[WIP] Added implementation of SGHMC and example. #415

dwadden · 2017-01-19T18:40:05Z

I'm working with @nfoti and @yianma. We have some questions on a few small parts of the code. See comments in changed files.

dwadden · 2017-01-19T18:41:10Z

edward/inferences/sghmc.py

+    decreases.
+    Implements the update equations from (15) of Chen et al., 2014.
+    """
+    # TODO: Would it be more memory-efficient to use tf.control_dependencies,


Does making aliases make copies of the tensors? If so, would it save memory to use tf.control_dependencies and mutate the objects directly?

i'm not sure i understood the question but for example:

>>> x = tf.constant(0) >>> y = x >>> x, y (<tf.Tensor 'Const:0' shape=() dtype=int32>, <tf.Tensor 'Const:0' shape=() dtype=int32>)

y has a reference to the same tensor and doesn't recreate it from scratch, if that's what you're asking.

Yep, that was the question. Just wanted to make sure it wasn't recreating.

dwadden · 2017-01-19T18:41:53Z

edward/inferences/sghmc.py

+
+    # Simulate Hamiltonian dynamics with friction.
+    friction = tf.constant(self.friction, dtype = tf.float32)
+    # TODO: Allow option for exponentially decaying learning rate, or similar.


For SGLD, we found that using a constant learning rate performed better than using exponential decay. Perhaps this should be an option passed into the inference method?

sure, that makes sense. note one difficulty i had when implementing SGLD was that i could not leverage any of the existing tensorflow optimizers, and if more learning rates were an option, one would need to implement them all from scratch. if there were a way to use the tensorflow optimizers, it would be nice so long as the implementation doesn't ruin the readability and/or speed of the code.

Got it re: tensorflow.

I'll make a new PR that adds an option to keep the learning rate constant, along with an example showing cases where this may be a bit easier to work with.

dwadden · 2017-01-19T18:42:22Z

edward/inferences/sghmc.py

+    z_sample : dict
+      Latent variable keys to samples.
+    """
+    # TODO: This appears to be identical across HMC, SGHMC, SGLD. Should it be


The _log_joint function looks like it's the same. Refactor?

yeah, that sounds useful to refactor (although i'm not sure where the best location is to put it)

OK. Will think about this and submit another PR.

dwadden · 2017-01-19T18:43:33Z

examples/normal_sghmc.py

+from edward.models import Empirical, MultivariateNormalFull
+
+# Helper functions
+# TODO: Where should these go?


I added some helper functions to visualize the contours of a bivariate Gaussian (or more generally, any Edward distribution with a pdf method). Should this be refactored?

that's awesome. maybe we can leave it here for now and think about refactoring it in another PR that focuses more on general visualizations for edward

dwadden · 2017-01-19T18:44:07Z

examples/normal_sghmc.py

+    if label:
+        plt.clabel(cs, inline=1, fontsize=10)
+
+def retrieve_samples(qz):


It took a bit of work to figure out how to grab the actual MCMC samples. Is there a method to do this that we missed, or should we move this elsewhere to allow for more general use?

if you just want the samples you can do qz.params

Nice, thanks

dustinvtran

thanks for submitting the PR! this is excellent work. excited to have SGMCMC methods to experiment with.

dustinvtran · 2017-01-19T20:14:36Z

examples/normal_sghmc.py

+    if label:
+        plt.clabel(cs, inline=1, fontsize=10)
+
+def retrieve_samples(qz):


if you just want the samples you can do qz.params

dustinvtran · 2017-01-19T20:24:34Z

edward/inferences/sghmc.py

+
+    # Simulate Hamiltonian dynamics with friction.
+    friction = tf.constant(self.friction, dtype = tf.float32)
+    # TODO: Allow option for exponentially decaying learning rate, or similar.


sure, that makes sense. note one difficulty i had when implementing SGLD was that i could not leverage any of the existing tensorflow optimizers, and if more learning rates were an option, one would need to implement them all from scratch. if there were a way to use the tensorflow optimizers, it would be nice so long as the implementation doesn't ruin the readability and/or speed of the code.

dustinvtran · 2017-01-19T20:28:14Z

edward/inferences/sghmc.py

+    decreases.
+    Implements the update equations from (15) of Chen et al., 2014.
+    """
+    # TODO: Would it be more memory-efficient to use tf.control_dependencies,


i'm not sure i understood the question but for example:

>>> x = tf.constant(0) >>> y = x >>> x, y (<tf.Tensor 'Const:0' shape=() dtype=int32>, <tf.Tensor 'Const:0' shape=() dtype=int32>)

y has a reference to the same tensor and doesn't recreate it from scratch, if that's what you're asking.

