itertools.izip brokeness

[rurpy]

The code below should be pretty self-explanatory.
I want to read two files in parallel, so that I
can print corresponding lines from each, side by
side. itertools.izip( ) seems the obvious way
to do this.

izip() will stop interating when it reaches the
end of the shortest file. I don't know how to
tell which file was exhausted so I just try printing
them both. The exhausted one will generate a
StopInteration, the other will continue to be
iterable.

The problem is that sometimes, depending on which
file is the shorter, a line ends up missing,
appearing neither in the izip() output, or in
the subsequent direct file iteration. I would
guess that it was in izip's buffer when izip
terminates due to the exception on the other file.

This behavior seems plain out broken, especially
because it is dependent on order of izip's
arguments, and not documented anywhere I saw.
It makes using izip() for iterating files in
parallel essentially useless (unless you are
lucky enough to have files of the same length).

Also, it seems to me that this is likely a problem
with any iterables with different lengths.
I am hoping I am missing something...

#---------------------------------------------------------
# Task: print contents of file1 in column 1, and
# contents of file2 in column two. iterators and
# izip() are the "obvious" way to do it.

from itertools import izip
import cStringIO, pdb

def prt_files (file1, file2):

for line1, line2 in izip (file1, file2):
print line1.rstrip(), "\t", line2.rstrip()

try:
for line1 in file1:
print line1,
except StopIteration: pass

try:
for line2 in file2:
print "\t",line2,
except StopIteration: pass

if __name__ == "__main__":
# Use StringIO to simulate files. Real files
# show the same behavior.
f = cStringIO.Strin gIO

print "Two files with same number of lines work ok."
prt_files (f("abc\nde\nfg h\n"), f("xyz\nwv\nstu \n"))

print "\nFirst file shorter is also ok."
prt_files (f("abc\nde\n") , f("xyz\nwv\nstu \n"))

print "\nSecond file shorter is a problem."
prt_files (f("abc\nde\nfg h\n"), f("xyz\nwv\n" ))
print "What happened to \"fgh\" line that should be in column
1?"

print "\nBut only a problem for one line."
prt_files (f("abc\nde\nfg h\nijk\nlm\n"), f("xyz\nwv\n" ))
print "The line \"fgh\" is still missing, but following\n" \
"line(s) are ok! Looks like izip() ate a line."

[Paul Rubin]

ru***@yahoo.com writes:

The problem is that sometimes, depending on which file is the
shorter, a line ends up missing, appearing neither in the izip()
output, or in the subsequent direct file iteration. I would guess
that it was in izip's buffer when izip terminates due to the
exception on the other file.

Oh man, this is ugly. The problem is there's no way to tell whether
an iterator is empty, other than by reading from it.

http://aspn.activestate.com/ASPN/Coo.../Recipe/413614

has a kludge that you can use inside a function but that's no good
for something like izip.

For a temporary hack you could make a wrapped iterator that allows
pushing items back onto the iterator (sort of like ungetc) and a
version of izip that uses it, or a version of izip that tests the
iterators you pass it using the above recipe.

It's probably not reasonable to ask that an emptiness test be added to
the iterator interface, since the zillion iterator implementations now
existing won't support it.

A different possible long term fix: change StopIteration so that it
takes an optional arg that the program can use to figure out what
happened. Then change izip so that when one of its iterator args runs
out, it wraps up the remaining ones in a new tuple and passes that
to the StopIteration it raises. Untested:

def izip(*iterlist) :
while True:
z = []
finished = [] # iterators that have run out
still_alive = [] # iterators that are still alive
for i in iterlist:
try:
z.append(i.next ())
still_alive.app end(i)
except StopIteration:
finished.append (i)
if not finished:
yield tuple(z)
else:
raise StopIteration, (still_alive, finished)

You would want some kind of extended for-loop syntax (maybe involving
the new "with" statement) with a clean way to capture the exception info.
You'd then use it to continue the izip where it left off, with the
new (smaller) list of iterators.

[bonono]

But that is exactly the behaviour of python iterator, I don't see what
is broken.

izip/zip just read from the respectives streams and give back a tuple,
if it can get one from each, otherwise stop. And because python
iterator can only go in one direction, those consumed do lose in the
zip/izip calls.

