*** complogger has joined the channel | 2009-11-24 10:53:31 |
<toffer> | yes i saw it | 2009-11-24 11:06:16 |
| but that's nothing new | 2009-11-24 11:06:49 |
<Shelwien> | well, after that i also messed with my extrapolation formula a little | 2009-11-24 11:07:14 |
<toffer> | and? | 2009-11-24 11:07:35 |
<Shelwien> | and after using n*(1-wr)+wr for update instead of n+1 | 2009-11-24 11:07:41 |
| i've got p^wr/(p^wr+(1-p)^wr) for extrapolation | 2009-11-24 11:08:01 |
<toffer> | you'd better place some bars on top of your pc | 2009-11-24 11:08:11 |
<Shelwien> | and that's exactly what i added to paq mixer too | 2009-11-24 11:08:28 |
<toffer> | n*(1-wr)+wr = exponential aging? | 2009-11-24 11:08:54 |
<Shelwien> | yeah, somewhat | 2009-11-24 11:10:40 |
| anyway, that p^wr mapping looks right on the plot too | 2009-11-24 11:11:09 |
| with this, maybe i'll have all the necessary components now | 2009-11-24 11:11:54 |
| rolling freqs, counter, mixer, and SSE | 2009-11-24 11:12:30 |
<toffer> | that is supposed ot be a counter probability estimation via liklihood? | 2009-11-24 11:13:09 |
| p^wr/(p^wr+(1-p)^wr) | 2009-11-24 11:13:24 |
*** Shelwien has left the channel | 2009-11-24 11:13:30 |
*** Guest9968193 has joined the channel | 2009-11-24 11:13:35 |
<Shelwien> | yeah | 2009-11-24 11:13:56 |
<toffer> | and p is? | 2009-11-24 11:14:08 |
<Shelwien> | initial probability estimation | 2009-11-24 11:14:21 |
<toffer> | i mean how do you derive it? | 2009-11-24 11:14:22 |
| ok | 2009-11-24 11:14:24 |
| so basically the primary estimate, via linear coutners i guess? | 2009-11-24 11:14:38 |
<Shelwien> | and with that it becomes a posterior probability approximation | 2009-11-24 11:14:44 |
| yeah, but with freqs | 2009-11-24 11:15:01 |
<toffer> | well, liklihoods | 2009-11-24 11:15:08 |
| how does p^wr/(p^wr+(1-p)^wr) look alike using stretch/squash? | 2009-11-24 11:15:29 |
<Shelwien> | i mean, i'd like to make it all freq-based this time | 2009-11-24 11:15:31 |
<toffer> | sq(wr*st(p)) ? | 2009-11-24 11:15:38 |
<Shelwien> | counter, mixer, and sse | 2009-11-24 11:15:39 |
| yeah ;) | 2009-11-24 11:15:46 |
| that's what i said - i added it to my version of paq mixer | 2009-11-24 11:16:01 |
<toffer> | now you still need a practical evaluation | 2009-11-24 11:16:27 |
<Shelwien> | afair there was no such thing there originally, at least not explicit | 2009-11-24 11:16:29 |
<toffer> | yes, you're right there | 2009-11-24 11:16:45 |
| but it can be implicitly contained within mixibg | 2009-11-24 11:17:03 |
| mixing | 2009-11-24 11:17:05 |
| sq(w0*s0 + ... = sq(wr*(w0'*s0 + ...)) | 2009-11-24 11:17:28 |
<Shelwien> | yeah, but that's not very compatible with optimizing | 2009-11-24 11:17:28 |
| also maybe i'd finally build my fsm with this | 2009-11-24 11:18:07 |
| if you remember, that thing with lossy compression | 2009-11-24 11:18:35 |
<toffer> | yeah | 2009-11-24 11:18:41 |
<Shelwien> | this extrapolation is exactly what it lacked before ;) | 2009-11-24 11:19:08 |
| bit histories there would be fairly short | 2009-11-24 11:19:34 |
| so that part becomes very important | 2009-11-24 11:19:46 |
<toffer> | i'd appreciate results with "real" data | 2009-11-24 11:24:09 |
| e.g. book1 | 2009-11-24 11:24:18 |
<Shelwien> | well, book1.bwt maybe | 2009-11-24 11:24:35 |
<toffer> | so it's just a direct model for context histories? | 2009-11-24 11:26:01 |
| why not integrate it into a cm compressor? | 2009-11-24 11:26:17 |
*** Shelwien has left the channel | 2009-11-24 11:26:27 |
*** Guest9968193 has joined the channel | 2009-11-24 11:26:30 |
| reboot again? | 2009-11-24 11:26:44 |
<Guest9968193> | no, irc server disconnects somehow | 2009-11-24 11:26:57 |
<toffer> | too quick i guess | 2009-11-24 11:26:58 |
| it must be the illuminates | 2009-11-24 11:27:06 |
<Guest9968193> | sure | 2009-11-24 11:27:11 |
<toffer> | ^^ | 2009-11-24 11:27:18 |
<Shelwien> | i don't have a good coder template for direct model | 2009-11-24 11:27:25 |
| and i don't like to do it with hashes | 2009-11-24 11:27:31 |
<toffer> | you could use something like the data structure of dmc | 2009-11-24 11:27:50 |
| well with some path compression | 2009-11-24 11:27:58 |
<Shelwien> | ideally it has to be sliding window + bytewise tree + unary model + bitwise entropy model | 2009-11-24 11:28:19 |
<toffer> | or use the model on multi symbol alphabets | 2009-11-24 11:28:23 |
<Shelwien> | but i didn't quite solve it neither with the tree, nor with unary model | 2009-11-24 11:28:59 |
<toffer> | unary model... reminds me of some older source where i started to write a unary coder | 2009-11-24 11:29:02 |