dustinvtran · 2017-01-19T20:31:22Z

edward/inferences/sghmc.py

+    z_sample : dict
+      Latent variable keys to samples.
+    """
+    # TODO: This appears to be identical across HMC, SGHMC, SGLD. Should it be


yeah, that sounds useful to refactor (although i'm not sure where the best location is to put it)

dustinvtran · 2017-01-19T20:32:06Z

examples/normal_sghmc.py

+from edward.models import Empirical, MultivariateNormalFull
+
+# Helper functions
+# TODO: Where should these go?


that's awesome. maybe we can leave it here for now and think about refactoring it in another PR that focuses more on general visualizations for edward

Figured out how to retrieve MCMC trace more easily. Clean up normal_sghmc - Make indent 2 spaces instead of 4 - Get rid of TODO's. Clean up sghmc.py Get rid of TODO's.

dwadden · 2017-01-20T22:57:45Z

Cleaned things up a bit. Got rid of the README notes since we've now discussed them. Squashed into previous commit, which I now know is a bad thing to do because our discussions are now "outdated". Lesson learned.

Let me know if there are other things you'd like changed on this PR. Otherwise, I'll start thinking about the other things we discussed: step size, code viz, and refactor of _log_joint.

dustinvtran · 2017-01-21T02:10:17Z

looks good to me. can you add an example with data? e.g., linear/logistic regression is fine.

dwadden · 2017-01-23T19:47:09Z

Will do.

dwadden · 2017-01-26T01:16:03Z

Git style question: I'm want to update this branch with the latest changes from master, which contains relevant bug fixes for #412 . I'm also going to add my integer type fix.

In general, this can be done by merging master into this branch, or rebasing the branch onto master. The problem with the former approach is it will make the history messy, while doing the latter makes it impossible to view changes that were made before the rebase (like what's happened above).

Which do you prefer?

dustinvtran · 2017-01-26T02:29:39Z

i personally prefer to rebase to keep the commit history clean. that said, we're always squashing+merging pull requests, so feel free to do either.

dwadden · 2017-01-27T20:10:34Z

Cool, rebase works for me.

dustinvtran · 2017-01-28T20:08:41Z

i ran the examples in tensorflow 0.11.0 with python 2 and 3. i also double-checked upon merging that they work in tensorflow 1.0.0rc0. merging now.

dwadden commented Jan 19, 2017

View reviewed changes

dustinvtran reviewed Jan 19, 2017

View reviewed changes

nfoti and others added 2 commits January 20, 2017 14:20

Added implementation of SGHMC and example.

b1ae2ef

Figured out how to retrieve MCMC trace more easily. Clean up normal_sghmc - Make indent 2 spaces instead of 4 - Get rid of TODO's. Clean up sghmc.py Get rid of TODO's.

Fix pep8 style violations.

5828889

Add a blank line to make flake8 happy.

18bc75f

dustinvtran force-pushed the master branch from a5ac146 to de5c24c Compare January 21, 2017 16:31

Example: Bayesian linear regression with SGHMC.

340779f

dustinvtran merged commit 7164162 into blei-lab:master Jan 28, 2017

dwadden deleted the sghmc branch January 29, 2017 11:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] Added implementation of SGHMC and example. #415

[WIP] Added implementation of SGHMC and example. #415

dwadden commented Jan 19, 2017 •

edited

Loading

dwadden Jan 19, 2017

dustinvtran Jan 19, 2017

dwadden Jan 20, 2017

dwadden Jan 19, 2017

dustinvtran Jan 19, 2017

dwadden Jan 20, 2017

dwadden Jan 19, 2017

dustinvtran Jan 19, 2017

dwadden Jan 20, 2017

dwadden Jan 19, 2017

dustinvtran Jan 19, 2017

dwadden Jan 20, 2017

dwadden Jan 19, 2017

dustinvtran Jan 19, 2017

dwadden Jan 20, 2017

dustinvtran left a comment

dustinvtran Jan 19, 2017

dustinvtran Jan 19, 2017

dustinvtran Jan 19, 2017

dustinvtran Jan 19, 2017

dustinvtran Jan 19, 2017

dwadden commented Jan 20, 2017 •

edited

Loading

dustinvtran commented Jan 21, 2017 •

edited

Loading

dwadden commented Jan 23, 2017

dwadden commented Jan 26, 2017

dustinvtran commented Jan 26, 2017

dwadden commented Jan 27, 2017

dustinvtran commented Jan 28, 2017

[WIP] Added implementation of SGHMC and example. #415

[WIP] Added implementation of SGHMC and example. #415

Conversation

dwadden commented Jan 19, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dustinvtran left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

dwadden commented Jan 20, 2017 • edited Loading

dustinvtran commented Jan 21, 2017 • edited Loading

dwadden commented Jan 23, 2017

dwadden commented Jan 26, 2017

dustinvtran commented Jan 26, 2017

dwadden commented Jan 27, 2017

dustinvtran commented Jan 28, 2017

dwadden commented Jan 19, 2017 •

edited

Loading

dwadden commented Jan 20, 2017 •

edited

Loading

dustinvtran commented Jan 21, 2017 •

edited

Loading