Apply some of the filter to input stream. We can use this for bitsofcotton/i2g filtered images.
There exists Denoising Diffusion Probabilistic Models (DDPM; Ho et al. 2020). So this is another try on them but different flavoured one, we only focus to apply each pixel context to color image into monochrome one, which have the structure completely depends on filters' multiple meaning or complexity.
We need to do ulimit or edit /etc/login.conf for large malloc use cases required by larger than medium sized input.
Using this with mimalloc or so can increase memory usage with multi thread on some systems.
Some of the lieonn.hh operator >> class doesn't work as expected, might be compilers' bug.
The artificially created image stream which have counter measure to the predictions cannot be predicted by the single predictors. Also, there's similar streams we want to predict.
There's a chase the internal states dimension the stream have vs. given input image stream number. So to predict, we should feed much of image number which wins to the internal states dimensions. This can be done by shrinking input image sizes.
Also, our predictor cannot predict some of the strong habit on the number of the input index. So we make the hypothesis there's no timing-related internal states on them.
We can use bitsofcotton/goki_check_cc:test.py:[pq]redg command to recursive predictions.
./predg(32|64)?(mp)? <in0.ppm> ...
./qredg(32|64)?(mp)? <in0out.ppm> ...
./ddpmopt(32|64)?(mp)? + <in0in.ppm> <in0out.pgm> ... > cache.txt
./ddpmopt(32|64)?(mp)? - <in0.ppm> ... < cache.txt
./tcont [xyit] <in0.ppm> ...
cp `./tcont i <in0.ppm> ... | sort | head -n ... | tr '\n' ' '` outdir
2023/03/01 2023/03/09 bug fix after close #1. 2023/03/13 bug fix after close #2. 2023/03/13 integrate all files into lieonn.hh after close #3. 2023/03/18 merge latest p0, after close #4. 2023/03/20 qred makes only row-direction prediction, after close #5. 2023/03/24 ddpmopt.cc +0 00 -0 command fix, code clean, after close #6. 2023/03/29 merge p0 pnext, after close #7. 2023/03/31 persistent prediction to get average 1/2 * 2/3 pixels. 2023/04/02 merge p2 result. 2023/04/03 better simple predv. 2023/04/04 update readme. 2023/04/05 fix makeProgramInvariant scale accuracy stability. 2023/04/19 add topt.cc. 2023/04/21 shape up around makeProgramInvariant/revertProgramInvariant, algorithm changed. 2023/04/23 qredg.cc prediction/original norm fix. 2023/05/18 predv function change to better analytical, update README.md. 2023/06/08 predv normalization fix. 2023/06/11 topt.cc output normalization fix. update readme. 2023/06/13 update readme. ddpmopt.cc large change. 2023/06/14 ddpmopt.cc fix now works, update readme. 2023/06/15 update readme, some fixes around predg, qredg, ddpmopt. 2023/06/24 fix qredg.cc autoLevel in lieonn. 2023/06/27 ddpmopt.cc sz 3 to 2 change, update readme. 2023/07/07 [pq]redg predict along with middle image/line. 2023/07/08 invariant causes +1. qredg.cc fix crash. 2023/07/12 update readme. 2023/08/02 update topt, lieonn as crush to the last. 2023/08/03 topt, ddpmopt fix on apply. update readme. 2023/08/04 update readme. fix predv result norm. 2023/08/26 ddpmopt - option calculation change, update readme. update [pq]redg for recursive ones to beat with geometric average limit. 2023/09/04 auto configure predg/qredg param, conservative. 2023/09/11 auto configure predg/qredg size, aggressive ga param. 2023/09/12 fix last broken predv func and qredg.cc. 2023/09/13 change ddpmopt retry and retry on geometric average. 2023/09/19 autogamma after doing predg. update readme. 2023/09/19 predvResizeMat resize size fix to most reasonable one. 2023/09/22 ddpmopt change not to use crush but with linearInvariant. 2023/09/24 fix last up, don't know why they worked well without crash on last debug. 2023/09/25 change output size strategy, not using resize, preferring complement to predict. 2023/10/03 update readme. 2023/10/05 update readme, should close except for some of the ddpmopt for pairs of the images. 2023/10/18 update readme. 2023/10/19 mipmap impl, update readme. 