| trees reuqire some work compared to hashes. especially when doing them "nicely" or efficient | 2009-11-24 11:29:31 |
<Shelwien> | unary model is most problematic with bitwise models | 2009-11-24 11:29:48 |
| Shkarin's approach with mtf ranking is good enough, but only works with plain freqs | 2009-11-24 11:30:00 |
<toffer> | erm? | 2009-11-24 11:30:11 |
| i thought he actually used frequency sorting | 2009-11-24 11:30:19 |
| (or prob. sorting) | 2009-11-24 11:30:25 |
| for decomposition | 2009-11-24 11:30:34 |
<Shelwien> | he does that sometimes too - on rescales etc | 2009-11-24 11:30:48 |
| but its more of mtf most of the time | 2009-11-24 11:31:08 |
<toffer> | well one can optimize a decomposition model which minimizes the number of guesses ^^ | 2009-11-24 11:31:33 |
| reminds me of guessing game | 2009-11-24 11:31:38 |
<Shelwien> | anyway, its very hard to reorder the symbols with bitwise statistics | 2009-11-24 11:31:43 |
<toffer> | i mean shannon's | 2009-11-24 11:31:47 |
<Shelwien> | sure, that's what i'm talking about basically | 2009-11-24 11:32:18 |
| but its impossible to implement with bitwise counters etc | 2009-11-24 11:32:29 |
<toffer> | in my eyes bitwise coding is pretty attractive (i mean using flat deco.) since you don't need escapes | 2009-11-24 11:32:33 |
| on the other hand you cannot mask symbols | 2009-11-24 11:32:50 |
<Shelwien> | its relatively ok until we keep the order static | 2009-11-24 11:32:53 |
<toffer> | so both cm and ppm coding have some redundancy | 2009-11-24 11:33:18 |
| order? | 2009-11-24 11:33:36 |
<Shelwien> | guessing order or whatever | 2009-11-24 11:33:45 |
| if its freqs like in ppmd, we can just swap them and all stats remain precise | 2009-11-24 11:34:10 |
| but swapping two unary probabilities... its very messy | 2009-11-24 11:35:00 |
| division etc | 2009-11-24 11:35:10 |
| and imprecise even while very slow | 2009-11-24 11:35:26 |
<toffer> | why'd you need to swap unary probs? | 2009-11-24 11:35:52 |
<Shelwien> | because order of appearance or whatever else | 2009-11-24 11:36:17 |
| won't be efficient forever | 2009-11-24 11:36:24 |
<toffer> | i thought these unary probs would be constructed adaptivly | 2009-11-24 11:36:56 |
<Shelwien> | that's the problem | 2009-11-24 11:37:09 |
| if its a:Na b:Nb ... | 2009-11-24 11:37:32 |
| then probs are Na/(Na+Nb+...) etc | 2009-11-24 11:37:41 |
<toffer> | and where's the problem? | 2009-11-24 11:37:59 |
<Shelwien> | and we can just swap the symbols in the order | 2009-11-24 11:38:02 |
| but, for example, if instead of Na | 2009-11-24 11:38:17 |
| we'd use a linear counter there | 2009-11-24 11:38:24 |
<toffer> | ok now i now what you mean | 2009-11-24 11:38:31 |
<Shelwien> | it would be a:Pa b:Pb | 2009-11-24 11:38:32 |
| and would require some crazy renormalization to swap | 2009-11-24 11:38:50 |
<toffer> | but that's not just related to probs - it depends on how symbols are processed | 2009-11-24 11:39:32 |
<Shelwien> | ? | 2009-11-24 11:39:49 |
<toffer> | that implies you process multi-symbol alphabets | 2009-11-24 11:40:40 |
| i mean not decomposing the symbol as a basis | 2009-11-24 11:41:11 |
| (like flat deco. in cm does) | 2009-11-24 11:41:31 |
<Shelwien> | text compression would never be completely precise | 2009-11-24 11:41:39 |
| unless we start estimating atleast the byte probabilities | 2009-11-24 11:41:58 |
<toffer> | well, yes | 2009-11-24 11:42:07 |
| but as i said ppm has redundancy in escapes | 2009-11-24 11:42:14 |
<Shelwien> | not really | 2009-11-24 11:42:29 |
| ppm is redundant by whole idea | 2009-11-24 11:42:35 |
<toffer> | ? | 2009-11-24 11:42:44 |
<Shelwien> | coding with one context | 2009-11-24 11:42:50 |
| well, its a cool speed optimization in the end | 2009-11-24 11:43:07 |
<toffer> | nah - that's a question of probability estimation | 2009-11-24 11:43:14 |
<Shelwien> | basically, the thing Shkarin tried to do in ppmonstr | 2009-11-24 11:43:20 |
<toffer> | i mean the fact that it requires to encode artificial symbols | 2009-11-24 11:43:29 |
<Shelwien> | is implementing a cached mixing of some kind | 2009-11-24 11:43:39 |
<toffer> | bitwise cm has its redundancy in the unavailability of masking | 2009-11-24 11:43:56 |
| well afair ppmii is static mixing - as you said | 2009-11-24 11:44:18 |
<Shelwien> | there's no redundancy in coding escapes | 2009-11-24 11:44:27 |
<toffer> | why? | 2009-11-24 11:44:34 |
<Shelwien> | with exclusions | 2009-11-24 11:44:34 |
| the only question is to properly compute the stats | 2009-11-24 11:44:51 |
