Commit ⋅ pyro-ppl/pyro

Showing 1 of 2 files from the diff.

Other files ignored by Codecov

docs/source/inference_algos.rst has changed.

@@ -21,11 +21,26 @@

Loading


class Trace_CRPS:
    """
    Posterior predictive CRPS loss.
    Posterior predictive CRPS loss with ``KL(q,p)`` regularization.

    This is a likelihood-free method; no densities are evaluated.
    This is a likelihood-free method, and can be used for likelihoods without
    tractible density functions. CRPS is a robust loss function, and is well
    defined for any distribution with finite absolute moment ``E[|data|]``.

    :param num_particles: The number of particles/samples used to form the
    This requires static model structure, fully reparametrized guide, and
    reparametrized likelihood distributions in the model. Model latent
    distributions may be non-reparametrized.

    Note that in the loss ``CRPS + KL(q,p)``, the ``CRPS`` term has data units
    whereas the ``KL(q,p)`` term has units of nats.  To calibrate these units,
    you can wrap likelihood sites in ``poutine.scale``.

    References
    [1] `Strictly Proper Scoring Rules, Prediction, and Estimation`
    Tilmann Gneiting, Adrian E. Raftery (2007)
    https://www.stat.washington.edu/raftery/Research/PDF/Gneiting2007jasa.pdf

    :param int num_particles: The number of particles/samples used to form the
        gradient estimators. Must be at least 2.
    :param int max_plate_nesting: Optional bound on max number of nested
        :func:`pyro.plate` contexts. This is only required when enumerating

@@ -37,23 +52,25 @@

Loading

    def __init__(self,
                 num_particles=2,
                 max_plate_nesting=float('inf')):
        assert isinstance(num_particles, int) and num_particles >= 2
        if not isinstance(num_particles, int) and num_particles >= 2:
            raise ValueError("Expected num_particles >= 2, actual {}".format(num_particles))
        self.num_particles = num_particles
        self.vectorize_particles = True
        self.max_plate_nesting = max_plate_nesting

    def _get_traces(self, model, guide, *args, **kwargs):
        if self.max_plate_nesting == float("inf"):
            # TODO factor this out as a stand-alone helper.
            ELBO._guess_max_plate_nesting(self, model, guide, *args, **kwargs)
        vectorize = pyro.plate("num_particles_vectorized", self.num_particles,
                               dim=-1 - self.max_plate_nesting)
                               dim=-self.max_plate_nesting)

        # Trace the guide as in ELBO.
        with poutine.trace() as tr, vectorize:
            guide(*args, **kwargs)
        guide_trace = tr.trace

        # Trace the model, saving obs in tr2 and posterior predictives in tr1.
        # Trace the model, drawing posterior predictive samples.
        with poutine.trace() as tr, poutine.uncondition():
            with poutine.replay(trace=guide_trace), vectorize:
                model(*args, **kwargs)

@@ -61,7 +78,6 @@

Loading

        for site in model_trace.nodes.values():
            if site["type"] == "sample" and site["infer"].get("was_observed", False):
                site["is_observed"] = True

        if is_validation_enabled():
            check_model_guide_match(model_trace, guide_trace, self.max_plate_nesting)


@@ -69,7 +85,6 @@

Loading

        model_trace = prune_subsample_sites(model_trace)
        guide_trace.compute_log_prob()
        model_trace.compute_log_prob(site_filter=lambda name, site: not site["is_observed"])

        if is_validation_enabled():
            for site in guide_trace.nodes.values():
                if site["type"] == "sample":

@@ -87,6 +102,10 @@

Loading

        return guide_trace, model_trace

    def __call__(self, model, guide, *args, **kwargs):
        """
        Computes the surrogate loss that can be differentiated with autograd
        to produce gradient estimates for the model and guide parameters.
        """
        guide_trace, model_trace = self._get_traces(model, guide, *args, **kwargs)

        # Extract observations and posterior predictive samples.

@@ -143,4 +162,7 @@

Loading

143	162		return loss
144	163
145	164		def loss(self, args, *kwargs):
	165	+	"""
	166	+	Not implemented. Added for compatibility with unit tests only.
	167	+	"""
146	168		raise NotImplementedError("Trace_CRPS implements only surrogate loss")

Files		•	•	•	Coverage
pyro	13,231	12,434	0	797	93.97%
Project Totals (203 files)	13,231	12,434	0	797	93.97%

3.5=.5
TRAVIS_OS_NAME=linux

View CI Build Download

3.5=.5
TRAVIS_OS_NAME=linux

View CI Build Download

3.5=.5
TRAVIS_OS_NAME=linux

View CI Build Download

3.5=.5
TRAVIS_OS_NAME=linux

View CI Build Download

ignore:
  - "pyro/docutil.py"
  - "pyro/logger.py"

coverage:
  range: 60..95
  round: nearest
  precision: 2

comment: false

Read the docs

Interact now

The inner-most circle is the entire project, moving away from the center are folders then, finally, a single file. The size and color of each slice is representing the number of statements and the coverage, respectively.

