Conversation
|
@strandn thanks for your contribution! I noticed there are changes in the core ( |
|
Thansk for your contribution! I am commenting with a sledgehammer: brace for impact. First of all, the Matrix change is not necessary-ish: usually compilers initialize to zero, but I appreciate your lack in trust about that: it is a good practice. But I think it needs to be refined a little (to be pedantic: The change in Tools.h has already been fixed, but it is not a problem: a rebase will remove the modification. I would like to ask you to do a small compacting of your code history, so that it will be easier to rebase and apply the new astyle. Luckyly you never merged master so it should be smooth. How much are you confortable with git? There are some smaller details, like the external dependency on ITensor, or the "using namespace x;" things, but we will walk you trough those later |
I am actually fixing this separately. I think it is a genuine bug, the point is that these functions are not regtested currently |
See one of the changes in #1385 To be backported
some implementation progress minor minor more progress function declarations in ttsketch class minor change include orders added keys and initializations to constructor progress on plumed-related functions all bare functionalities implemented (without parallelization)--compilation not attempted yet edit .gitignore Delete .vscode directory replace eigen with internal plumed matrix functionality few fixes minor typo fix compilation errors fix warnings fix CONV flag related issue change STRIDE to SAMPLESTRIDE small bug another small bug debugging prints use cout instead gsl eps bug more cout statements flush stdout small bug in update minor add couts back add prints in parasketch add prints in createttcoeff reduce number of prints clean prints bugs with bias derivative and vshift some more small fixes, plus more prints for debugging purposes minor comment out prints add some extra debugging prints for visualizing bias potential need to enable convolution immediately after sketch algorithm minor minor add count feature typo bug fix modify location of count print another small bug still more related bugs small fix--hopefully works now another tiny fix fixes related to vshift minor prints in histogram file revert prints clean up additional output regarding gradtop widen integration limits for Gaussian smoothing bug with min and max CV values typo further fix Gaussian integration if variables are periodic minor debug 4D case minor less output oops output in envi R part more refined output more print adjustments omit details from output more output changes fix bug related to rank revert changes switch back to new version get rid of vbias dumping change integration limits update default quadrature params minor several changes; need to improve ACA algorithm minor massive changes some more progress further progress ready to test code compilation fixes small fixes minor add parallelization small output change debugging output minor bug fix more debugging minor another fix yet more debugging more output even more output another fix more prints another fix? minor idk why gsl is failing fixed issue with epsabs that didnt fix the problem; lets examine the integrand function more closely output integrand more finely play around with key try key=4 now try relaxing tolerance try epsabs=1e-7 change back to 1e-8, see if other things can cause improvement try qags integration switch to 1e-6 try decreasing epsrel change epsabs too will use plumed input file to tune params disable gsl error handle add include new thing to debug issue with update vb again minor debuggging prints i think its fixed now remove debugging prints write vbias to separate file instead of stdout temporarily disable omp uncomment omp lines separate gsl workspace for each thread vbias append to file minor typo ok should work now output vbias with original basis small fix fix issue with vshift minor clean up modifications in vbias fix minor issues minor change the way vbias is outputted another version revert to old version minor comment out vbias printing try slight modification in f added h5 support+write TT to disk output arg labels to file debug outputting change prints this should fix it also need to store samples minor fixes minor resolve bug--can only output ITensor objects to HDF5 typo another typo minor yet another small fix more debugging output add more debugging statemement should hopefully fix the problem hangs--add print statements again this should do the trick different bug fix clean up change yet again to new version, include append change add some log output minor nsamples should start from 0 instead of 1 fix potential indexing problem minor might not need in the end progress on fes code a bit more progress first attempt implemented fix compilation issues issue with inheritance try multiple inheritance some more fix attempts minor add component debugging output some more debugging prints typo should not use default getStride() oops forgot to add key fixed--remove debugging output free energy must do -ktlog minor fix input stuff minor maybe better now? small fix minor try update instead of calculate minor revert to old version don't use all samples if not