<toffer> | normally one wouldn't require to encode symbols not actually present in the message. thus inserting these symbols adds redundancy | 2009-11-24 11:45:07 |
<Shelwien> | if you encode escape from a context | 2009-11-24 11:45:21 |
| and means that symbols present in that context | 2009-11-24 11:45:41 |
<toffer> | i'm not talking about probabilty estimation of escapes, etc. but the fact that ppm needs to code artificial symbols | 2009-11-24 11:45:50 |
<Shelwien> | would be masked out and unavailable for encoding anymore | 2009-11-24 11:45:54 |
<toffer> | i know | 2009-11-24 11:46:04 |
| but that's still artificial data to code | 2009-11-24 11:46:12 |
| added by the algorithm | 2009-11-24 11:46:18 |
| not present in the data | 2009-11-24 11:46:21 |
<Shelwien> | escapes are no more artificial that some msbs in bitwise coding | 2009-11-24 11:46:29 |
| they're used to select a symbol set | 2009-11-24 11:46:50 |
<toffer> | i didn't say that cm has no additional redundancy | 2009-11-24 11:46:52 |
| i just said that ppm has | 2009-11-24 11:47:00 |
<Shelwien> | well, i still don't agree that its a redundancy | 2009-11-24 11:47:18 |
| it only makes it harder to properly compute the probabilities in some cases | 2009-11-24 11:47:46 |
| while bitwise CM has a real redundancy | 2009-11-24 11:48:04 |
<toffer> | i think it's simply bias of a certain viewpoint | 2009-11-24 11:48:26 |
<Shelwien> | because it has to assign probabilities to unused symbols | 2009-11-24 11:48:45 |
<toffer> | lz requires to code e.g. flags to distingush between a (non)-match | 2009-11-24 11:48:53 |
| that's redundancy generated by the algorithm, too | 2009-11-24 11:49:01 |
<Shelwien> | well, LZ is not a compression algorithm at all | 2009-11-24 11:49:13 |
| its an ad hoc transform of data into a match sequence | 2009-11-24 11:49:32 |
| redundant of course | 2009-11-24 11:49:47 |
<toffer> | if the resulting sequence is shorter than it's compression | 2009-11-24 11:49:49 |
| anyway | 2009-11-24 11:49:51 |
<Shelwien> | its not always shorter | 2009-11-24 11:50:06 |
<toffer> | that was just an example of redundancy added by a compression algorithm | 2009-11-24 11:50:07 |
<Shelwien> | and its unknown whether it would be shorter | 2009-11-24 11:50:15 |
| sure | 2009-11-24 11:50:21 |
<toffer> | compression is bijective, so no algorithm is guaranteed to compress everything | 2009-11-24 11:50:36 |
<Shelwien> | and as I said, bitwise coding has some redundancy too | 2009-11-24 11:50:37 |
| due to precision limits | 2009-11-24 11:50:45 |
| while PPM doesn't actually have any | 2009-11-24 11:50:57 |
<toffer> | and i never said that it doesn't contain redundancy - i just sad ppm has | 2009-11-24 11:51:02 |
<Shelwien> | as there it all depends on implementation | 2009-11-24 11:51:05 |
| well, but bytewise PPM is worse for direct compression of binaries | 2009-11-24 11:51:38 |
| and as to bijectivity | 2009-11-24 11:52:38 |
| i mean, there's no certain way to know that LZ transform would allow to compress anything | 2009-11-24 11:53:05 |
| there're only some very approximate metrics | 2009-11-24 11:53:18 |
| while redundancy is much easier to limit with PPM/CM | 2009-11-24 11:53:58 |
<toffer> | erm wasn't there a proof somewhere that lz is equivalent to statistical compression in the infinite case?! | 2009-11-24 11:53:59 |
<Shelwien> | sure, but who cares about that | 2009-11-24 11:54:25 |
| we don't have enough memory to assign a symbol to every encountered substring | 2009-11-24 11:54:58 |
<toffer> | conclusion about statistical algorithms thus can be ported | 2009-11-24 11:54:59 |
<Shelwien> | not really, that's not practical | 2009-11-24 11:55:20 |
| lz optimizers still have to use ad hoc metrics and multiple passes | 2009-11-24 11:55:41 |
<toffer> | despite not beeing practical it's still possible | 2009-11-24 11:56:22 |
| anyway - it's nice to have some more time for talks like that ^^ | 2009-11-24 11:56:37 |
*** Shelwien has left the channel | 2009-11-24 11:57:06 |
*** Guest9968193 has joined the channel | 2009-11-24 11:57:09 |
| and some more coffee | 2009-11-24 11:57:14 |
| these discs are really annoying | 2009-11-24 11:57:28 |
<Shelwien> | yeah, dunno wtf. i said this | 2009-11-24 11:57:52 |
| <Shelwien> not really, that's not practical | 2009-11-24 11:57:57 |
| <Shelwien> lz optimizers still have to use ad hoc metrics and multiple passes | 2009-11-24 11:57:57 |
| <Shelwien> ... | 2009-11-24 11:57:57 |
| <Shelwien> also, masking with bitwise coding is still possible | 2009-11-24 11:57:57 |
<toffer> | toffer> despite not beeing practical it's still possible | 2009-11-24 11:59:37 |
