+", help="select the examples to execute"
)
parser.add_argument(
"--guide",
help="autoguide (http://num.pyro.ai/en/latest/autoguide.html)",
default="AutoNormal",
)
# Override posteriorDB configs
parser.add_argument("--steps", type=int, help="number of svi steps")
parser.add_argument("--samples", type=int, help="number of samples")
args = parser.parse_args()
if args.posteriors:
assert all(p in golds for p in args.posteriors), "Bad posterior name"
golds = args.posteriors
logging.basicConfig(level=logging.INFO)
numpyro.set_host_device_count(20)
if not os.path.exists("logs"):
os.makedirs("logs")
today = datetime.datetime.now()
logpath = f"logs/status_svi_{args.backend}_{args.mode}"
if args.backend != "stan":
logpath += f"_{args.guide}"
logpath += f"_{today.strftime('%y%m%d_%H%M%S')}.csv"
with open(logpath, "a") as logfile:
print(",status,rel_err,n_eff,exception", file=logfile, flush=True)
for name in (n for n in golds):
# Configurations
posterior = get_posterior(name)
if args.test:
args.steps = 100
args.samples = 100
if args.steps is None:
args.steps = 100000
if args.samples is None:
args.samples = posterior.reference_draws_info()["diagnostics"]["ndraws"]
try:
# Compile
compile_model(posterior=posterior, backend=args.backend, mode=args.mode)
# Run and Compare
compare(
posterior=posterior,
backend=args.backend,
mode=args.mode,
Autoguide=getattr(autoguide, args.guide),
num_steps=args.steps,
num_samples=args.samples,
logfile=logfile,
)
except:
exc_type, exc_value, _ = sys.exc_info()
err = " ".join(traceback.format_exception_only(exc_type, exc_value))
err = re.sub(r"[\n\r\",]", " ", err)[:150] + "..."
logger.error(f"Failed {name} with {err}")
print(f'{name},error,,,"{err}"', file=logfile, flush=True)