Hutter Prize, 4.17% improvement is here

[Alexander Rhatushnyak]

"Also, you can provide your own dictionary ..." -- that was a 3-minute answer, and here's a 30-minute:

Why would I need Someone Else's code if I were young and smart and looking for a problem to solve, a challenge to take?
Doesn't the fact that fast and small and simple SSE is still pretty useful (after all the enormous work performed by cmix and similar)
tell you that PAQ-like algorithms (including phda and cmix) don't recognize any patterns in the
model_number / probability_from_previous_models two-dimensional matrix?
They only have a one-dimensional view of it, first horizontal, then vertical.
But your algorithm does not have to look at this matrix at all !
Come on, guys, forget about phda and cmix, invent something radically new!
Like BWT and ANS were amazingly new solutions to old and boring problems.
And for super-brief introductions to a couple problems that look really important,
search for my last name on youtube.com (2nd half of the talk)

[byronknoll]

Here are the results for version 1.6:


enwik9 compressed size: 117039346 bytes
size of decompression program in .zip: 41911 bytes
total size (compressed file + decompression program): 117081257 bytes
compression time: 84713.758 seconds
decompression time: 88401.702 seconds
compression memory: 4996672 KiB
decompression memory: 4993904 KiB


enwik8 compressed size: 15040647 bytes
size of decompression program in .zip: 564616 bytes
total size (compressed file + decompression program): 15605263 bytes
compression time: 8946.996 seconds
decompression time: 8904.394 seconds
compression memory: 3799820 KiB
decompression memory: 3836272 KiB


Description of test machine:
processor: Intel Core i7-7700K
memory: 32GB DDR4
OS: Ubuntu 16.04

[Alexander Rhatushnyak]

The set of enwik-specific transforms
from the 2017 Hutter Prize winning entry
goes open-source today.
Archive is attached, read_me is included.

[JamesB]

What size difference do these transforms make to phda? I'm curious to know how well it does as a generic text compressor vs dedicated to this one corpus.

I guess the flip-side of this is to test what these transforms do to the performance of e.g. cmix. It's already matching phda9 for size, albeit considerably off for CPU and memory. It'd take a long while to know though!

Anyway, thanks for making them public.

[Alexander Rhatushnyak]

Here are the results for version 1.6:
...
enwik8 compressed size: 15040647 bytes

If you compress preprocessed_enwik8 instead of enwik8: 15015895 bytes.
The difference is much bigger in the compressor that may use only 1 GiB of RAM.

[byronknoll]

@Alex, thanks for posting these transforms!

I guess the flip-side of this is to test what these transforms do to the performance of e.g. cmix. It's already matching phda9 for size, albeit considerably off for CPU and memory. It'd take a long while to know though!

cmix currently compresses enwik8 to 14965334.
preprocessed_enwik8 compresses to 14953755.

[Alexander Rhatushnyak]

36445248 enwik8.gz --- 103.44%
35233985 preprocessed_enwik8.gz

29008758 enwik8.bz2 --- 101.78%
28502591 preprocessed_enwik8.bz2

24861205 enwik8.7z --- 101.41%
24515695 preprocessed_enwik8.7z

gzip 1.6 -9
bzip2 1.0.6 -9
7z 9.20 -t7z -mx=9


Update:
Also, with phda9 the size of enwik8 before DRT decreases by ~5.1% (as in the read_me),
but size after DRT decreases by ~8.36% : 60520510 ==> 55852090 bytes.

[byronknoll]

Using the instructions in enwik8preproc.zip, I extracted the two files from the phda November 2017 release:
1) preprocessed enwik8
2) phda dictionary

I tried testing these with cmix:
cmix currently compresses enwik8 to 14947555 bytes.
phda preprocessed enwik8 compresses to 14866787 bytes.
cmix using the phda dictionary compresses enwik8 to 14878492 bytes.

The dictionary used by phda is amazing! @Alex, I am hoping to use this dictionary for cmix - let me know if this is not OK. Recently I have been focusing on the DRT, including trying to improve the dictionary. I am learning a lot through trying different experiments, but I think I am a long way from being able to generate a dictionary even close to this level of performance.

[Alexander Rhatushnyak]

I am hoping to use this dictionary for cmix - let me know if this is not OK.

OK, but if you tried building a dictionary by yourself, that would be even better

phda preprocessed enwik8 compresses to 14866787 bytes.
cmix using the phda dictionary compresses enwik8 to 14878492 bytes.