| anyway - it's nice to have some more time for talks like that ^^ | 2009-11-24 11:59:39 |
<Shelwien> | err... so you would have it now? ;) | 2009-11-24 11:59:57 |
| not working or anything? ;) | 2009-11-24 12:00:14 |
<toffer> | well i got at least 6 weeks of vacation now | 2009-11-24 12:00:57 |
| afterwards i gonna work at the fraunhofer institute here | 2009-11-24 12:01:19 |
<Shelwien> | cool | 2009-11-24 12:01:24 |
<toffer> | and find a project for phd | 2009-11-24 12:01:27 |
<Shelwien> | maybe you'd be able to motivate me enough to actually write something ;) | 2009-11-24 12:01:38 |
| for now i'm mainly working on that conexware backup | 2009-11-24 12:02:05 |
<toffer> | the thing which is really cool about that is that when i'm 26 i'd have my phd. most people finishing their studies are at lest 25-26 here | 2009-11-24 12:02:11 |
<Shelwien> | yeah, i guess | 2009-11-24 12:02:39 |
| unfortunately phd or anything doesn't really mean anything here atm | 2009-11-24 12:03:10 |
<toffer> | i got around army - since i was in hospital durint draft. and in east germany there're 12 years of school compared to 13 in west germany | 2009-11-24 12:03:41 |
| why? | 2009-11-24 12:03:50 |
<Shelwien> | no good jobs i guess | 2009-11-24 12:04:13 |
<toffer> | ah | 2009-11-24 12:04:16 |
<Shelwien> | these phd guys trade on the flea market here | 2009-11-24 12:04:57 |
<toffer> | >.< | 2009-11-24 12:05:19 |
| i'd leave the country than | 2009-11-24 12:05:27 |
| having a phd raises your salary | 2009-11-24 12:06:02 |
<Shelwien> | well, on other hand, that's what they're worth too | 2009-11-24 12:06:09 |
<toffer> | and provides more possibilities | 2009-11-24 12:06:12 |
<Shelwien> | because smart people already left, yeah | 2009-11-24 12:06:21 |
<toffer> | alltogetehr that's really weird | 2009-11-24 12:07:13 |
*** Shelwien has left the channel | 2009-11-24 12:08:01 |
*** Guest9968193 has joined the channel | 2009-11-24 12:08:05 |
<Shelwien> | well, anyway, are there any modelling jobs? ;) | 2009-11-24 12:09:57 |
<toffer> | yes | 2009-11-24 12:11:57 |
<Shelwien> | and does it really look like nobody knows about CM there? ;) | 2009-11-24 12:12:28 |
<toffer> | yep | 2009-11-24 12:13:29 |
<Shelwien> | huh. i guess you have a bright future then ;) | 2009-11-24 12:13:55 |
<toffer> | ^^ | 2009-11-24 12:16:05 |
*** Shelwien has left the channel | 2009-11-24 12:17:22 |
*** Guest9968193 has joined the channel | 2009-11-24 12:17:26 |
<Shelwien> | damn, that's on my side i guess | 2009-11-24 12:17:53 |
| luckily i have chanlog here ;) | 2009-11-24 12:18:06 |
<toffer> | sorry, had to phone with my family | 2009-11-24 12:18:31 |
| anyway that cm principle is mostly unknown | 2009-11-24 12:18:43 |
| or present in an ad hoc manner | 2009-11-24 12:18:51 |
<Shelwien> | yeah, thing is, i discussed it with some academic people here too | 2009-11-24 12:19:07 |
<toffer> | not taking a proper optimization of the whole forecast system into accoutn | 2009-11-24 12:19:08 |
<Shelwien> | same thing | 2009-11-24 12:19:09 |
<toffer> | well i'd characterize cm in the following way: | 2009-11-24 12:19:42 |
<Shelwien> | err, not only optimization | 2009-11-24 12:19:43 |
<toffer> | - combine several predictions adaptively | 2009-11-24 12:20:01 |
| - have specialized, adaptive, independent modeling components | 2009-11-24 12:20:29 |
| - separate the data via discrete states (contexts) | 2009-11-24 12:20:52 |
<Shelwien> | well, for me the main point is the shannon metric | 2009-11-24 12:21:29 |
<toffer> | but a fuzzy system for example can, under certain assumptations, be interpreted as a cm system, too | 2009-11-24 12:21:34 |
<Shelwien> | that's one thing which they tend to always forget | 2009-11-24 12:21:40 |
| yeah, that's what i mean too | 2009-11-24 12:22:10 |
| and another quirk is caring too much about precision and theory | 2009-11-24 12:22:47 |
<toffer> | well i'd think dropping theory would be helpful | 2009-11-24 12:23:46 |
<Shelwien> | somehow it looks like they never believe that the modelled data can be imprecise or partly random | 2009-11-24 12:23:48 |
<toffer> | mh | 2009-11-24 12:24:05 |
<Shelwien> | not, dropping theory is wrong too | 2009-11-24 12:24:14 |
<toffer> | i'd not agree - there actual models for random processes | 2009-11-24 12:24:34 |
<Shelwien> | but its important to have a right image in mind | 2009-11-24 12:24:36 |
| i mean, what we do is really simple math-wise | 2009-11-24 12:25:33 |
<toffer> | something i often encouter while looking at some papers | 2009-11-24 12:25:35 |
<Shelwien> | like linear interpolation etc | 2009-11-24 12:25:52 |
