NegativeBinomial distribution by kashif · Pull Request #9345 · pytorch/pytorch

kashif · 2018-07-11T13:16:35Z

implement distribution
add tests
docs

cc @ingmarschuster

vishwakftw · 2018-07-11T13:44:38Z

cc @fritzo

facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ingmarschuster

This is fine modulo correcting the semantics of unnormalized probability and normalization constant 👍

vishwakftw · 2018-07-12T21:35:49Z

13:40:52 ======================================================================
13:40:52 FAIL: test_negative_binomial_log_prob (__main__.TestDistributions)
13:40:52 ----------------------------------------------------------------------
13:40:52 Traceback (most recent call last):
13:40:52   File "test_distributions.py", line 918, in test_negative_binomial_log_prob
13:40:52     self._check_log_prob(NegativeBinomial(total_count, probs), ref_log_prob)
13:40:52   File "test_distributions.py", line 634, in _check_log_prob
13:40:52     asset_fn(i, val.squeeze(), log_prob)
13:40:52   File "test_distributions.py", line 916, in ref_log_prob
13:40:52     self.assertAlmostEqual(log_prob, expected, places=3)
13:40:52   File "/var/lib/jenkins/workspace/test/common.py", line 348, in assertAlmostEqual
13:40:52     self.assertEqual(x, y, prec, msg, allow_inf)
13:40:52   File "/var/lib/jenkins/workspace/test/common.py", line 296, in assertEqual
13:40:52     self.assertEqual(x.item(), y, prec, message, allow_inf)
13:40:52   File "/var/lib/jenkins/workspace/test/common.py", line 340, in assertEqual
13:40:52     super(TestCase, self).assertLessEqual(abs(x - y), prec, message)
13:40:52 AssertionError: 2.9444389791664403 not less than or equal to 0.001

Build failures are legitimate.

kashif · 2018-07-12T21:51:25Z

@vishwakftw indeed... I think it might be that the scipy nbinom uses the probability and number of successes... almost bed time here so will investigate in the morning.

fritzo

Looks good. I have not checked the math.

Could you please add a docs section to docs/source/distributions.rst ?

torch/distributions/negative_binomial.py

+            raise ValueError("Either `probs` or `logits` must be specified, but not both.")
+        if probs is not None:
+            self.total_count, self.probs, = broadcast_all(total_count, probs)
+            self.total_count = self.total_count.type_as(self.logits)


torch/distributions/negative_binomial.py

+        else:
+            self.total_count, self.logits, = broadcast_all(total_count, logits)
+            self.total_count = self.total_count.type_as(self.logits)
+            is_scalar = isinstance(self.logits, Number)


facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

fritzo · 2018-07-17T21:31:31Z

cc @alicanb

torch/distributions/negative_binomial.py

+
+    def sample(self, sample_shape=torch.Size()):
+        with torch.no_grad():
+            rate = torch.distributions.Gamma(concentration=self.total_count,


alicanb · 2018-07-17T22:45:59Z

Code is currently failing due to divide by 0 in sample_gamma, we might need argcheck here. Gamma concentration parameter should be positive, but it's possible that it can be fed 0 here.

kashif · 2018-07-18T06:36:46Z

thanks @alicanb for the review... will get it done

kashif · 2018-07-18T12:43:45Z

@alicanb is there a way to replicate the divide by 0 error?

facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ezyang · 2018-07-20T15:53:46Z

@kashif We can get you the docker image for ASAN builds, will that help?

kashif · 2018-07-20T15:59:20Z

@ezyang no I have a GPU box, can you tell me how exactly to run the tests with GPU enabled? Thanks!

torch/distributions/negative_binomial.py

+    @lazy_property
+    def _gamma(self):
+        return torch.distributions.Gamma(concentration=self.total_count,
+                                         rate=(1. - self.probs) / self.probs)


facebook-github-bot

@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

aten/src/ATen/native/Distributions.h

facebook-github-bot

ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

facebook-github-bot

ezyang is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

kashif · 2018-08-01T16:26:01Z

cool thanks all!