So the gain from enwik-specific preprocessing is 11705 bytes,
very close to what was observed earlier:

cmix currently compresses enwik8 to 14965334.
preprocessed_enwik8 compresses to 14953755.

[byronknoll]

OK, but if you tried building a dictionary by yourself, that would be even better

Thanks! Yeah, I will definitely continue trying to build a dictionary and improve DRT.

[Mike]

Version 1.6 is here: http://qlic.altervista.org/phda9.zip
As usual, four executables are barely tested,
hopefully they will work as expected.
Compressed size of enwik9 should be 117'039'xxx.

UPDATE: 117'039'346.

New version from Alexander (same link)!

[Alexander Rhatushnyak]

Version 1.7 is here: http://qlic.altervista.org/phda9.zip
Improvement is below 0.1% on enwik9,
but bigger on smaller text files,
e.g. 0.4% on book1 from CC.

[byronknoll]

Here are the results for version 1.7:


enwik9 compressed size: 116940874 bytes
size of decompression program in .zip: 43,274 bytes
total size (compressed file + decompression program): 116984148 bytes
compression time: 83712.733 seconds
decompression time: 87596.519 seconds
compression memory: 4999504 KiB
decompression memory: 4996880 KiB


enwik8 compressed size: 15023870 bytes
size of decompression program in .zip: 565,352 bytes
total size (compressed file + decompression program): 15589222 bytes
compression time: 8907.486 seconds
decompression time: 8868.092 seconds
compression memory: 3802892 KiB
decompression memory: 3838836 KiB


Description of test machine:
processor: Intel Core i7-7700K
memory: 32GB DDR4
OS: Ubuntu 18.04

[Matt Mahoney]

Updated LTCB. http://mattmahoney.net/dc/text.html

[Jyrki Alakuijala]

Updated LTCB. http://mattmahoney.net/dc/text.html

How do we get the large window brotli results on LTCB?

These results have been available since 2016. https://groups.google.com/forum/m/#!...li/aq9f-x_fSY4

LTCB reports 223'597'884 bytes

Actually delivered with large window: 199'118'013 bytes

[Alexander Rhatushnyak]

Version 1.8 is here: http://qlic.altervista.org/phda9.zip
Memory usage is higher now, 6 GiB rather than 4.75.
No other news.
As usual, barely tested executables.

[byronknoll]

Here are the results for version 1.8:


enwik9 compressed size: 116544849 bytes
size of decompression program in .zip: 42,944 bytes
total size (compressed file + decompression program): 116587793 bytes
compression time: 86182.993 seconds
decompression time: 86305.520 seconds
compression memory: 6319256 KiB
decompression memory: 6316396 KiB


enwik8 compressed size: 15010414 bytes
size of decompression program in .zip: 558,298 bytes
total size (compressed file + decompression program): 15568712 bytes
compression time: 9162.225 seconds
decompression time: 9258.572 seconds
compression memory: 4800208 KiB
decompression memory: 4836472 KiB


Description of test machine:
processor: Intel Core i7-7700K
memory: 32GB DDR4
OS: Ubuntu 18.04

[Matt Mahoney]

The race continues. http://mattmahoney.net/dc/text.html

[Alexander Rhatushnyak]

I am going to publish an article
presenting this idea.
Most likely on arxiv.org.
Looking for a co-author or two.
I believe what's already in the draft
is worth at least 1/3 of this article.

If you want to be a co-author, it's
up to you how much you contribute,
maybe as little as one sentence.
A graph based on empirical data
would clearly look better,
but requires some work.
No problem if you can't provide it,
please do what you can. Thank you!

[Alexander Rhatushnyak]

Same draft as a Google Doc: here

[Mike]

New rule published on 2019-12-03 (and cosmetic changes on 2020-01-15):

Documented source code must be made publicly available under some OSI-approved license before the prize money will paid out.

[Shelwien]

I guess we need to work hard and post enough improvements for paq, so that Alex could make enough money to get bored of it and open-source phda.

Also the recent entry uses 5G of memory (with swap to disk) and an archive for it basically has to be encoded on the same machine (because of floats):
http://mattmahoney.net/dc/text.html#1165
So its only possible to compete using the same tricks (its fairly similar to 5x difference in window size).