<toffer> | was that just assumptations were made without analysis | 2009-11-24 12:25:55 |
| yes | 2009-11-24 12:25:57 |
| that's simple indeed | 2009-11-24 12:26:01 |
<Shelwien> | well, i mean, they try to overcomplicate things right away | 2009-11-24 12:26:26 |
| to make it look smart ;) | 2009-11-24 12:26:31 |
<toffer> | depends | 2009-11-24 12:26:36 |
| i guess it looks like that if i'd post something on the forum sometimes | 2009-11-24 12:26:50 |
| but to some degree that is just more compact and simpler | 2009-11-24 12:27:07 |
| math wise | 2009-11-24 12:27:09 |
<Shelwien> | err... your posts are just hard to understand most of the time | 2009-11-24 12:27:44 |
| because you probably just type them and send as is ;) | 2009-11-24 12:27:53 |
<toffer> | ?! | 2009-11-24 12:28:10 |
<Shelwien> | formatting etc ;) | 2009-11-24 12:28:28 |
<toffer> | i'm used to write tex | 2009-11-24 12:28:40 |
* Shelwien wonders if vbulletin has a tex plugin | 2009-11-24 12:29:03 |
* Shelwien thinks they should for their $200 | 2009-11-24 12:29:11 |
| ^^ | 2009-11-24 12:29:24 |
| that'd be nice | 2009-11-24 12:29:28 |
| well but i guess me and probably you would be the only guys using it | 2009-11-24 12:29:40 |
| or osman maybe, too | 2009-11-24 12:29:49 |
<Shelwien> | shkarin was the only one who actually used it until now ;) | 2009-11-24 12:29:59 |
<toffer> | haven't seen him for a while | 2009-11-24 12:30:02 |
<Shelwien> | so i had to render it and post instead of him ;) | 2009-11-24 12:30:12 |
<toffer> | when writing stuff properly i mostly usei t | 2009-11-24 12:30:28 |
| it | 2009-11-24 12:30:30 |
<Shelwien> | osman? he appeared here | 2009-11-24 12:30:43 |
| had a talk with him about molecular biology ;) | 2009-11-24 12:30:54 |
| apparently he tries to write a good dna model | 2009-11-24 12:31:08 |
<toffer> | interesting, too | 2009-11-24 12:31:17 |
| but i don't got any knowledge there | 2009-11-24 12:31:27 |
<Shelwien> | yeah, he found some interesting molecular dependencies there | 2009-11-24 12:31:36 |
<toffer> | i could hardly help | 2009-11-24 12:31:38 |
| and i guess understanding everything would take some effort | 2009-11-24 12:31:57 |
| but i remember to read something about linear functions | 2009-11-24 12:32:10 |
<Shelwien> | and came here asking how to approximate a data set with lines ;) | 2009-11-24 12:32:16 |
<toffer> | that's what i mean | 2009-11-24 12:32:40 |
<Shelwien> | yeah | 2009-11-24 12:32:42 |
| apparently cumulative phases there are linear | 2009-11-24 12:32:58 |
<toffer> | phases of what? | 2009-11-24 12:33:27 |
<Shelwien> | dna triplets are mapped onto some complex numbers | 2009-11-24 12:33:29 |
| and their phases summed | 2009-11-24 12:33:33 |
<toffer> | ah | 2009-11-24 12:33:33 |
| ok | 2009-11-24 12:33:35 |
| didn't know about that | 2009-11-24 12:33:44 |
<Shelwien> | and somehow there're long smooth lines | 2009-11-24 12:33:52 |
| like millions of nucleotids | 2009-11-24 12:34:02 |
<toffer> | at least i could help there - since it's a quite easy prediction task | 2009-11-24 12:34:17 |
<Shelwien> | but that's fairly hard to use in a model imho | 2009-11-24 12:34:24 |
<toffer> | (or modeling) | 2009-11-24 12:34:27 |
<Shelwien> | as its only linear on average | 2009-11-24 12:34:40 |
<toffer> | well if you can decompose the modeled entity into several subentities | 2009-11-24 12:34:50 |
<Shelwien> | it doesn't block any given symbol from appearing at any place | 2009-11-24 12:34:58 |
| afair osman was going to use it as context though | 2009-11-24 12:35:21 |
<toffer> | you can use probability densities and the decomposition to merge the densities of the subentities back | 2009-11-24 12:35:24 |
| guess to completely understand i'd need to have a look at all of that | 2009-11-24 12:36:08 |
<Shelwien> | well, he didn't post any code | 2009-11-24 12:36:22 |
<toffer> | but i didn't know that dna triplets are mapped onto the complex plane | 2009-11-24 12:36:35 |
<Shelwien> | !grep nucleotid | 2009-11-24 12:36:42 |
| hmm, that was a month ago i guess | 2009-11-24 12:37:03 |
| he does read the forum though | 2009-11-24 12:37:18 |
<toffer> | exactly a month ago | 2009-11-24 12:37:19 |
| looks like the channel still isn't that busy | 2009-11-24 12:39:24 |
| what about inviting some people like malcom tylor? | 2009-11-24 12:39:33 |
| or charles bloom | 2009-11-24 12:39:41 |
| i accidently read some blog entries from bloom | 2009-11-24 12:40:07 |
<Shelwien> | won't work with bloom i think | 2009-11-24 12:40:11 |
<toffer> | why? | 2009-11-24 12:40:17 |
<Shelwien> | he was cool though | 2009-11-24 12:40:18 |
