- James only (missed meetings since 29 July)
- radial axis on inside of donut; 'arc' helper
- running in VS Code
- tests (when running locally) check against expected image output (PNG)
- James only (missed meetings since 29 July)
- archimedian (arithmetic) spiral plotting: r = a + b * theta
- generalises polar_plot (b = 0)
- plotting variance
- added spiral stream plot to example notebook
- finalise
README.md - publish final version to PyPI
- performance of spiral stream plot
- finalise docs
- not sure of use case here -- somewhat subsumed by notebooks
- maybe WAYS-related use case would help?
- reproducibility best practices usually want code not outputs
- how used in R version?
Present: Greg, Ed, Roly
Meeting cancelled (James ill)
Meeting cancelled (James unable to make it, Greg ill)
Present: Greg, Roly
- Greg: key changes from last time (see 15 July meeting notes)
- radial axis on inside of donut
- running in VS Code
- archimedian (arithmetic) spiral plotting: r = a + b * theta
- generalises polar_plot (b = 0)
- so far just plot the chains, no variance
- total ~125 LOC
- What is "inner burn" vs. "outer burn" in trace dial colours?
Present: James, Roly
- example notebooks
- consolidate polar plot infrastructure with existing plot design (including axis handling)
- one segment of "donut" per slice
- slice histograms aggregate all chains
- step plot for each chain
- using go.Histogram, we run into z-ordering problem trying to overlay step plots on top:
- implement our own histograms using go.Bar for plotting, and numpy for binning
- need this for consistency with step plots anyway (which also use numpy for binning)
- histograms overplot rather than group
- switch to Cartesian coordinates to render traces, donut
- solves z-order problem
- allows correct positioning of histograms (but have to enforce square figure)
- Polar trace plot (one per chain)
- "Ring" plot (pie chart with hole)
- Generalise histogram from first plot to one histogram per chain
- How do we decide number of burn-in iterations?
- What are the "polygons" in the R trace_dial plot?
- What is "inner" vs "outer" burn segment?
- What determines the colours used for burn-in and rest of sample?
- no new progress to report (first pass complete)
- Saving history in Backfillz object, use cases:
- summary of plots I’ve done so far (as show method)
- recall a particular plot
- save/load Backfillz object
- example notebooks, showing usage of Backfillz library + history object
- add density function plot to histogram (one per chain)
- histograms to aggregate all chains
- performance experiment with 1,000,000 iterations
- labels on "joining segments" to right of y-axis
- disable zoom/drag functionality
- plot Raftery-Lewis diagnostic
- additional x-axis above density plots
- Improve class-oriented design:
- avoid brittle dependency on magic numbers assigned to subplot axes by Plotly
- parameterise on height of Raftery-Lewis section and width of other 3 sections
- make a pass over R code to check for minor viz details/settings
- record plot information in Backfillz
- further design improvements to make subplot titles and row/column specifications less brittle
- Is Raftery-Lewis the right diagnostic, given no longer supported by PyMC3? (And is R dependency ok?)
- Should each RL plot have its own x-axis? Perhaps should be max of expected/actual iterations for all chains?
- Use cases to drive ledger requirements/design
- histogram per slice on RHS (currently for one chain only)
- single x-axis shared by histograms
- one "joining segment" per slice (shaded for now)
- rectangle drawn around in slice in trace plot
- Class-oriented design to make various things explicit:
- particular "view" of the MCMC data taken by this visualisation
- overall organisation of top-plot into subplots
- allows code to be mostly self-documenting
- Plotly vs. Bokeh w.r.t. "compositionality"
- neither allow arbitrary nesting of figures but only one level of containment
- in Bokeh, gridplots and row/column plots (which aren't themselves "plots")
- in Plotly, single flexible subplot grid with cell merging
- Bokeh: difficult to precisely place subplots because decorations affect size of core plot region
- Plotly places subplots relative to parent and then attaches decorations independently
- neither allow arbitrary nesting of figures but only one level of containment
- Use PyStan to generate sample model from 8 Schools example
- save (pickle) the sample model and test generated model against saved version
- first pass over slice histogram plot:
- LEFT: line plot of all draws (no per-chain colouring yet)
- MIDDLE: initial stab at "joining segments"