2023/10/20 revert, mipmap doesn't work well. 2023/10/22 instead of complement trace of input, we should do shrink after output is done. This is from some of the numerical tests, so whole image orthogonality isn't matter even when input data is small enough, instead of them, we try to vanish some prediction errors with geometric mean. 2023/10/23 ddpmopt strategy algorithm in/output large change. 2023/10/26 update readme. 2023/10/27 update readme. close. 2023/10/30 copy structure reliably with randtools meaning. 2023/11/13 update readme. 2023/11/19 revert each line complement condition, force to use complement. 2023/11/20 only complement with 2, they smoothes output enough with our measure (> 2/3). 2023/11/21 update readme. 2023/11/23 fix known tips, there was at most double error. 2023/11/25 implement and revert the test to complement before to revertProgramInvariant, they doesn't improve well differed to shrinking. 2023/12/06 fix ddpmopt.cc as makeProgramInvariant, revertProgramInvariant to better compatible with [0,1]. 2023/12/07 fix ddpmopt.cc makeProgramInvariant double apply serious bug in crush. Also, update readme as compatible with in/output. realclose. 2023/12/15 use the tactics not to apply twice make/revertProgramInvariant on prediction, the invariant is already taken, however, this can causes P^t 1 == 0 condition on linearInvariant, we don't fix them. 2024/03/19 half p8 compatible change on prediction P1I to P01 with predg... qredg... binary. 2024/04/02 fix P01, vanish predvc. 2024/04/04 only use large accuracy on calculating pnextcache, but this is broken with cache naming. 2024/04/07 rgb2xyz, xyz2rgb on first/last of predg/qredg if color is 3ch. 2024/04/09 fix maeProgramInvariant x_k==0 to wrap into x_k:=1, refix pnext with lower digits. 2024/04/10 we count the function entropy enough beat with in/output on [pq]redg. 2024/04/12 update readme. 2024/04/14 take a median after predv before revertProgramInvariant. 2024/04/15 update readme, it is the specification of this some output to be broken. 2024/04/17 update readme. 2024/04/18 add tcont.cc, real close with omake. 2024/04/18 won't update without lieonn.hh change. 2024/04/29 update readme. 2024/05/05 p01 class fix step-step to 1-step correction. 2024/05/07 correct output depth limit to have a meaning. update readme. 2024/06/01 add predg0... qredg0... compile option, these using latest results on {p0,p1,p2}. 2024/06/02 revert. it's nonsense. 2024/06/05 fix p01 crash in rare cases. 2024/06/07 fix number of predictions to reasonable one. add another implementation on python predg.cc, only QR decomposition is differ but this has a better results?? update readme. 2024/06/09 factorize into each bit and predict with them. leave with this but this have color intensity == {0,1} confusion bug. 2024/06/09 fix last bug. average step skips. add readme. 2024/06/11 update readme. 2024/06/12 update readme. 2024/06/13 update readme. 2024/06/14 update readme. 2024/06/15 conclude 2024/06/11-2024/06/14 conditions readme. 2024/06/16 revert P210 to original, then, P01, P0 pred temporarily, update readme. 2024/06/17 merge p2 logic with p10 class. 2024/06/18 code cleaning, update readme. 2024/06/18 speed remedy. 2024/06/19 add restriction for getting average on PprogressionOnce010n. 2024/06/20 our machine is infected, take a most logically valid in our program predictions. 2024/06/21 fix fatal error on PprogressionOnce::next, they doesn't use predictors. 2024/06/21 revert and brush up, add fiocursed.cc series, brush readme. 2024/06/22 update readme, updates around pprogression causes only 1 step after/before prediction. 2024/06/22 p01 fatal fix. make/revert program invariant change friendly to predictions. 2024/06/23 large change around class instance initializer, also have progression short range fix. 2024/06/23 fatal fix around last update. 2024/06/24 fatal fix around last update, rotate predMat sloppy case. 2024/06/26 some of the assertion fix, update readme. 2024/06/26 update readme, fix around Ppersistent buf.full condition. 2024/06/27 fix predv last norm condition calculations. predMat bugs might be fixed. 2024/06/29 update readme and comments. 2024/06/30 re-insert periods with better stable method. update readme.