| well, the same reason why he doesn't write compressors anymore ;) | 2009-11-24 12:40:46 |
| some people like taylor, or shkarin, or mahoney could be nice though | 2009-11-24 12:41:30 |
<toffer> | i still don't know why | 2009-11-24 12:41:49 |
| is he dead or something?! | 2009-11-24 12:41:57 |
<Shelwien> | but in fact, i don't really know what to discuss with them | 2009-11-24 12:42:21 |
| bloom? no, alive | 2009-11-24 12:42:31 |
| just has a life or something | 2009-11-24 12:42:34 |
<toffer> | i do too | 2009-11-24 12:42:40 |
<Shelwien> | and he's a physicist anyway, not a programmer | 2009-11-24 12:42:45 |
<toffer> | but still i'm messing with all of that | 2009-11-24 12:42:49 |
<Shelwien> | well, he did too, for quite some time | 2009-11-24 12:43:03 |
<toffer> | tylor seems to be a very weird person | 2009-11-24 12:43:25 |
<Shelwien> | too paranoid, like many | 2009-11-24 12:44:09 |
| but talks at least | 2009-11-24 12:44:13 |
| but the problem, as i said, is that i don't know what to talk to them atm | 2009-11-24 12:45:03 |
| i don't really have any questions anymore | 2009-11-24 12:45:18 |
<toffer> | just some basic stuff can produce new ideas, etc | 2009-11-24 12:45:29 |
| more like interchanging opinions | 2009-11-24 12:45:47 |
<Shelwien> | at least, such that i can go asking about without thinking | 2009-11-24 12:45:47 |
| yeah, but they're too old for that or something | 2009-11-24 12:46:22 |
| and some are also too paranoid | 2009-11-24 12:46:29 |
| anyway, for example, Matt seems only interested in some philosophic questions ;) | 2009-11-24 12:47:14 |
| well, whatever | 2009-11-24 12:48:22 |
| they'd probably appear sometimes | 2009-11-24 12:48:28 |
| i just have to decompile something - that's easy ;) | 2009-11-24 12:48:40 |
<toffer> | ^^ | 2009-11-24 12:48:44 |
| anyway that ccm decompilation was a nice toss | 2009-11-24 12:49:08 |
<Shelwien> | yeah, i even implemented the parallel rc in that wave | 2009-11-24 12:49:36 |
| btw, instead of making a direct CM compressor | 2009-11-24 12:50:43 |
| i'm thinking more about hutter prize thing | 2009-11-24 12:51:02 |
| if you remember, there was my multipass idea | 2009-11-24 12:51:18 |
| kinda an extended unary coding | 2009-11-24 12:51:24 |
| where we write down a set of replace regexps | 2009-11-24 12:51:46 |
| and encode the information lost in replacements, to make them reversible | 2009-11-24 12:52:09 |
| so, its still a CM, of course | 2009-11-24 12:52:43 |
| but we can encode whole words or anything | 2009-11-24 12:53:07 |
| using bidirectional contexts | 2009-11-24 12:53:18 |
| and of any complexity too | 2009-11-24 12:53:22 |
| including stuff like distances to other words etc | 2009-11-24 12:53:34 |
| doesn't it sound interesting? | 2009-11-24 12:54:12 |
<toffer> | not yet | 2009-11-24 12:54:20 |
| ^^ | 2009-11-24 12:54:22 |
| at least at the moment | 2009-11-24 12:54:28 |
| i'm mostly thinking about some m1 fixups and extensions | 2009-11-24 12:54:46 |
<Shelwien> | yeah, somebody should finally beat ccm at least ;) | 2009-11-24 12:55:19 |
| there's also a lot to do about rangecoding in fact | 2009-11-24 12:55:59 |
<toffer> | somehow i decided to re-use a paq style vectorizable mixer. since with too may models the sse chains get too large resulting in a slowdown | 2009-11-24 12:56:06 |
<Shelwien> | %) | 2009-11-24 12:56:20 |
<toffer> | i mean | 2009-11-24 12:56:24 |
| i found out that having a 0.5 or 1 k stretch table and a 4k squash table is sufficient | 2009-11-24 12:56:40 |
| precision wise | 2009-11-24 12:56:50 |
| thus alltogether i'd need 9 or 10k of auxilary memory | 2009-11-24 12:57:20 |
<Shelwien> | however, in my case (mix_Test) it didn't visibly affect the speed | 2009-11-24 12:57:26 |
<toffer> | but a single sse chain requires ~ 80x80x2 bytes ~ 12k | 2009-11-24 12:57:44 |
<Shelwien> | when i increased the tables to 64k from 12-bit ones | 2009-11-24 12:57:45 |
| well, caches are big etc ;) | 2009-11-24 12:58:01 |
<toffer> | the most important thing for caching is the context models | 2009-11-24 12:58:21 |
| having more than ~ 4 sse maps results in a slowdown | 2009-11-24 12:58:34 |
| m1x2 with a match model already used 4 | 2009-11-24 12:58:43 |
| and that paq mixing can be vectorized | 2009-11-24 12:58:55 |
| and i wonder why noone ever tried to vectorize counter updates | 2009-11-24 12:59:06 |
| i'm just writing an inline asm piece for that | 2009-11-24 12:59:29 |
<Shelwien> | i tried to explain that to "encode" | 2009-11-24 12:59:50 |
| he even has these dual counters | 2009-11-24 13:00:00 |
| really easy to vectorize | 2009-11-24 13:00:06 |