Summary: - [x] implement distribution - [x] add tests - [x] docs cc ingmarschuster Pull Request resolved: pytorch/pytorch#9345 Differential Revision: D8807023 Pulled By: ezyang fbshipit-source-id: 7bf7f352dd455e0909c58dd94e1bdebba0e8b5c8

Summary: - [x] implement distribution - [x] add tests - [x] docs cc ingmarschuster Pull Request resolved: pytorch#9345 Differential Revision: D8807023 Pulled By: ezyang fbshipit-source-id: 7bf7f352dd455e0909c58dd94e1bdebba0e8b5c8

added initial NegativeBinomial

7de0b65

kashif requested review from apaszke, colesbury, ezyang, gchanan, soumith and zdevito as code owners July 11, 2018 13:16

fixed typos and import

dd9a3d9

Kashif Rasul added 6 commits July 11, 2018 15:59

fix sampling from Gamma

0a4824a

lint

e23bd4e

neg. binomial has no enum support

0dfe1bb

typo!

63c130b

simplyfy

b8438e1

fixed sample shape

88ec5bf

facebook-github-bot reviewed Jul 11, 2018

View reviewed changes

Kashif Rasul added 4 commits July 12, 2018 14:17

added initial tests

f37b773

fixed typo

ea0af37

fixed indent

e6bd0ed

import F.logsigmoid

7bd3766

ingmarschuster reviewed Jul 12, 2018

View reviewed changes

fritzo reviewed Jul 12, 2018

View reviewed changes

facebook-github-bot reviewed Jul 14, 2018

View reviewed changes

added constraint and fixed type inf.

73c90ec

kashif requested review from ssnl and zou3519 as code owners July 16, 2018 11:10

Kashif Rasul added 2 commits July 16, 2018 13:12

added docs

e2be7ac

fix initial tests

831d4cd

alicanb reviewed Jul 17, 2018

View reviewed changes

torch/distributions/negative_binomial.py Outdated

def sample(self, sample_shape=torch.Size()):

with torch.no_grad():

rate = torch.distributions.Gamma(concentration=self.total_count,

This comment was marked as off-topic.

Sign in to view

Kashif Rasul added 3 commits July 18, 2018 11:22

added back greater_than and half_open_interval constraints

c480c43

use constraints and moved Gamma to a property

d3b987b

fixed doc string

fadc416

facebook-github-bot reviewed Jul 20, 2018

View reviewed changes

fritzo reviewed Jul 20, 2018

View reviewed changes

torch/distributions/negative_binomial.py Outdated

@lazy_property

def _gamma(self):

return torch.distributions.Gamma(concentration=self.total_count,

rate=(1. - self.probs) / self.probs)

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

use logits instead of probs

08c118a

facebook-github-bot reviewed Jul 23, 2018

View reviewed changes

fritzo reviewed Jul 23, 2018

View reviewed changes

aten/src/ATen/native/Distributions.h Outdated

This comment was marked as off-topic.

Sign in to view

return 0 if alpha is 0

82e3e8e

kashif force-pushed the neg-binomial branch from d6fddab to 82e3e8e Compare July 23, 2018 18:54

Merge branch 'master' into neg-binomial

77785d7

weiyangfb added the ready for review (this tag is deprecated) All PRs are ready for review unless they are draft, WIP, or have undismissed requested changes label Jul 24, 2018

facebook-github-bot reviewed Jul 29, 2018

View reviewed changes

ezyang approved these changes Jul 29, 2018

View reviewed changes

facebook-github-bot reviewed Jul 31, 2018

View reviewed changes

facebook-github-bot closed this in ee964c5 Aug 1, 2018

kashif deleted the neg-binomial branch August 1, 2018 16:25

ezyang added open source merged labels Jun 24, 2019

Conversation

kashif commented Jul 11, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vishwakftw commented Jul 11, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ingmarschuster left a comment

Choose a reason for hiding this comment

Uh oh!

vishwakftw commented Jul 12, 2018

Uh oh!

kashif commented Jul 12, 2018

Uh oh!

fritzo left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

fritzo commented Jul 17, 2018

Uh oh!

This comment was marked as off-topic.

Uh oh!

alicanb commented Jul 17, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kashif commented Jul 18, 2018

Uh oh!

kashif commented Jul 18, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ezyang commented Jul 20, 2018

Uh oh!

kashif commented Jul 20, 2018

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

kashif commented Aug 1, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

kashif commented Jul 11, 2018 •

edited

Loading

alicanb commented Jul 17, 2018 •

edited

Loading