SVG

The top section represents the entire project. Proceeding with folders and finally individual files. The size and color of each slice is representing the number of statements and the coverage, respectively.

SVG

Each block represents a single file in the project. The size and color of each block is represented by the number of statements and the coverage, respectively.

SVG

Learn more about the coverage sunburst graph

69	85		model_trace = prune_subsample_sites(model_trace)
70	86		guide_trace.compute_log_prob()
71	87		model_trace.compute_log_prob(site_filter=lambda name, site: not site["is_observed"])
72		-
73	88		if is_validation_enabled():
74	89		for site in guide_trace.nodes.values():
75	90		if site["type"] == "sample":

87	102		return guide_trace, model_trace
88	103
89	104		def __call__(self, model, guide, args, *kwargs):
	105	+	"""
	106	+	Computes the surrogate loss that can be differentiated with autograd
	107	+	to produce gradient estimates for the model and guide parameters.
	108	+	"""
90	109		guide_trace, model_trace = self._get_traces(model, guide, args, *kwargs)
91	110
92	111		# Extract observations and posterior predictive samples.

Navigation	Overlay
`t` Navigate files	`h` Toggle hits
`y` Change url to tip of branch	`m` Toggle misses
`b / v` Jump to prev/next hit line	`p` Toggle partial
`z / x` Jump to prev/next missed or partial line	`1..9` Toggle flags
`shift + o` Open current page in GitHub	`a` Toggle all on
`/ or ?` Show keyboard shortcuts dialog	`c` Toggle context lines or commits

21	21
22	22		class Trace_CRPS:
23	23		"""
24		-	Posterior predictive CRPS loss.
	24	+	Posterior predictive CRPS loss with ``KL(q,p)`` regularization.
25	25
26		-	This is a likelihood-free method; no densities are evaluated.
	26	+	This is a likelihood-free method, and can be used for likelihoods without
	27	+	tractible density functions. CRPS is a robust loss function, and is well
	28	+	defined for any distribution with finite absolute moment ``E[\|data\|]``.
27	29
28		-	:param num_particles: The number of particles/samples used to form the
	30	+	This requires static model structure, fully reparametrized guide, and
	31	+	reparametrized likelihood distributions in the model. Model latent
	32	+	distributions may be non-reparametrized.
	33	+
	34	+	Note that in the loss ``CRPS + KL(q,p)``, the ``CRPS`` term has data units
	35	+	whereas the ``KL(q,p)`` term has units of nats. To calibrate these units,
	36	+	you can wrap likelihood sites in ``poutine.scale``.
	37	+
	38	+	References
	39	+	[1] `Strictly Proper Scoring Rules, Prediction, and Estimation`
	40	+	Tilmann Gneiting, Adrian E. Raftery (2007)
	41	+	https://www.stat.washington.edu/raftery/Research/PDF/Gneiting2007jasa.pdf
	42	+
	43	+	:param int num_particles: The number of particles/samples used to form the
29	44		gradient estimators. Must be at least 2.
30	45		:param int max_plate_nesting: Optional bound on max number of nested
31	46		:func:`pyro.plate` contexts. This is only required when enumerating

37	52		def __init__(self,
38	53		num_particles=2,
39	54		max_plate_nesting=float('inf')):
40		-	assert isinstance(num_particles, int) and num_particles >= 2
	55	+	if not isinstance(num_particles, int) and num_particles >= 2:
	56	+	raise ValueError("Expected num_particles >= 2, actual {}".format(num_particles))
41	57		self.num_particles = num_particles
42	58		self.vectorize_particles = True
43	59		self.max_plate_nesting = max_plate_nesting
44	60
45	61		def _get_traces(self, model, guide, args, *kwargs):
46	62		if self.max_plate_nesting == float("inf"):
	63	+	# TODO factor this out as a stand-alone helper.
47	64		ELBO._guess_max_plate_nesting(self, model, guide, args, *kwargs)
48	65		vectorize = pyro.plate("num_particles_vectorized", self.num_particles,
49		-	dim=-1 - self.max_plate_nesting)
	66	+	dim=-self.max_plate_nesting)
50	67
51	68		# Trace the guide as in ELBO.
52	69		with poutine.trace() as tr, vectorize:
53	70		guide(args, *kwargs)
54	71		guide_trace = tr.trace
55	72
56		-	# Trace the model, saving obs in tr2 and posterior predictives in tr1.
	73	+	# Trace the model, drawing posterior predictive samples.
57	74		with poutine.trace() as tr, poutine.uncondition():
58	75		with poutine.replay(trace=guide_trace), vectorize:
59	76		model(args, *kwargs)

61	78		for site in model_trace.nodes.values():
62	79		if site["type"] == "sample" and site["infer"].get("was_observed", False):
63	80		site["is_observed"] = True
64		-
65	81		if is_validation_enabled():
66	82		check_model_guide_match(model_trace, guide_trace, self.max_plate_nesting)
67	83

1	ignore:
2	- "pyro/docutil.py"
3	- "pyro/logger.py"
4
5	coverage:
6	range: 60..95
7	round: nearest
8	precision: 2
9
10	comment: false

Updating our web app