I wonder why they can't do the sane thing and just reboot the contest with different sample data...
say, same 100M of enwik text, but all markup, bibliography, references filtered out.
And with higher memory limit, since 1G for 100M file means that only paq is applicable
(because it basically stores statistics in compressed format (hashtable/FSM)).

[Alexander Rhatushnyak]

Also the recent entry uses 5G of memory (with swap to disk)

Sorry do you mean the latest HP entry? or the latest phda9 ?
The latest HP entry uses less than 1 GiB of RAM, and a 176 MiB scratch file.
The latest phda9 uses ~6.02 GiB and nothing like "swap to disk".

and an archive for it basically has to be encoded on the same machine (because of floats): http://mattmahoney.net/dc/text.html#1165

Really? Have you ever tried?

[Shelwien]

> Sorry do you mean the latest HP entry? or the latest phda9 ?

latest HP entry.

> The latest HP entry uses less than 1 GiB of RAM, and a 176 MiB scratch file.

5G is reported on LTCB.
Or is it asymmetric and encoder uses 5G while decoder fits in 1G?
Then its my mistake.
I presumed that it would fit under time constraint even when working
with swap on 1G of physical memory.

> Really? Have you ever tried?

Tried this now - it didn't run because of too old glibc.
I'd try again tomorrow whether it would work by patching glibc version in unpacked decoder.

Also I remember some archive which still had windows decoder, but it didn't decode.

And I doubt that you would be able to compile that decoder from source now and get it to decode the entry archive.
Its not like cmix etc don't have the same problem, but in this case it makes contest entries hard to verify,
since math functions (log etc) would likely work differently in different glibc versions and/or on different linux distributions.

-------------

Anyway, I presume that you don't want to open-source phda.
It may be your right (or not, since paq and cmix are GPL), and in any case nobody can force you.
But can you please tell that to contest organizers, so that they can reboot the contest and let others participate?
Because competing with you in current contest (if its even possible) requires for everybody to stop posting any CM improvements for years.

[image28]

Early in the morning down under, thought I would quickly introduce myself.

As in my bio, dug up some old code from my 20's yesterday after finding this site.
Still early stages yet. wrote a quick bash test to compress the base enwiki9(*) 1gb file down two ~500mb using mostly grep regex's to generate dictionaries. while I sift through 20 years of backups.

(*) updated, had en8 but was talking about enwiki9,
had a look at improving the regex's a minute ago (seperating the xml tags may yield a better results)
e.g
cat enwik9 | grep -iEo "(<|<\/)[A-Z0-9]*>" | sort | uniq > enwik9-tags
cat enwik9 | grep -iEv "(<|<\/)[A-Z0-9]*>" | grep -iEo "[A-Z0-9]* " | sort | uniq > emwik9-words
Forgot the regex I used to extract all other symbols :T


One of my best algo's may be ready to go if I can find it. Only ever tested it on already compressed data mp3/xvid files, but could out compress tar/gz/bz2/rar/zip in those test cases.
( Once ran it for 5 days of iterations before it was compressing less than 8 bits per iteration. Decompression was much faster ).

Will try find something to post each week over the coming weeks if ya'll are keen!

Anyway, Hello, happy Wednesday/Thursday depending where you are,
Kevin

[image28]

Friday here, running my first test on enwik9, going out to enjoy the afternoon while it runs. Had to slow the algo down to fit the ram requirements. Algorithm I'm working on still has a few issues. Final version should be around 100 lines of vanilla c code ( no includes base linux gcc ), if I don't decide to rewrite it in asm. Compresses fill in iteration's, each iteration compressing the file to 3/4th's the previous iteration size. Ran into some issues with fseek and not liking my use of signed integers ( u_int32_t ). Worked around it by re-reading the file to position repeatedly. ( quick hack to make sure it could compress and decompress a test file. Slowing the program right down in the process ). Currently uses around 5GB of ram. Will update with the time it takes tomorrow more.

Happy Friday!

[well]

howdoyoudo,people!
unpacked file in attach causes segfault in lastest phda9, both with preproc and without at 1%!
thanks for attantion!
p.s.who will say that i'm useless person? i'm talanted tester
https://www.youtube.com/watch?v=0N9PCPW2fpY

[well]

i have been testing phda9
$./phda9 C9 enwik9 out
is it right? C9 - is it right option for enwik9?as you can see behind phda9 is not universal compressor, so i ask for right using instructions!
$./phda9 C enwik9 out
91% Segmentation fault
$
so C9 - is strictly for enwik9?