inference.py 13.2 KB
Newer Older
Chris Jewell's avatar
Chris Jewell committed
1
"""MCMC Test Rig for COVID-19 UK model"""
2
import optparse
Chris Jewell's avatar
Chris Jewell committed
3
import os
4
from time import perf_counter
Chris Jewell's avatar
Chris Jewell committed
5

6
7
import tqdm
import yaml
Chris Jewell's avatar
Chris Jewell committed
8
import h5py
9
import numpy as np
10
11
import tensorflow as tf
import tensorflow_probability as tfp
Chris Jewell's avatar
Chris Jewell committed
12

13
from covid import config
Chris Jewell's avatar
Chris Jewell committed
14
from covid.impl.util import compute_state
15
from covid.impl.mcmc import UncalibratedLogRandomWalk, random_walk_mvnorm_fn
16
from covid.impl.event_time_mh import UncalibratedEventTimesUpdate
17
from covid.impl.occult_events_mh import UncalibratedOccultUpdate, TransitionTopology
18
19
from covid.impl.gibbs import DeterministicScanKernel, GibbsStep, flatten_results
from covid.impl.multi_scan_kernel import MultiScanKernel
20

21
import model_spec
22

23
# Uncomment to block GPU use
24
# os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
25
26
27
28
29
if tf.test.gpu_device_name():
    print("Using GPU")
else:
    print("Using CPU")

30

31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
tfd = tfp.distributions
tfb = tfp.bijectors
DTYPE = config.floatX


if __name__ == "__main__":

    # Read in settings
    parser = optparse.OptionParser()
    parser.add_option(
        "--config",
        "-c",
        dest="config",
        default="example_config.yaml",
        help="configuration file",
46
    )
47
48
49
50
51
52
53
54
55
56
57
58
59
    options, cmd_args = parser.parse_args()
    print("Loading config file:", options.config)

    with open(options.config, "r") as f:
        config = yaml.load(f, Loader=yaml.FullLoader)

    covar_data = model_spec.read_covariates(config["data"])

    # We load in cases and impute missing infections first, since this sets the
    # time epoch which we are analysing.
    cases = model_spec.read_cases(config["data"]["reported_cases"])

    # Impute censored events, return cases
60
    events = model_spec.impute_censored_events(cases)
61
62
63
64
65
66
67
68
69
70

    # Initial conditions are calculated by calculating the state
    # at the beginning of the inference period
    #
    # Imputed censored events that pre-date the first I-R events
    # in the cases dataset are discarded.  They are only used to
    # to set up a sensible initial state.
    state = compute_state(
        initial_state=tf.concat(
            [covar_data["N"], tf.zeros_like(events[:, 0, :])], axis=-1
71
        ),
72
        events=events,
73
        stoichiometry=model_spec.STOICHIOMETRY,
74
    )
75
76
77
78
79
80
81
82
83
84
85
86
87
    start_time = state.shape[1] - cases.shape[1]
    initial_state = state[:, start_time, :]
    events = events[:, start_time:, :]
    num_metapop = covar_data["N"].shape[0]

    ########################################################
    # Build the model, and then construct the MCMC kernels #
    ########################################################
    model = model_spec.CovidUK(
        covariates=covar_data,
        initial_state=initial_state,
        initial_step=0,
        num_steps=events.shape[1],
88
    )
89

90
91
92
93
    # Full joint log posterior distribution
    # $\pi(\theta, \xi, y^{se}, y^{ei} | y^{ir})$
    def logp(theta, xi, events):
        return model.log_prob(
Chris Jewell's avatar
Chris Jewell committed
94
            dict(beta1=theta[0], beta2=theta[1], gamma=theta[2], xi=xi, seir=events,)
95
        )
96