<toffer> | ^^ | 2009-11-24 13:00:09 |
<Shelwien> | well, result is expected ;) | 2009-11-24 13:00:23 |
<toffer> | maybe rewriting as a loop is enough to let the compiler do the work | 2009-11-24 13:00:23 |
<Shelwien> | no, its not that easy | 2009-11-24 13:00:42 |
<toffer> | i tried to rewrite things as loops and looked at the vectorizer output | 2009-11-24 13:00:42 |
| but mostly things weren't vectorized | 2009-11-24 13:00:55 |
| so i need to write that on my own | 2009-11-24 13:01:01 |
<Shelwien> | you have to pay attention to types and pointers | 2009-11-24 13:01:03 |
<toffer> | i know | 2009-11-24 13:01:12 |
| but there's ftree-vectorizer-verbose=# in gcc | 2009-11-24 13:01:24 |
<Shelwien> | types = SSE2 has short ints, but not bytes or words somehow | 2009-11-24 13:01:32 |
<toffer> | for me mmx is sufficient | 2009-11-24 13:01:45 |
<Shelwien> | well, intelc doesn't support mmx anymore ;) | 2009-11-24 13:02:00 |
<toffer> | sse seems to be mostly related to floats | 2009-11-24 13:02:02 |
<Shelwien> | its good for ints too | 2009-11-24 13:02:12 |
| twice the vector size | 2009-11-24 13:02:31 |
<toffer> | yep 128 bit | 2009-11-24 13:02:38 |
| but in fact on a 64 bit cpu even "manual" vectorization can be possible | 2009-11-24 13:02:55 |
<Shelwien> | don't forget that vectorized mixing in ccm ;) | 2009-11-24 13:03:23 |
<toffer> | well in m1 there's something like that too. several things packed into a uin32 | 2009-11-24 13:04:02 |
| uint32 | 2009-11-24 13:04:06 |
<Shelwien> | btw, i still want to make that probability mapping model for jpeg | 2009-11-24 13:06:09 |
| i mean, completely unpack it into pixels, and models these pixels with CM | 2009-11-24 13:06:33 |
| and then transform the pixel p.d's into DCT coef p.d's | 2009-11-24 13:07:06 |
| should be nice compression-wise | 2009-11-24 13:07:22 |
| but i really wonder about an efficient implementation for that transform | 2009-11-24 13:07:39 |
<toffer> | well i got no hint. what is the pdf of cos X for the random variable X ^^ | 2009-11-24 13:08:27 |
| guess i'll go to gym than | 2009-11-24 13:09:00 |
*** Shelwien has left the channel | 2009-11-24 13:09:04 |
*** Shelwien has joined the channel | 2009-11-24 13:09:09 |
<Shelwien> | ;) | 2009-11-24 13:09:19 |
<toffer> | i need some more kilos to approach my physical limit which would be around 86kg ^^ | 2009-11-24 13:09:34 |
| and somehow that is really a nice variation | 2009-11-24 13:10:09 |
<Shelwien> | %) | 2009-11-24 13:10:21 |
| adding kilos is not a compression, yeah ;) | 2009-11-24 13:10:29 |
<toffer> | ^^ | 2009-11-24 13:10:35 |
| adding non-fat kilos became pretty hard the more you approach your physical limit | 2009-11-24 13:10:57 |
| in 1 year i gained 20 kilos in the next just 4 | 2009-11-24 13:11:13 |
| and now i got stuck at around 82 | 2009-11-24 13:11:22 |
| so guess it'd take at least 1-2years | 2009-11-24 13:11:36 |
| to reach the limit | 2009-11-24 13:11:40 |
<Shelwien> | ...and then you won't need any trains or cars anymore, because you would be strong enough to just run anywhere? ;) | 2009-11-24 13:14:52 |
<toffer> | not really | 2009-11-24 13:15:37 |
| but it's fun | 2009-11-24 13:15:40 |
| like others go cycling or stuff like that | 2009-11-24 13:15:48 |
| and matt participates at marathons | 2009-11-24 13:16:14 |
<Shelwien> | well, yeah | 2009-11-24 13:17:44 |
| just that i don't have any friends with that hobby here, somehow ;) | 2009-11-24 13:17:59 |
| which just makes you more exotic ;) | 2009-11-24 13:18:46 |
<toffer> | don't know about the availability of stuff like that at your side | 2009-11-24 13:18:59 |
| but here're quite alot guys doing that | 2009-11-24 13:19:07 |
<Shelwien> | here too, of course | 2009-11-24 13:19:57 |
| just not anyone i know | 2009-11-24 13:20:05 |
<toffer> | ^^ | 2009-11-24 13:20:12 |
| maybe it`s because i'm a bit younger than you are | 2009-11-24 13:20:31 |
<Shelwien> | nah, i was like that all the time | 2009-11-24 13:20:44 |
<toffer> | dunnot know than | 2009-11-24 13:21:18 |
| what else do you do? any sports? | 2009-11-24 13:21:27 |
<Shelwien> | well, shoot things with a bow ;) | 2009-11-24 13:21:42 |
<toffer> | bow? | 2009-11-24 13:21:48 |
| shooting? | 2009-11-24 13:21:53 |
<Shelwien> | climb trees and eat apples ;) | 2009-11-24 13:21:55 |
<toffer> | shoot apples off people's heads ^^ | 2009-11-24 13:22:19 |
<Shelwien> | nah, they don't allow me somehow ;) | 2009-11-24 13:22:46 |
<toffer> | ... might be caused by the possibility of weird "side effects" | 2009-11-24 13:23:27 |
<Shelwien> | nah, they're just paranoid ;) | 2009-11-24 13:23:48 |
