Tasks for first working prototype

[fritzo]

• parametrize coefficients
• handle constraints
• make github repo
• add syntactic sugar for constraints
• implement block gibbs
• pairwise / quadratic / cross terms
• ingest tsv file in example.py (fritz)
• move to HMC (fritz)
• port notebook pipeline to scripts
• add helpers to specify and validate initial data
  • {"sequences": _, "batch_id": _, "response": _}
• add some unit tests (fritz)
• wrap things up in higher-level components with validation logic (fritz)
  • e.g. wrap thompson_sample() in get_next_batch()
• improve README.md (fritz)
• improve docstrings (fritz)
• add tf8 plots to README (martin)
• make tf8 example reproducible, add plots (martin)
• make observation model configurable
  • currently a quantized response model
  • maybe add a couple options specified by string?
• extend model language to include user-provided features
  • rename FEATURES to GROUPS or FEATURE_BLOCKS?
  • continuous-valued embeddings
  • cluster ids
• model criticism (martin)
  • assess model fit, heldout error, find outliers
  • explore which features are active (rank or visualize?)
• warnings and errors
  • warn if GIBBS_BLOCKS or FEATURE_BLOCKS are too large
  • print number of parameters and the user-facing code that led to the most parameters
• choose better hyperparameter priors
• support coefficient sparsity (fritz)
  • allow Laplace priors for coefficients (probably default to this)
  • use different scale parameters for single vs pairwise coefficients