Memoizing Generators

[Mark]

Hello all,

Peter Norvig has a great recipe for memoizing Python functions available at
http://www.norvig.com/python-iaq.html . I made good use of this until I
tried to memoize a generator. At this point, it breaks down because the
_generator_ object is stored as the value for the *args key.

So, I came up with this:

class MemoizeGenerato r:
def __init__(self, fn):
self.cache={}
self.fn=fn
def __call__(self,* args):
if self.cache.has_ key(args):
for _i in self.cache[args]:
yield _i
else:
self.cache[args]=()
for _i in self.fn(*args):
self.cache[args]+=(_i,)
yield _i

Basically, if the generator hasn't been called before with these args, it
will run the generator's iterator through and append a tuple in the
dictionary value field for these keys. If the args have been used before,
then it iterates through the previously computed list. It works like a
charm for me.

Now, the two remaining issues are 1) can we "unify" the generator
memoization with the "standard" memoization and 2) can we deal with lists
as well (i.e. my generator was returning tuples ... since I'm scared of
lists in recursive generator contexts *wink*). One other memoization
issues: what if the args (which will become keys) are mutables ... in
particular, lists. I suppose we could just tuple() them up?

Regards,
Mark

[Michael Sparks]

Re: Memoizing Generators

> Now, the two remaining issues are 1) can we "unify" the generator[color=blue]
> memoization with the "standard" memoization and 2) can we deal with lists
> as well (i.e. my generator was returning tuples ... since I'm scared of
> lists in recursive generator contexts *wink*).
> One other memoization issues: what if the args (which will become
> keys) are mutables ... in particular, lists. I suppose we could just
> tuple() them up?[/color]

How would you deal with the following generator? (On the assumption you want a
a general purpose memoisation mechanism :)

def retrieveData(so urce=None, command=True,bl ocksize=1024):
if source:
if not command:
fh = open(source, "rb",0)
else:
fh = os.popen(self.c ommand)
else:
fh = sys.stdin
try:
try:
while 1:
data = fh.read(blocksi ze)
if data:
yield data
else:
raise Finality
except Exception, e:
if source:
fh.close()
raise e
except Finality:
pass

If your source is a network connection, or a real user this becomes further
indeterminate.. . Not a suggestion to not try, just worth considering :)

The other problem as I see it with your implementation is it expects the
generator to terminate. This is far from guaranteed - generators are useful
for dealing with infinite sequences/lists and working with functions that
might not halt. (eg calculate the best approximation of pi that you can based
on time available, not based on number of iterations/memory available)

eg
def runGenerator(fn ,args,timeAllot ed):
tstart = time.time()
gen = fn(args).next
r = gen()
while 1:
if time.time()-tstart >=timeAlloted :
break
r = gen()
return r

Consider the possibility that the function is a "next move" in chess and the
time is set by the first player to "hard" - ie lots of time. The next time a
player comes along with the time set to "easy", the memoised version won't be
aware of this in this scenario and returns the hard result (but very
quickly...), or worse hits stop iteration - which the original version
wouldn't. Or even worse - occasionally hits the StopIteration and normally
returns the "hard" result aroung 80-90% of the time.

That said memoising _good_ chess results from certain positions _might_ be
considered a good thing. It's still an issue though.

Regards,


Michael.

[Mark]

Re: Memoizing Generators

Hi Michael,


Michael Sparks wrote:
[color=blue]
> How would you deal with the following generator? (On the assumption you
> want a a general purpose memoisation mechanism :)[/color]
[color=blue]
> <*snip* a generator reading from an input source>[/color]
[color=blue]
> If your source is a network connection, or a real user this becomes
> further indeterminate.. . Not a suggestion to not try, just worth
> considering :)[/color]

I think one of the standard assumptions with memoization is that you have a
deterministic algorithm ... specifically, given an input, the output is
fixed. So, your example of entering a file and getting different data,
throws typical (in my opinion ... which is by no means sacred) memoization
out the window.

[color=blue]
> def runGenerator(fn ,args,timeAllot ed):
> tstart = time.time()
> gen = fn(args).next
> r = gen()
> while 1:
> if time.time()-tstart >=timeAlloted :
> break
> r = gen()
> return r[/color]

Now, this is a problem I like. It would be pretty damn awsome to memoize a
function with 1) a time constraint and 2) a parameter that would indicate
when a time constraint is no longer sufficient and thus more work needs to
be done. Finally, if the generator could _restart_ at the point of work
already done, you would be saving the work done to that point which is
memoization's purpose:

Something like:

runGenerator(PI digits, None, 10)

and then later

runGenerator(PI digits, None, 20)

Then our memoizer would say, well, I can get more digits in the extra 10
seconds .... so, I need to recompute this answer. But wait, hopefully,
I've stored both the numerical result for 10 seconds along with the
generator that got me there. So, all I need to do is look up the 10 second
generator and run it for another 10 seconds. Humm, can we copy generators?
I don't seem to think so.

So, the memoizer class would need some more logic and storage to determine
whether to return a stored result, computer a brand new result, or
bootstrap off a previous generator. There would also need to be a
mechanism for determining just how long is "long enough" that more time
needs to be allocated (a really hard problem might not get any further in
12 seconds than it does in 9 seconds). And we'd need a copy of the
generator so we could go from, say 10 second generator to 20 second, but
later go from 10 to 15 seconds. Although, we could get around the problem
by using the maximally accurate result computed so far (presuming we can
look it up in the user specified time interval). Then we'd only need the
generator at the farther point ... and we could just keep running it when
the user gives us more time. Yeah, that could work nicely!
[color=blue]
>
> Consider the possibility that the function is a "next move" in chess and
> the time is set by the first player to "hard" - ie lots of time. The next
> time a player comes along with the time set to "easy", the memoised
> version won't be aware of this in this scenario and returns the hard[/color]

We could paramaterize the generator (and/or the memoizer) based on the
difficulty setting .... couldn't we?
[color=blue]
> result (but very quickly...), or worse hits stop iteration - which the[/color]

This is another interesting problem .... the "quick result" is actually a
means of making the game harder.
[color=blue]
> Regards,
> Michael.[/color]

Great ideas. I might code up the timing based memoizer for kicks.

Regards,
Mark