I think you need to use map(None,...) which would not drop anything,
just None filled. Though you don't have a relatively lazy version as
imap(None,...) doesn't behave like map but a bit like zip.

ru***@yahoo.com wrote:

The code below should be pretty self-explanatory.
I want to read two files in parallel, so that I
can print corresponding lines from each, side by
side. itertools.izip( ) seems the obvious way
to do this.

izip() will stop interating when it reaches the
end of the shortest file. I don't know how to
tell which file was exhausted so I just try printing
them both. The exhausted one will generate a
StopInteration, the other will continue to be
iterable.

The problem is that sometimes, depending on which
file is the shorter, a line ends up missing,
appearing neither in the izip() output, or in
the subsequent direct file iteration. I would
guess that it was in izip's buffer when izip
terminates due to the exception on the other file.

This behavior seems plain out broken, especially
because it is dependent on order of izip's
arguments, and not documented anywhere I saw.
It makes using izip() for iterating files in
parallel essentially useless (unless you are
lucky enough to have files of the same length).

Also, it seems to me that this is likely a problem
with any iterables with different lengths.
I am hoping I am missing something...

#---------------------------------------------------------
# Task: print contents of file1 in column 1, and
# contents of file2 in column two. iterators and
# izip() are the "obvious" way to do it.

from itertools import izip
import cStringIO, pdb

def prt_files (file1, file2):

for line1, line2 in izip (file1, file2):
print line1.rstrip(), "\t", line2.rstrip()

try:
for line1 in file1:
print line1,
except StopIteration: pass

try:
for line2 in file2:
print "\t",line2,
except StopIteration: pass

if __name__ == "__main__":
# Use StringIO to simulate files. Real files
# show the same behavior.
f = cStringIO.Strin gIO

print "Two files with same number of lines work ok."
prt_files (f("abc\nde\nfg h\n"), f("xyz\nwv\nstu \n"))

print "\nFirst file shorter is also ok."
prt_files (f("abc\nde\n") , f("xyz\nwv\nstu \n"))

print "\nSecond file shorter is a problem."
prt_files (f("abc\nde\nfg h\n"), f("xyz\nwv\n" ))
print "What happened to \"fgh\" line that should be in column
1?"

print "\nBut only a problem for one line."
prt_files (f("abc\nde\nfg h\nijk\nlm\n"), f("xyz\nwv\n" ))
print "The line \"fgh\" is still missing, but following\n" \
"line(s) are ok! Looks like izip() ate a line."

[David Murmann]

ru***@yahoo.com schrieb:

[izip() eats one line]

as far as i can see the current implementation cannot be changed
to do the Right Thing in your case. pythons iterators don't allow
to "look ahead", so izip can only get the next element. if this
fails for an iterator, everything up to that point is lost.

maybe the documentation for izip should note that the given
iterators are not necessarily in a sane state afterwards.

for your problem you can do something like:

def izipall(*args):
iters = [iter(it) for it in args]
while iters:
result = []
for it in iters:
try:
x = it.next()
except StopIteration:
iters.remove(it )
else:
result.append(x )
yield tuple(result)

note that this does not yield tuples that are always the same
length, so "for x, y in izipall()" won't work. instead, do something
like "for seq in izipall(): print '\t'.join(seq)" .

hope i was clear enough, David.

[Peter Otten]

ru***@yahoo.com wrote:

The problem is that sometimes, depending on which
file is the shorter, a line ends up missing,
appearing neither in the izip() output, or in
the subsequent direct file iteration.**I*w ould
guess that it was in izip's buffer when izip
terminates due to the exception on the other file.

With the current iterator protocol you cannot feed an item that you've read
from an iterator by calling its next() method back into it; but invoking
next() is the only way to see whether the iterator is exhausted. Therefore
the behaviour that breaks your prt_files() function has nothing to do with
the itertools.
I think of itertools more as of a toolbox instead of a set of ready-made
solutions and came up with

from itertools import izip, chain, repeat

def prt_files (file1, file2):
file1 = chain(file1, repeat(""))
file2 = chain(file2, repeat(""))
for line1, line2 in iter(izip(file1 , file2).next, ("", "")):
print line1.rstrip(), "\t", line2.rstrip()

which can easily be generalized for an arbitrary number of files.

