Living Code

Just a quick reminder: Henry Prêcheur will be presenting Unit Testing in Python at the VanPyZ meeting tomorrow, February 3rd. The meeting will be at Workspace in Gastown (see map on VanPyZ page). Meetings are from 7-8:30, then we head out for beers afterwards.

Hope to see you there!

Goal: Iterate through all my (OS X Safari) browser windows and make a list of titles and urls which is then placed in the clipboard ready to be pasted into an email or blog post.

This is an update to Tab Dumping in Safari. That still works well as the basis for extending any Cocoa-based application at runtime, but it relies on SIMBL, which while it is a great bit of code, essentially is abusing the InputManager interface. Some developers and users shun such hacks, and at least one Apple application checks for them at startup and warns you from using them.

I have been running the WebKit nightlies, which are like Safari, but with newer code and features (most importantly to me right now, a Firebug-like developer toolkit). WebKit warns at startup that if you’re running extensions (such as SIMBL plugins) it may make the application less stable. I was running both Saft and my own tab dumping plugin, and WebKit was crashing a lot. So I removed those and the crashes went away. I miss a handful of the Saft extensions (but not having to update it for every Safari point release), and I found I really miss my little tab-dumping tool.

I toyed with the idea of rewriting it as a service, which would then be available from the services menu, but couldn’t figure out how to access the application’s windows and tabs from the service. So I tried looking at Safari’s scriptable dictionary, using the AppleScript Script Editor. Long ago, John Gruber had written about the frustration with Safari’s tabs not being scriptable, but a glance at the scripting dictionary showed me this was no longer the case (and probably hasn’t been for years, I haven’t kept track).

I am a complete n00b at AppleScript. I find the attempt at English-like syntax just confuses (and irritates) me no end. But what I wanted looked achievable with it, so I armed myself with some examples from Google searches, and Apple’s intro pages and managed to get what I wanted working. It may not be the best possible solution (in fact I suspect the string concatenation may be one of the most pessimal methods), but it Works For Me™.

In Script Editor, paste in the following:

set url_list to ""
-- change WebKit to Safari if you are not running nightlies
tell application "WebKit"
  set window_list to windows
  repeat with w in window_list
    try
      set tab_list to tabs of w
      repeat with t in tab_list
        set url_list to url_list & name of t & "\n"
        set url_list to url_list & URL of t & "\n\n"
      end repeat
    on error
      -- not all windows have tabs
    end try
  end repeat
  set the clipboard to url_list
end tell

I had to use AppleScript Utility to add the script menu to my menu bar. From there it was easy to create script folders that are specific to both WebKit and Safari and save a copy of the script (with the appropriate substitution, see comment in script) into each folder. Now I can copy the title and URL of all my open tabs onto the clipboard easily again, without any InputManager hacks.

I had some recollection that is a way to do this from Python, so I looked and found Appscript. I was able to install this with a simple easy_install appscript and quickly ported most of the applescript to Python. The only stumbling block was that I couldn’t find a way to access the clipboard with appscript, and I didn’t want to have to pull in the PyObjC framework just to write to the clipboard. So I used subprocess to call the command-line pbcopy utility.

#!/usr/local/bin/python
from appscript import app
import subprocess
tab_summaries = []
for window in app('WebKit').windows.get():
    try:
        for tab in window.tabs.get():
            name = tab.name.get().encode('utf-8')
            url = tab.URL.get().encode('utf-8')
            tab_summaries.append('%s\n%s' % (name, url))
    except:
        # not all windows have tabs
        pass
clipboard = subprocess.Popen('pbcopy', stdin=subprocess.PIPE)
clipboard.stdin.write('\n\n'.join(tab_summaries))

The remaining hurdle was simply to put the Python script I’d written into the same Scripting folder as my AppleScript version. For me this was ~/Library/Scripts/Applications/WebKit/. When run from the scripts folder, your usual environment is not inherited, so the #! line must point to the version of Python you are using (and which has Appscript installed). You should also make the script executable. Adding .py or any other extension is not necessary.

Overall, while I found AppleScript to be very powerful, and not quite as painful as I remembered, I found the Python version (warts and all) to be easier to work with. Combined with the fact that the script folder will run non-Applescript scripts, this opens up new worlds for me. I have hesitated in the past to write a lot of SIMBL-based plugins, tempting though it may be, because they are hacks, and they run in every Cocoa-based application. But adding application-specific (or universal) scripts, in Python, is pure, unadulterated goodness.

