warrenagin
commented
5 years ago Take a look at Pixiedust from IBM Watson. It creates an in-notebook dashboard to do visualizations. Warren Agin Warren Agin
On Sat, Nov 24, 2018, 5:10 PM Paul Gowder <notifications@github.com wrote:
Suggestions wanted: what python plotting library should this class use? Exploratory visualization is going to be a major theme of this course, and I intend to use it to provide concrete statistical instruction as well---for example, my current plan is to build out understanding of statistical distributions via histograms (i.e., get them to produce histograms of simulated data for a normal, see how it changes with different parameters, then ditto with other distributions, understand z scores and p values by manipulating the normal histograms, etc.)
Options:
Seaborn
Positives: easy to produce attractive, interesting plots, popular so there's tons of info about how to use.
Negatives: the API seems to have changed a lot in the last major version, and the version that comes pre-installed on Azure notebooks is obsolete; I had a kind of ugly experience https://github.com/paultopia/quantitative-methods-for-lawyers/blob/master/lessons/tbd/gotcha.ipynb trying to make sample code work. And the point of using a single cloud environment is to protect students from this kind of ugliness.
Plotly
Positives: dead simple. Produces interactive plots. Those plots are attractive.
Negatives: not terribly customizable, it might be hard to build out incremental teaching material using it---not sure if there's a way to, for example, put vertical lines and shading at different percentiles of a plot to indicate where standard p-value cutoff points are. Also, it's fundamentally a commercial product and it feels designed to nudge people into paid accounts.
Bokeh
Positives: also pretty simple, also interactive, albeit less attractive.
Negatives: I'm not sure how customizable it is, mainly because I don't personally know the library all that well. I'd have to dig in a good bit, and that would be time consuming. Not matplotlib based I believe, which potentially limits powerful options.
Raw matplotlib
Positives: you can do anything.
Negatives: It's stupidly difficult.
Pandas built-in plotting
Positive: it's a shockingly good plotting library on its own, and probably customizable because it just wraps matplotlib.
Negative: None that I can think of. It's Pandas.
Altair
Positives: Not sure, I don't know it really.
Negatives: not all dependencies are installed from the get-go. I don't know it really. Probably shouldn't try to use a library I have never used when visualization is a core part of the class.
Ggplot
Positives: I guess ggplot-style stuff is kind of an industry standard?
Negatives: I kind of hate the grammar of graphics. There are a bunch of different python ports of ggplot (though I've heard good things about plotnine)
Make my own damn library
Positives: total control. I can make sure students have the same version by only having one version available. I can just build in simple functions to generate every single thing I want. I should probably have a course library anyway, as this seems to be a fairly effective teaching method for incremental learning (e.g., fast.ai library wrapping pytorch, how to design programs class with racket learning languages, both of those courses are legendarily good). I already have a minimal plotting library https://github.com/paultopia/plottyprint built in raw matplotlib that I can build off of.
Disadvantages: a ton of work. Doesn't completely avoid problem of obsolete/wonky versions of things, since matplotlib versions tend to be problematic, I can't imagine what it's like trying to upgrade matplotlib on azure notebooks, and so I might have to pin this library to whatever version of matplotlib is on there. Which is likely to be incredibly annoying and inconvenient. Though, who knows, maybe azure comes with a modern version of matplotlib? We should be so lucky... (Actually, I just checked, it seems to come with 3.0.0, which is recent enough probably.. and plottyprint works, so there's that!)
I'm leaning toward the own library solution, with pandas as a backup, but thoughts?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/paultopia/quantitative-methods-for-lawyers/issues/8, or mute the thread https://github.com/notifications/unsubscribe-auth/AQTLwAudu5KdhbwMQfy7IZdRoecfgkxRks5uycPFgaJpZM4YxnFB .
paultopia
WithPrecedent
Suggestions wanted: what python plotting library should this class use? Exploratory visualization is going to be a major theme of this course, and I intend to use it to provide concrete statistical instruction as well---for example, my current plan is to build out understanding of statistical distributions via histograms (i.e., get them to produce histograms of simulated data for a normal, see how it changes with different parameters, then ditto with other distributions, understand z scores and p values by manipulating the normal histograms, etc.)
Options:
Seaborn
Positives: easy to produce attractive, interesting plots, popular so there's tons of info about how to use.
Negatives: the API seems to have changed a lot in the last major version, and the version that comes pre-installed on Azure notebooks is obsolete; I had a kind of ugly experience trying to make sample code work. And the point of using a single cloud environment is to protect students from this kind of ugliness.
Plotly
Positives: dead simple. Produces interactive plots. Those plots are attractive.
Negatives: not terribly customizable, it might be hard to build out incremental teaching material using it---not sure if there's a way to, for example, put vertical lines and shading at different percentiles of a plot to indicate where standard p-value cutoff points are. Also, it's fundamentally a commercial product and it feels designed to nudge people into paid accounts.
Bokeh
Positives: also pretty simple, also interactive, albeit less attractive.
Negatives: I'm not sure how customizable it is, mainly because I don't personally know the library all that well. I'd have to dig in a good bit, and that would be time consuming. Not matplotlib based I believe, which potentially limits powerful options.
Raw matplotlib
Positives: you can do anything.
Negatives: It's stupidly difficult.
Pandas built-in plotting
Positive: it's a shockingly good plotting library on its own, and probably customizable because it just wraps matplotlib.
Negative: None that I can think of. It's Pandas.
Altair
Positives: Not sure, I don't know it really.
Negatives: not all dependencies are installed from the get-go. I don't know it really. Probably shouldn't try to use a library I have never used when visualization is a core part of the class.
Ggplot
Positives: I guess ggplot-style stuff is kind of an industry standard?
Negatives: I kind of hate the grammar of graphics. There are a bunch of different python ports of ggplot (though I've heard good things about plotnine)
Make my own damn library
Positives: total control. I can make sure students have the same version by only having one version available. I can just build in simple functions to generate every single thing I want. I should probably have a course library anyway, as this seems to be a fairly effective teaching method for incremental learning (e.g., fast.ai library wrapping pytorch, how to design programs class with racket learning languages, both of those courses are legendarily good). I already have a minimal plotting library built in raw matplotlib that I can build off of.
Disadvantages: a ton of work. Doesn't completely avoid problem of obsolete/wonky versions of things, since matplotlib versions tend to be problematic, I can't imagine what it's like trying to upgrade matplotlib on azure notebooks, and so I might have to pin this library to whatever version of matplotlib is on there. Which is likely to be incredibly annoying and inconvenient. Though, who knows, maybe azure comes with a modern version of matplotlib? We should be so lucky... (Actually, I just checked, it seems to come with 3.0.0, which is recent enough probably.. and plottyprint works, so there's that!)
I'm leaning toward the own library solution, with pandas as a backup, but thoughts?