97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
    # Build Metropolis within Gibbs sampler
    #
    # Kernels are:
    #     Q(\theta, \theta^\prime)
    #     Q(\xi, \xi^\prime)
    #     Q(Z^{se}, Z^{se\prime}) (partially-censored)
    #     Q(Z^{ei}, Z^{ei\prime}) (partially-censored)
    #     Q(Z^{se}, Z^{se\prime}) (occult)
    #     Q(Z^{ei}, Z^{ei\prime}) (occult)
    def make_theta_kernel(scale, bounded_convergence, name):
        return GibbsStep(
            0,
            tfp.mcmc.MetropolisHastings(
                inner_kernel=UncalibratedLogRandomWalk(
                    target_log_prob_fn=logp,
                    new_state_fn=random_walk_mvnorm_fn(scale, p_u=bounded_convergence),
                )
            ),
            name=name,
        )
117

118
119
120
121
122
123
124
125
126
    def make_xi_kernel(scale, bounded_convergence, name):
        return GibbsStep(
            1,
            tfp.mcmc.RandomWalkMetropolis(
                target_log_prob_fn=logp,
                new_state_fn=random_walk_mvnorm_fn(scale, p_u=bounded_convergence),
            ),
            name=name,
        )
Chris Jewell's avatar
Chris Jewell committed
127

128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
    def make_partially_observed_step(
        target_event_id, prev_event_id=None, next_event_id=None, name=None
    ):
        return GibbsStep(
            2,
            tfp.mcmc.MetropolisHastings(
                inner_kernel=UncalibratedEventTimesUpdate(
                    target_log_prob_fn=logp,
                    target_event_id=target_event_id,
                    prev_event_id=prev_event_id,
                    next_event_id=next_event_id,
                    initial_state=initial_state,
                    dmax=config["mcmc"]["dmax"],
                    mmax=config["mcmc"]["m"],
                    nmax=config["mcmc"]["nmax"],
                )
            ),
            name=name,
        )
147

148
149
150
151
152
153
154
155
    def make_occults_step(prev_event_id, target_event_id, next_event_id, name):
        return GibbsStep(
            2,
            tfp.mcmc.MetropolisHastings(
                inner_kernel=UncalibratedOccultUpdate(
                    target_log_prob_fn=logp,
                    topology=TransitionTopology(
                        prev_event_id, target_event_id, next_event_id
156
                    ),
157
158
159
160
161
162
163
                    cumulative_event_offset=initial_state,
                    nmax=config["mcmc"]["occult_nmax"],
                    t_range=(events.shape[1] - 21, events.shape[1]),
                    name=name,
                )
            ),
            name=name,
164
165
        )

166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
    # MCMC tracing functions
    def trace_results_fn(_, results):
        """Returns log_prob, accepted, q_ratio"""

        def is_accepted(result):
            if hasattr(result, "is_accepted"):
                return tf.cast(result.is_accepted, DTYPE)
            return is_accepted(result.inner_results)

        def f(result):
            log_prob = result.proposed_results.target_log_prob
            accepted = is_accepted(result)
            q_ratio = result.proposed_results.log_acceptance_correction
            if hasattr(result.proposed_results, "extra"):
                proposed = tf.cast(result.proposed_results.extra, log_prob.dtype)
                return tf.concat([[log_prob], [accepted], [q_ratio], proposed], axis=0)
            return tf.concat([[log_prob], [accepted], [q_ratio]], axis=0)

        def recurse(f, list_or_atom):
            if isinstance(list_or_atom, list):
                return [recurse(f, x) for x in list_or_atom]
            return f(list_or_atom)

        return recurse(f, results)

    # Build MCMC algorithm here.  This will be run in bursts for memory economy
    @tf.function(autograph=False, experimental_compile=True)
    def sample(n_samples, init_state, sigma_theta, sigma_xi):
        with tf.name_scope("main_mcmc_sample_loop"):

            init_state = init_state.copy()

            kernel = DeterministicScanKernel(
                [
                    make_theta_kernel(sigma_theta, 1.0, "theta_kernel"),
                    make_xi_kernel(sigma_xi, 1.0, "xi_kernel"),
                    MultiScanKernel(
                        config["mcmc"]["num_event_time_updates"],
                        DeterministicScanKernel(
                            [
                                make_partially_observed_step(0, None, 1, "se_events"),
                                make_partially_observed_step(1, 0, 2, "ei_events"),
                                make_occults_step(None, 0, 1, "se_occults"),
                                make_occults_step(0, 1, 2, "ei_occults"),
                            ]
                        ),
                    ),
                ],
                name="gibbs_kernel",
            )
216

