* https://hackmd.io/@story645/scipy2021mpl * tag: @story645, @matplotlib --- # Thanks * CZI EOSS 1 & 3 for supporting Thomas Caswell, Michael Grossberg, and me * Thomas Caswell & Michael Grossberg * Sam & Melissa * YT for making the best presentations to steal ideas from * the Matplotlib team :hearts: --- ## Rearchitecture Goals * Identify the different types of transformations * Formalize the rules of these transformations * Implement in a way that respects seperation <!--why? makes it easier to code against, implement, maintain, correctly write these abstractions, improves consistency in complex cases--> --- ### What is Matplotlib's job? <!--provide components for converting data to graphics-->  --- ### How does Matplotlib do it? <!-- the artist object-->  ---- <!-- stress it takes in an object, it forwards to an object, list of objects -->  --- ### What is the proposed API? <!-- is about functions, drop ins, decoupled flexibility -->  --- ### Graphic -> Data: Topology <!-- make the point here that it's about breaking out of the structure, is for a general model -->  ---- <!--- dig down into scatter to explain what we mean by topology --->  ---- <!---implicit assumption scatter is backed by discrete points-->  ---- <!-- made explicit via arrows rather than hard coding any assumption-->  ---- <!--- can verify w/ shuffle -->  ---- <!-- what happens when we connect?-->  ---- <!--- still have map back to data-->  ---- <!--- but introduce map that doesn't exist --->  <!--wrap w/ how this allows us to express graphic to data relationships w/o relying on one notion of continuity--> --- ### Data -> Graphic: Equivariance <!--visualization is structure preserving maps, by structure we mean things like measurment type-->  ---- <!--precip is ratio, distance between Jan 1 and Jan 2 and Jan 2 and Jan 3 is proportionally same in data and viz-->  ---- <!-- scaling the data induces scaling in the viz-->  --- ### Why? Multivariate Multidimensional Data <!--space + label or change to histogram, add in all points but keep 3 points, top and right to precip -->  ---- <!--- bookeeeping problem-->  ---- <!--- topology & generalized data model--->  ---- <!--- equivariance and consistent updates-->  ----  --- ### Takeaways <!--reframe goals, + overall status--> * decoupling the data, visual encoding, and plot assembly steps * provides easier to code against interface * improves reliability, consistency, maintability * conceptual prototype: https://github.com/story645/proposal --- ### Thank you!! * contact: @story645, haizenman@gmail.com * discourse: https://discourse.matplotlib.org/ * gitter: https://gitter.im/matplotlib/matplotlib --- ## Appendix --- ### Functional API model ```python def fiberbundle(section): index = section.section.index def projection(column_name): def values(): # if column_name is None: return pd.Series(index=index, dtype=float) return section.section[column_name][index] return values return projection return subset_k #s ``` ---- ```python @dataclass class Section: F = {'fruit': np.dtype('O'), 'calories': np.dtype('int64'), 'juice': np.dtype('bool')} K = pd.core.indexes.range.RangeIndex section: pd.DataFrame def __post_init__(self): #check the fiber matches assert self.section.dtypes.to_dict() == self.F assert isinstance(self.section.index, self.K) ```` ---- ```python data = {'fruit': ['apple', 'orange', 'lemon', 'lime'], 'calories':[95, 67, 17, 20], 'juice':[True, True, False, False]} df = pd.DataFrame(data) fruit.fiberbundle(fruit.Section(df))(None)('juice')() ``` --- ```python nus = {'position': lambda x : x, 'length' : position.Nominal({'apple': 0, 'orange': 2, 'lemon': 4, 'lime': 6}), 'floor': lambda _: 0, 'width': lambda _ : .8, 'facecolor' : color.Nominal({True: 'orange', False: 'blue'}), 'edgecolor' : lambda _ : 'k', 'linewidth' : lambda _ : 2, 'linestyle' : lambda _ : (None, None)} ``` --- ```python class BarArtist(martist.Artist): #view is fruit.fiberbundle def __init__(self, view, orientation, aes, position, length, floor, width, facecolors, edgecolors, linewidths, linestyles): # some function unpacking/ set parametrs def assemble(pos, ..., fc): xval = pos paths = [[(x, y), (x, y+yo), (x+xo, y+yo), (x+xo, y), (x, y)] for (x, xo, y, yo) in zip(xval, xoff, yval, yoff)] return paths, ..., fc def draw(self, renderer, *args, **kwargs): #conversion works something like pos = position(view(self.data)(None)(aes['position'])) # .... fc = facecolors(view(self.data)(None)(aes['facecolors'])) # all the visual conversion gets pushed to child artists self.assemble(pos, ..., fc) ``` --- ### Topological Equivariant Artist Model ---- <!--sorry didn't have time for alt text yet, these are diagrams of the math. Please bug me for more info if interested-->  ----  ----  ----  ----  ----  ----  ----  ----