I really need to start blogging these before they happen, but I will at least try to summarize the January Vancouver Python user group meeting. Our featured speaker wasn’t able to make it, and it was only with the nudging of Andy and the gracious help of Jim and Dane at Workspace that we even had a January meeting. But a meeting was had, a surprising number of people braved the heavy rain, and we had a good time talking about our varied interests and explorations, mostly with a Python theme.

What I did over winter vacation

We had an open discussion, starting with the nominal topic of the night, Python packaging. We discussed some of the pros and cons of distutils, setuptools, and easy_install. Our casual conclusion was that all of these tools work better if you first use Ian Bicking’s VirtualEnv to prevent library pollution in your main Python install. VirtualEnv puts a copy of Python, your required libraries, and any dependencies they have into a separate directory and then runs from that, keeping a project nicely sandboxed. The OS X Python application builder py2app, works similarly, but after the fact as a final build step so you can deliver a complete application without worrying what version of Python and libraries are installed on a user’s computer. Ian Bicking also has package installer you can use with (or without) VirtualEnv, called Pip, and if you check his two links here you may notice a pattern forming for his blog titles.

From there we wandered around, topic-wise, discussing XML/HTML parsing libraries such as Beautiful Soup and lxml, from there to screen scrapers like mechanize, to web spiders like scrapy.

Conversation drifted into 3D for a bit, touching on VRML/X3D and the open-source FreeWRL viewer (built by the Canadian Government, yay!). While it isn’t really a Python project, there is a Python library to generate VRML as part of the Scientific library (not to be confused with SciPy. And if you don’t want to go the XML route, you can stay in Python because Scientific can also generate VPython code (although VPython does not yet support things like transparency or texture mapping).

We kept coming back to web frameworks, discussing the tradeoffs between Django and Zope/Plone and the impact of “cloud computing” platforms such as Amazon Web Services (AWS) and Google App Engine. While the AWS services give a lot more flexibility, they require a substantial amount of planning and configuration. App Engine has the easy deployment of a PHP app, but with the convenience of Python and the power of Big Table, but the rather substantial disadvantage of still being a beta environment that you can’t actually buy services on yet (no scaling up).

There are some workarounds for these and other problems (the App Engine 10 app limit, for instance). Hadoop was mentioned as an open-source alternative to App Engine. Amazon Machine Images can give you a head start at deploying on AWS, although you will still need to make arrangements for data persistence. 10gen appears to be building using the App Engine model (more or less), but as a smaller player, they may be more responsive to user feedback. LowEndBox is not a provider, but a blog tracking ultra-cheap hosting with root access, so you could conceivably build your own AWS on a shoestring.

Since I spent part of the holidays building Lego Mindstorms models with my kids (a machine gun and a puppy), I kept trying to steer the conversation towards robotics. Unfortunately, the Python Robotics project does not support Mindstorms yet, but they do support Roombas, so I may still be in luck. I haven’t found a Python project for programming Mindstorms yet, but you could probably wrap Not Exactly C with Python fairly easily, and after the meeting I found this script to use Bluetooth to control Mindstorms from Python.

After that we continued the discussion at the pub and I don’t have any browser history to help me remember that part of the discussion. I do remember that before the meeting started, I did plug Scratch again as the best way to introduce 6-12 year olds to programming.

Finally, I want to thank everyone on the VanPyZ mailing list for helping to organize and to re-establish the group’s web site after recent hack attacks and crashing, especially Henry Prêcheur for restoring the content and moving it to the official Python wiki for the new VanPyZ page.

This is something of a followup to earlier posts Drawing Hexmaps and Saving PNG from PyGame. A recent comment from Roger Aisling on the second of those posts asked about drawing with opacity in the Python Image Library (PIL). A check of the PIL API didn’t turn up anything, but there is an optional add-on to the PIL called aggdraw, which is based on Anti-Grain Geometry (AGG), a graphics library that seems to cover similar ground to Cairo. I’ve been meaning to play with both of these, but don’t know enough about either project to know what is better about one vs. the other. I know Cairo is used by Mozilla, GTX+, Mono, Inkscape, and WebKit (under GTX+). I wasn’t able to find much about where AGG is used.

Anyway, here are some examples showing how to draw polygons with variable opacity (or alpha). I’ve got examples for NodeBox, aggdraw, and pycairo. I was going to do some for the browser canvas, Processing, PyObjC, and SVG, but a) I wanted to get this post out, and b) PyCairo can already output to SVG and Cocoa (and Processing will be very similar to NodeBox). So if there is any interest in those examples, or others I may have forgotten, let me know in the comments.

While the images below look identical (at least they do to me), each was drawn with its respective library.