217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
            samples, results = tfp.mcmc.sample_chain(
                n_samples, init_state, kernel=kernel, trace_fn=trace_results_fn
            )

            return samples, results

    ####################################
    # Construct bursted MCMC loop here #
    ####################################

    # MCMC Control
    NUM_BURSTS = config["mcmc"]["num_bursts"]
    NUM_BURST_SAMPLES = config["mcmc"]["num_burst_samples"]
    NUM_EVENT_TIME_UPDATES = config["mcmc"]["num_event_time_updates"]
    THIN_BURST_SAMPLES = NUM_BURST_SAMPLES // config["mcmc"]["thin"]
    NUM_SAVED_SAMPLES = THIN_BURST_SAMPLES * NUM_BURSTS

    # RNG stuff
    tf.random.set_seed(2)

    current_state = [
        np.array([0.85, 0.3, 0.25], dtype=DTYPE),
Chris Jewell's avatar
Chris Jewell committed
239
        np.zeros(model.model["xi"]().event_shape[-1], dtype=DTYPE),
240
241
242
243
244
245
246
247
248
249
        events,
    ]

    # Output Files
    posterior = h5py.File(
        os.path.expandvars(config["output"]["posterior"]),
        "w",
        rdcc_nbytes=1024 ** 2 * 400,
        rdcc_nslots=100000,
        libver="latest",
250
    )
251
252
253
    event_size = [NUM_SAVED_SAMPLES] + list(current_state[2].shape)

    posterior.create_dataset("initial_state", data=initial_state)
Chris Jewell's avatar
Chris Jewell committed
254
255
256
257

    # Ideally we insert the inference period into the posterior file
    # as this allows us to post-attribute it to the data.  Maybe better
    # to simply save the data into it as well.
258
259
260
261
262
263
264
    posterior.create_dataset(
        "inference_period", data=[b"2020-06-16", b"2020-09-08"]
    ).attrs["description"] = "inference period [start, end)"
    theta_samples = posterior.create_dataset(
        "samples/theta",
        [NUM_SAVED_SAMPLES, current_state[0].shape[0]],
        dtype=np.float64,
265
    )
266
267
    xi_samples = posterior.create_dataset(
        "samples/xi", [NUM_SAVED_SAMPLES, current_state[1].shape[0]], dtype=np.float64,
268
    )
269
270
271
272
273
274
275
    event_samples = posterior.create_dataset(
        "samples/events",
        event_size,
        dtype=DTYPE,
        chunks=(32, 32, 32, 1),
        compression="szip",
        compression_opts=("nn", 16),
276
    )
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304

    output_results = [
        posterior.create_dataset("results/theta", (NUM_SAVED_SAMPLES, 3), dtype=DTYPE,),
        posterior.create_dataset("results/xi", (NUM_SAVED_SAMPLES, 3), dtype=DTYPE,),
        posterior.create_dataset(
            "results/move/S->E", (NUM_SAVED_SAMPLES, 3 + num_metapop), dtype=DTYPE,
        ),
        posterior.create_dataset(
            "results/move/E->I", (NUM_SAVED_SAMPLES, 3 + num_metapop), dtype=DTYPE,
        ),
        posterior.create_dataset(
            "results/occult/S->E", (NUM_SAVED_SAMPLES, 6), dtype=DTYPE
        ),
        posterior.create_dataset(
            "results/occult/E->I", (NUM_SAVED_SAMPLES, 6), dtype=DTYPE
        ),
    ]
    posterior.swmr_mode = True