Peter

[Paul Rubin]

bo****@gmail.co m writes:

But that is exactly the behaviour of python iterator, I don't see what
is broken.
What's broken is the iterator interface is insufficient to deal with
this cleanly.
And because python iterator can only go in one direction, those
consumed do lose in the zip/izip calls.
Yes, that's the problem. It's proven useful for i/o streams to support
a pushback operation like ungetc. Maybe something like it can be done
for iterators.
I think you need to use map(None,...) which would not drop anything,
just None filled. Though you don't have a relatively lazy version as
imap(None,...) doesn't behave like map but a bit like zip.

I don't understand what you mean by this? None is not callable.

How about this (untested):

def myzip(iterlist) :
"""return zip of smaller and smaller list of iterables as the
individual iterators run out"""
sentinel = object() # unique sentinel
def sentinel_append (iterable):
return itertools.chain (iterable, itertools.repea t(sentinel))
for i in itertools.izip( map(sentinel_ap pend, iterlist)):
r = [x for x in i.next() if x is not sentinel]
if r: yield r
else: break

[bonono]

Paul Rubin wrote:

I think you need to use map(None,...) which would not drop anything,
just None filled. Though you don't have a relatively lazy version as
imap(None,...) doesn't behave like map but a bit like zip.
I don't understand what you mean by this? None is not callable.

zip([1,2,3],[4,5]) gives [(1,4),(2,5)]

map(None,[1,2,3],[4,5]) gives [(1,4),(2,5),(3, None)]

So the result of map() can be filtered out for special processing. Of
course, your empty/sentinel filled version is doing more or less the
same thing.

How about this (untested):

def myzip(iterlist) :
"""return zip of smaller and smaller list of iterables as the
individual iterators run out"""
sentinel = object() # unique sentinel
def sentinel_append (iterable):
return itertools.chain (iterable, itertools.repea t(sentinel))
for i in itertools.izip( map(sentinel_ap pend, iterlist)):
r = [x for x in i.next() if x is not sentinel]
if r: yield r
else: break

[Paul Rubin]

bo****@gmail.co m writes:

map(None,[1,2,3],[4,5]) gives [(1,4),(2,5),(3, None)]

I didn't know that until checking the docs just now. Oh man, what a
hack! I always thought Python should have a built-in identity
function for situations like that. I guess it does the above instead.
Thanks. Jeez ;-)

[bonono]

Paul Rubin wrote:

bo****@gmail.co m writes:

map(None,[1,2,3],[4,5]) gives [(1,4),(2,5),(3, None)]

I didn't know that until checking the docs just now. Oh man, what a
hack! I always thought Python should have a built-in identity
function for situations like that. I guess it does the above instead.
Thanks. Jeez ;-)

Of course, for OP's particular case, I think a specialized func() is
even better, as the None are turned into "" in the process which is
needed for string operation.

map(lambda *arg: tuple(map(lambd a x: x is not None and x or "", arg)),
["a","b","c"],["d","e"])

[Diez B. Roggisch]

> What's broken is the iterator interface is insufficient to deal with

this cleanly.
I don't consider it broken. You just think too much in terms of the OPs
problems or probably other fields where the actual data is available for
"rewinding" .

But as iterators serve as abstraction for lots of things - especially
generatiors - you can't enhance the interface.

Yes, that's the problem. It's proven useful for i/o streams to support
a pushback operation like ungetc. Maybe something like it can be done
for iterators.
No. If you want that, use

list(iterable)

Then you have random access. If you _know_ there will be only so much data
needed to "unget", write yourself a buffered iterator like this:

buffered(iterab le, size)

Maybe something like that _could_ go in the itertools. But I'm not really
convinced, as it is too tied to special cases - and besides that very
easily done.

How about this (untested):

def myzip(iterlist) :
"""return zip of smaller and smaller list of iterables as the
individual iterators run out"""
sentinel = object() # unique sentinel
def sentinel_append (iterable):
return itertools.chain (iterable, itertools.repea t(sentinel))
for i in itertools.izip( map(sentinel_ap pend, iterlist)):
r = [x for x in i.next() if x is not sentinel]
if r: yield r
else: break

If that fits your semantics - of course. But the general zip shouldn't
behave that way.

Regards,

Diez