For each library I will present the drawing function and the main function that sets things up and drives it. At the end I will present the utility functions used which were the same for all libraries. No optimizations here, just the simplest thing I could do that worked.

NodeBox

I began with NodeBox, because it is so easy work with. In Processing they call programs “sketches” and in NodeBox that is very much what it feels like. The first thing I worked up was a way to draw “random” polygons in the same way for each libary. I could seed the random with a set value, but decided to use random to generate starting values, then just re-use those initial values. To get 50 different polygons, I used a list of primes, chose from the list at random, then used those selected values to increment each polygon vertex and color value (red, green, blue, and alpha). All of which was probably more work than was strictly necessary, but should be stable across different versions of Python and different operating systems.

The drawing routines all have the same signature. They take a tuple of point tuples (x,y) and a tuple of colors (r,g,b,a) as ints between 0 and 255.

def nodebox_poly(pts, clr):
    fill(*ints2floats(clr)) # convert ints in range 0-255 into floats between 0.0 and 1.0
    beginpath()             # every time we want to change colours or linewidth we need to start a new path
    moveto(*pts[-1])     # start and end at the last point in the list
    for pt in pts:
        lineto(*pt)
    endpath()

def nodebox_main():
    size(WIDTH, HEIGHT) # nodebox wants to know how big a canvas you are working with
    draw_all(nodebox_poly) # No additional setup needed

PyCairo

The lines in cairo_main where we create the surface and where we save it are all we would have to change if we wanted to generate a Cocoa view, an SVG document, a PDF document, etc. Two of the main differences between this and the NodeBox example are that there is a bit more setup (Cairo doesn’t default to a white background) and we need to have a reference to the context for drawing, where NodeBox has an implicit drawing context.

def cairo_poly(pts, clr):
    ctx.set_source_rgba(*ints2floats(clr))
    ctx.move_to(*pts[-1])
    for pt in pts:
        ctx.line_to(*pt)
    ctx.close_path()
    ctx.fill()

def cairo_main():
    # Setup Cairo
    import cairo
    global ctx
    surface = cairo.ImageSurface(cairo.FORMAT_ARGB32, WIDTH, HEIGHT)
    ctx = cairo.Context(surface)
    # fill background white
    cairo_poly(((0,0),(WIDTH,0),(WIDTH,HEIGHT),(0,HEIGHT)),(255,255,255,255))
    draw_all(cairo_poly)
    surface.write_to_png('cairo_example.png')

aggdraw

The aggdraw example is very similar to Cairo. In fact, if I’d done either of these first, I would have passed the context object around and simply ignored it in NodeBox to eliminate a global variable. Aggdraw can draw in memory without using PIL, but doesn’t appear to have a way to write to a PNG image that way, so we draw to a PIL image (which requires us to flush all drawing operations before writing the file).

def aggdraw_poly(pts, clr):
    import aggdraw
    global ctx
    b = aggdraw.Brush(clr, clr[-1])
    pts2 = []
    for p in pts:
        pts2.extend(p)
    ctx.polygon(pts2, b)

def aggdraw_main():
    import aggdraw, Image
    global ctx
    img = Image.new('RGBA', (WIDTH, HEIGHT), "white")
    ctx = aggdraw.Draw(img)
    draw_all(aggdraw_poly)
    ctx.flush()
    img.save('aggdraw_example.png')

Utilites

These were the same for all of the examples. If you put it all in one file (or use mine, linked below) you can choose what library to use by switching which main function is called.

MIN_ALPHA = 50
MAX_ALPHA = 100

WIDTH = 500
HEIGHT = 250

#
#   Utilities
#
def hex2tuple(hex_color):
    return tuple([int(hex_color[i:i+2], 16) for i in range(1,9,2)])

def tuple2hex(tuple_color):
    return "#%0.2X%0.2X%0.2X%0.2X" % tuple_color

def ints2floats(tuple_color):
    return tuple([c / 255.0 for c in tuple_color])

def inc_point(p, dp):
    return (p[0] + dp[0]) % WIDTH, (p[1] + dp[1]) % HEIGHT

def inc_triangle(t, dt):
    return tuple([inc_point(t[i], dt[i]) for i in range(3)])

def inc_color(c, dc):
    new_c = [(c[i] + dc[i]) % 256 for i in range(3)]
    new_a = (c[3] + dc[3]) % MAX_ALPHA
    if new_a < MIN_ALPHA: new_a += MIN_ALPHA
    new_c.append(new_a)
    return tuple(new_c)

def draw_all(draw_fn):
    triangle = start_t
    color = start_c
    for i in range(50):
        triangle = inc_triangle(triangle, dt)
        color = inc_color(color, dc)
        draw_fn(triangle, color)

