Edward Roadmap

[dustinvtran]

Following the 2017 TensorFlow Dev Summit, here is an outline of Edward going forward at least for Spring 2017. Of course, comments are always welcome. I'm happy to change priorities subject to interest. Related issues are added in parentheses.

• More community support
  • Mailing list (https://discourse.edwardlib.org)
  • Pre-trained models in the model zoo (model zoos and pre-trained models #28)
  • Jupyter notebooks for all tutorials (tutorials #35)
  • Drafting stages on a fully open source handbook
  • Repo for sharing Edward content (edwardlib/papers)
• More robustification
  • Integration tests (Simple tests of full models #174)
  • Improved error and warning messages (e.g., Improve type support in inferences and criticisms #561, Add informative message to all raise errors #563)
  • More TensorBoard support
  • Many bug fixes
• More automation
  • Automated choice of posterior approximation [waiting on transformations]
  • Automated transformations (Constrained parameters in Edward? #439)
  • Variance reduction (variance reduction techniques #4)
  • Convergence diagnostics (convergence diagnostics #5)
  • Automated hyperparameter tuning
• More applications
  • Sigmoid belief network with Gumbel-Softmax/Concrete distribution (Jang et al., 2017; Maddison et al., 2017) (sigmoid belief net #299) [waiting on tf1.1.0]
  • Large-scale LDA using (vanilla) stochastic variational inference (Hoffman et al., 2013) (a small LDA example doesn't seem to work #423, Implementing LDA with Edward #463) [waiting on conjugacy]
  • Generative model applied to ImageNet and other large-scale examples (add model examples #22, Distributed training example + tests #137)
  • Hidden Markov model (Attempting basic hidden Markov model #450) [waiting on conjugacy]
  • More inference & criticism tutorials (tutorials #35)
  • More benchmarks, including a JMLR paper and dedicated webpage
  • Example of inference with missing data & causality (guidelines on how to handle missing data #399)
• More algorithms
  • Conjugacy & symbolic algebra
    • Complete conditional distribution
    • Gibbs sampling (Gelfand and Smith, 1990)
    • (vanilla) stochastic variational inference (Hoffman et al., 2013)
  • Expressive variational families (Rezende and Mohamed, 2015; Ranganath et al., 2016) (variational models #7, example of normalizing flows #284, Edward support for recent extensions of VAE  #447)
  • More GAN algorithms, and in general for implicit probabilistic models (Inference on deterministic functions of RandomVariables #444)
  • Operator variational inference (Ranganath et al., 2016)

Here are features I probably don't have time for. However, they are very much on my mind, and are impactful additions if someone wants to take the helm. (Any of the above is also appreciated!)

• Better support for Bayesian nonparametric modeling & inference (truncated DP mixture model example #183)
• Better support for undirected models, including potential functions (Building a Potential-like custom random variable #443)
• Sparse random variables (sparse random variables #358)
• Reinforcement learning applications, interfaced with OpenAI Gym and Universe
• Dynamic computation graphs
• Built-in visualization support (visualization of specified models #351, visualization of fitted models #352, visualization of inference/convergence diagnostics #353, visualization of criticism techniques #354, visualizations of distributions #355)
• Built-in dataset support (predefined toy, simulated and benchmark data #125)
• Module for contributed models (Module for contributed models #380)

[PKU-YYang]

Hi Dustin:

Just wondering if there is any Control variates (#4) tricks already implemented in Edward? I noticed there is a branch called rao-blakcwellization, not sure if it works properly now.

Thanks.

[dustinvtran]

Not at the moment. I was working out RB as you pointed out, which turns out to be quite involved. Control variates are much easier to implement; I haven't looked into it quite yet.

[ip01]

Hi Dustin, do you, or perhaps Francisco, or whomever, plan to implement Generalized Reparametrization Gradient in Edward? If yes - could it be usefull as effective black-box technique for large-scale topic models (Dirichlet G-REP'ed via exp-covariance, Multinomial via Gumbel-Softmax)?

[dustinvtran]

I've never experimented with it so I'm not sure. pinging @franrruiz. I do think there's no one-size-fits-all solution—especially for discrete random variables, and whether it be score functions+control variates, g-rep, gumbel softmax, or reparameterization gradients. However, it would be certainly useful to have them all to experiment.

[RoyiAvital]

Is anyone keep working on this?