<toffer> | sure ^^ | 2009-11-24 13:24:30 |
<Shelwien> | then, cracking passwords (atm)... or it doesn't count as sport? ;) | 2009-11-24 13:25:10 |
<toffer> | i meant physical activities | 2009-11-24 13:25:22 |
<Shelwien> | then sleeping ;) | 2009-11-24 13:26:02 |
<toffer> | i mostly just like that cause it can be challenging and you can switch off your mind | 2009-11-24 13:26:03 |
<Shelwien> | yeah, i understand | 2009-11-24 13:26:29 |
<toffer> | just a physical compensation | 2009-11-24 13:26:33 |
<Shelwien> | i try some things sometimes too, but don't keep for long | 2009-11-24 13:26:39 |
<toffer> | what for example? | 2009-11-24 13:27:29 |
<Shelwien> | well, cycling (until bike died) or running around | 2009-11-24 13:28:18 |
<toffer> | would you try some training than? | 2009-11-24 13:28:48 |
| to me it looks like you appreciate such activities, too | 2009-11-24 13:29:02 |
| i mean in the intention i explained | 2009-11-24 13:29:13 |
<Shelwien> | well, i'd prefer to try some martial arts more likely ;) | 2009-11-24 13:29:56 |
<toffer> | that'd be interesting, too | 2009-11-24 13:30:33 |
<Shelwien> | tried some fencing before | 2009-11-24 13:30:49 |
<toffer> | hey, me too. but just for 2 or 3 weeks as a tryout | 2009-11-24 13:31:10 |
<Shelwien> | well, just didn't find a good enough company anywhere i tried, i guess | 2009-11-24 13:31:16 |
<toffer> | liked it | 2009-11-24 13:31:21 |
<Shelwien> | yeah | 2009-11-24 13:31:27 |
<toffer> | no company - come on i can hardly imagine that | 2009-11-24 13:31:42 |
<Shelwien> | there're people, but i don't quite like them ;) | 2009-11-24 13:32:01 |
<toffer> | but what i didn't like is that people mostly targeted a area which wasn't armored | 2009-11-24 13:32:24 |
| why= | 2009-11-24 13:32:28 |
| ? | 2009-11-24 13:32:30 |
| like right below your arm | 2009-11-24 13:32:45 |
<Shelwien> | good instincts? ;) | 2009-11-24 13:33:08 |
<toffer> | not everybody is evil or an illuminate | 2009-11-24 13:33:27 |
| ^^ | 2009-11-24 13:33:28 |
<Shelwien> | nah, i mean, its more healthy to go for a kill right away, right? ;) | 2009-11-24 13:34:03 |
| especially, comparing to hitting the "armored areas" ;) | 2009-11-24 13:34:22 |
<toffer> | erm i dunnot understand | 2009-11-24 13:34:27 |
<Shelwien> | nevermind ;) | 2009-11-24 13:34:38 |
<toffer> | well i was talking about the "company" | 2009-11-24 13:34:54 |
<Shelwien> | i understand | 2009-11-24 13:35:03 |
| no, i my case its more about vodka i guess ;) | 2009-11-24 13:35:15 |
| *in | 2009-11-24 13:35:21 |
<toffer> | vodka? | 2009-11-24 13:35:44 |
<Shelwien> | its like part of every sport here or i dunno ;) | 2009-11-24 13:35:56 |
<toffer> | erm?! | 2009-11-24 13:36:03 |
| so boozing is part of every sport?! | 2009-11-24 13:36:27 |
<pinc> | he is definitely kidding )) vodka itself is a national sport | 2009-11-24 13:36:32 |
<toffer> | ^^ | 2009-11-24 13:36:40 |
<Shelwien> | well, like, if you meet some people regularly, that means you have to celebrate stuff with them ;) | 2009-11-24 13:36:41 |
<toffer> | yeah but that'd not be a problem | 2009-11-24 13:37:01 |
| and afair you drink alcohol | 2009-11-24 13:37:13 |
<pinc> | we ought to %) | 2009-11-24 13:37:34 |
<Shelwien> | i do, but don't like it to become regular | 2009-11-24 13:37:43 |
<toffer> | and i don't think things get too messy | 2009-11-24 13:37:44 |
| that certainly is different in here | 2009-11-24 13:38:14 |
<Shelwien> | well, i'm sure its possible to find something here too | 2009-11-24 13:38:23 |
| not that i really tried | 2009-11-24 13:38:27 |
| frankly, i don't have enough time even for computer-specific hobbies | 2009-11-24 13:39:21 |
<toffer> | but your work time is pretty flexible | 2009-11-24 13:39:47 |
*** Shelwien has left the channel | 2009-11-24 13:40:44 |
*** Guest9968193 has joined the channel | 2009-11-24 13:40:47 |
<Guest9968193> | <Shelwien> sure, but somehow always something happens | 2009-11-24 13:41:02 |
| <Shelwien> well, at least now i'm staying at home most of the time | 2009-11-24 13:41:03 |
<Shelwien> | like, in 2004-2007 i didn't have any time for compression at all | 2009-11-24 13:41:35 |
<toffer> | anyway you have to know what to do | 2009-11-24 13:43:09 |
| i'm lucky that i've got the time for that stuff | 2009-11-24 13:43:28 |
<Shelwien> | hopefully i still have time to leave it to chance ;) | 2009-11-24 13:44:30 |
<toffer> | i think i gonna leave now for gym | 2009-11-24 13:47:04 |
<Shelwien> | well, i'll sleep then ;) | 2009-11-24 13:47:19 |
<toffer> | ok | 2009-11-24 13:47:27 |
<Shelwien> | !next | 2009-11-24 13:47:30 |