#
#   Starting and incrementing values
#
start_c = hex2tuple('#0xE6A20644')
start_t = (127, 132), (341, 171), (434, 125)
dt = (107, 23), (47, 73), (13, 97)
dc = 61, 113, 109, 41

As you can see, these libraries are all quite capable of producing nice vector-based, anti-aliased polygons with alpha blending. I have heard anecdotal evidence that AGG is faster than Cairo, but Cairo appears to be more widely used. Aside from that I don't know why one would be preferred over the other (unless you want to get into language wars: AGG is written in C++, Cairo is written in C). Both have bindings in a number of languages besides Python, if that's your thing. NodeBox is still my favourite tool for noodling around in with graphics programming. All of these are great fun, easy to use, and handy to have in your bag of tricks.

I tested this under OS X (required for NodeBox), using pycairo installed via fink (using fink's Python 2.5) and aggdraw installed via easy_install (using the builtin in Python 2.5).

Here is my test file containing all of the code above: Alpha polygons example code

This is a collaboration that Daniela and I did together. Her words and voice, my scripting and adding some titles. The words to the poem can be found on her site here (or actually, on any page of her site since it is currently in her sidebar). The animation itself was done in NodeBox (so this example is Mac-only, but it could probably be adapted easily to Processing.

from math import radians, sin, cos
from random import seed

POEM = '''
in earth dreams
her limbs 		are all sky
her body		inhabited
by wild birds		and busy
mammals
an endangered 		species
she lost		 her head
where reason 		gives
the impression 		everything is
clear
she motions 	to dance
a rhythm			spins
and turns		attracted
to light
if she had eyes 		I would see
the stars		she holds
dear
then 		there are 		all
the words		she escapes
slips past 		their curved
seashell ears	as they listen
to capture 		her
on the page		 my words
are already	changing her
while she weaves 	gossamer
and soft moss		to hang
in autumn 	woods
'''.split()

size(320,240)
speed(30)

class Vector(object):

    def __init__(self, idx):
        self._starting_angle = self.angle = radians(random(360))
        self.rate = random(0.5,2.5)
        self.distance = sin(radians(idx * 2) + 180) * 100
        self.x = 0
        self.y = 0

    def step(self):
       self.angle = self._starting_angle + FRAME * radians(self.rate)
       self.x = cos(self.angle) * self.distance
       self.y = sin(self.angle) * self.distance
       return self.rate, self.x, self.y

class Word(object):

    index = []

    def __init__(self, idx, word):
		self.red = random()
		self.green  = random(0.4, 0.75)
		self.blue = 0.4
		self.vector = Vector(idx)
		self.word = word
		self.size = 10 + 2 * len(word)
		self.duration = 3 * (len(word) + 2)
		Word.index += [idx] * self.duration

    def step(self, alpha=1.0):
		fill(self.red, self.green, self.blue, alpha)
		rate, x, y = self.vector.step()
		push()
		rotate(rate * FRAME)
		font('Georgia', self.size)
		text(self.word, x, y)
		pop()

def setup():
    global words
    seed(0)
    words = [Word(idx, word) for idx,word in enumerate(POEM)]

def draw():
    translate(WIDTH/2 -50,HEIGHT/2)
    background(1.0, 0.97, 0.86)
    # Add one word at a time, then show all words for 100 frames more
    if FRAME < len(Word.index):
        last = Word.index[FRAME]
        for word in words[:last]:
            word.step()
    # Now fade out and have blank screen for titles
    elif FRAME < len(Word.index) + 100:
        alpha = 1.0 - 0.01 * (FRAME - len(Word.index))
        for word in words:
            word.step(alpha)

I used iMovie '08 to mix the voice and animation and to add the titles. I wasn't very happy with the changes to iMovie, I found the earlier versions were much more flexible and easy to use. On the other hand, I tried Norrkross Movie, that I actually had to buy, and was unable to do it at all. Next time I will just add the titles in NodeBox and only use iMovie to add the soundtrack.

My Python script could probably be cleaned up too. This was what I had after many iterations and experiments to get the timing, color, and rotation where we were both happy with it.

This was my first time uploading to YouTube and the resulting movie is very jumpy, in the original movie I uploaded was not. If anyone has suggestions how to avoid that (whether pre-processing or a better movie hosting service), please let me know in the comments. Overall I was happy with how it turned out and I'm looking forward to our next collaboration.