Add 'n' parameter to RandomVariable to access sample_n() functions #323
Conversation
|
interesting proposal. i like it! i wonder if it breaks any of the inference algorithms? my guess is no, because they all work on samples via a random variable's also: the argument |
|
Good catch on the Re: |
|
sure, |
Changed from 'n' to 'sample_n'. Improved handling of rv copying.
dustinvtran
force-pushed
the
master
branch
2 times, most recently
from
a5ac146
to
de5c24c
Compare
* use python idiom instead of if x in scale args * let scale arg take tensors * add test for scale with tensor
Change `xrange` to `range` for python3 compatibility.
* Invgamma-Normal Metropolis Hastings * Update invgamma_normal_mh.py * Update invgamma_normal_mh.py
…491) * add implicit_klqp.py * add example; let loss_obj be an argument * fix pep8 * let user pass in custom ratio loss * improve docstrings * generalize ImplicitKLqp's scale arg to be a dict * add simple unit test * tweak example * robustify ratio loss arg; update docstrings
* Fix get_variables for recursive ops * Allow copy of recrusive graphs * Add tests for recursive args in copy / get_variables * Fix pep8 error * extend recursion safety to ancestors, children, descendants, and parents * Also mirror existing test for get_variables
* let session runs work directly on RandomVariable * add unit test * remove use of .value() in examples and docs
* refactor log_*_exp to use tf.reduce_logsumexp * replace all uses to reduce_log*exp
* tease out Laplace into new file laplace.py * replace hessian utility fn with tf.hessians * update laplace to work with MultivariateNormal* * add unit test; update docs * clean up code on pointmass vs normal * revise docs
* hierarchical logistic regression edward * pep8
* add tex/iclr2017.tex * prescribe edward's development version for now * add iclr2017.ipynb; revise snippets
* remove edward/{stats/,models/models.py
* remove all model wrapper examples
* remove model wrapper docs
* remove model wrapper tests
* update docs/
* update tests/
* update .travis.yml,setup.py
* remove ed.placeholder,ed.MFVI
* update edward/criticisms/
* update edward/inferences/
* ppca tutorial * revising ppca tutorial * minor changes post code review
* Add the docker for using the every environment * update Dockerfile and Makefile for gpu environment * Update Makefile for gpu environment * Update Makefile and README.md for specify the gpu environment * Update Makefile for specify the gpu environment * Update README.md for gpu environment * Add Dockerfile and Update Makefile and README.md for cpu environment
* make miscellaneous revisions to doc * force 'from * import *' statements in all code snippets, no 'import *' except getting started * update some examples
…ature/sample_n Rewrite to catch up with a couple months of changes.
|
@dustinvtran I've finally gotten back to this. I think it's doing what it should, but it'd be good to get another set of eyes on it. We can now pass in a tuple |
|
Cool! Is there a way to see the diff? The one here says 199 files changed. |
|
A rebase to master should fix the number of changed files, shouldn't it? |
|
I'm going to close this and open a new one based on the conjugacy PR #588. |
This is a small PR that adds a keyword argument 'n' to Edward
RandomVariablesthat lets them access thetf.contrib.distributions.Distribution.sample_n()function. This lets us replace (IMO) hacky syntax likeed.Normal(mu=tf.ones([10, 1])*mu_vec, sigma=1.)withed.Normal(mu=mu_vec, sigma=1., n=10), which is (IMO) clearer, very marginally more efficient, and makes doing conjugacy algebra on the graph a little easier in some cases.