    print("Initial logpi:", logp(*current_state))

    theta_scale = tf.constant(
        [
            [1.12e-3, 1.67e-4, 1.61e-4],
            [1.67e-4, 7.41e-4, 4.68e-5],
            [1.61e-4, 4.68e-5, 1.28e-4],
        ],
        dtype=DTYPE,
305
    )
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
    theta_scale = theta_scale * 0.2 / theta_scale.shape[0]

    xi_scale = tf.eye(current_state[1].shape[0], dtype=DTYPE)
    xi_scale = xi_scale * 0.0001 / xi_scale.shape[0]

    # We loop over successive calls to sample because we have to dump results
    #   to disc, or else end OOM (even on a 32GB system).
    # with tf.profiler.experimental.Profile("/tmp/tf_logdir"):
    for i in tqdm.tqdm(range(NUM_BURSTS), unit_scale=NUM_BURST_SAMPLES):
        samples, results = sample(
            NUM_BURST_SAMPLES,
            init_state=current_state,
            sigma_theta=theta_scale,
            sigma_xi=xi_scale,
        )
        current_state = [s[-1] for s in samples]
        s = slice(i * THIN_BURST_SAMPLES, i * THIN_BURST_SAMPLES + THIN_BURST_SAMPLES)
        idx = tf.constant(range(0, NUM_BURST_SAMPLES, config["mcmc"]["thin"]))
        theta_samples[s, ...] = tf.gather(samples[0], idx)
        xi_samples[s, ...] = tf.gather(samples[1], idx)
        # cov = np.cov(
        #     np.log(theta_samples[: (i * NUM_BURST_SAMPLES + NUM_BURST_SAMPLES), ...]),
        #     rowvar=False,
        # )
        print(current_state[0].numpy(), flush=True)
        # print(cov, flush=True)
        # if (i * NUM_BURST_SAMPLES) > 1000 and np.all(np.isfinite(cov)):
        #     theta_scale = 2.38 ** 2 * cov / 2.0

        start = perf_counter()
        event_samples[s, ...] = tf.gather(samples[2], idx)
        end = perf_counter()

        flat_results = flatten_results(results)
        for i, ro in enumerate(output_results):
            ro[s, ...] = tf.gather(flat_results[i], idx)

        posterior.flush()
        print("Storage time:", end - start, "seconds")
        print(
            "Acceptance theta:",
            tf.reduce_mean(tf.cast(flat_results[0][:, 1], tf.float32)),
        )
        print(
            "Acceptance xi:", tf.reduce_mean(tf.cast(flat_results[1][:, 1], tf.float32))
        )
        print(
            "Acceptance move S->E:",
            tf.reduce_mean(tf.cast(flat_results[2][:, 1], tf.float32)),
        )
        print(
            "Acceptance move E->I:",
            tf.reduce_mean(tf.cast(flat_results[3][:, 1], tf.float32)),
        )
        print(
            "Acceptance occult S->E:",
            tf.reduce_mean(tf.cast(flat_results[4][:, 1], tf.float32)),
        )
        print(
            "Acceptance occult E->I:",
            tf.reduce_mean(tf.cast(flat_results[5][:, 1], tf.float32)),
        )
Chris Jewell's avatar
Chris Jewell committed
368

369
370
371
372
373
374
    print(f"Acceptance theta: {output_results[0][:, 1].mean()}")
    print(f"Acceptance xi: {output_results[1][:, 1].mean()}")
    print(f"Acceptance move S->E: {output_results[2][:, 1].mean()}")
    print(f"Acceptance move E->I: {output_results[3][:, 1].mean()}")
    print(f"Acceptance occult S->E: {output_results[4][:, 1].mean()}")
    print(f"Acceptance occult E->I: {output_results[5][:, 1].mean()}")
Chris Jewell's avatar
Chris Jewell committed
375

376
    posterior.close()