using final vb missing comma a few changes add some debugging output add parallelization add include remove output modify f, i think vbias 0 introduces errors change back test disabling omp when iterating through samples never mind try nomax version change f again move parsing of keys elsewhere print stuff to file to fix segfault semicolon dont check filename unless restart minor remove output stuff fix file input one more fix confused about error--add some prints hopefully this fixes it minor big revert back to old formalism small fixes add macros move implementation to cpp file commit cpp file cpp not cc bug with wt try adding support for multiple walkers fix minor issue some improvements in restart part debugging output some more output bug fix? change output again lets see what all_traj looks like this should do it change output back one more fix ok now it should work minor somethings wrong--add debugging output minor output to log instead of stdout flush more often changes to output more detailed output about vector sizes some more output still more output something is wrong with h5 write output to cerr instead serialize file read write to file minor small fix confusing minor exrta output small try using gdb print to log instead temporarily disable ttWrite and ttRead typo uncomment write print contents of tt move barrier fix typo try putting h5 in sub directory temporarily comment out h5_write now comment out close commenting out close(f) fixes it? no real need for barriers comment out Vbias printing bug fix make WIDTH a vector added ability to output adjusted vbias with smaller values bug in adj bias more fixing typo forgot to multiply adj vmax by kt fix bug try new adj vbias version many changes, including trimming now before linsolve, and added covmat calculations some compilation fixes this is better maybe? flush log a few times add a bunch of debugging prints additional prints possible fix somethings wrong with Pinv? more output cant matrixout to cout print vbias for aladip small bug check mu quick check minor some clean up comment out printing of vbias print integrals print tt i think this fixes integrals clean up typo fix bug with empirical cov implementation attempt for fast aca compilation fixes minor a few more small changes shift to -2kT separate conv for vb and g minor improvements to quadrature grid minor make grid finer comment out printing of F minor minor output tweaks never mind add option to not include all samples in aca calculations add feature to print ttcross errors minor another small fix typo try reducing number of grid points bug in l2 error calculations need to print some stuff need to take sqrt try 5x instead of 2x shift f from -2kT to 0 minor dont compute vshift on restart if aca some clean up forgot to add samples to aca object when restarting if error, decrease count by 1 need to check for errors multiple times needed to check reading of tt as well major changes--mainly memorystride and vshift minor small fix 1-indexing minor fix gradient filenames further fix io related stuff only barrier if multiple walkers some debugging output bug fix 6d bias output for debugging minor finer print change limits of integration for convolution remove vbias output uncomment vshift output more vbias output comment out output enable output small fix comment out output output for 3d and 4d bias minor comment out vbias output something is wrong with restart, adding some debugging output change location of output somehow it works minor output flush output some more output add barrier maybe more output more changes to output minor maybe fix? minor this should do it add vbias output for 6d calculations adj vbias minor vbias output for aladip all ranks should output vbias minor change position of vbias output also print adj vbias minor fix restart portion comment out vbias printing slight change with adj vmax vbias is now sum of log ttsketch again bug fix printing of 2d marginals typo debugging output minor typo more output more output refined output should fix filename location change filename change small clean up fix the way vshift and adj vshift is handled on restart change convergence scheme for aca oops when aca, search through grid samples minor revert another small revert never mind small fix another small fix print additional files that have gaussian kernels included also print 2d vbias typos move print code to right placee modify metad file to print vbias cross sections vscode is being weird include fstream clean up mtd minor only print vbias stuff for ala4 fix ttcross convergence check minor typo somethings wrong undo some small changes confused another fix attempt minor remove extra output again oops rearrange code re rearrange revert cross to fix issue add erased function small fix big changes--revert ttcross and implement gaussian basis fix redundant declaration fix ttsketch conv output gaussian marginals implementation small fix+clean up fix 2d marginal output + more gaussian basis clean up minor fix should be able to run with gaussian basis now fix some more output accidentally removed code something is wrong this should do it minor added feature to compute l2 errors for each aca rank fix errors more small fixes set upper limit on aca rank invert/pseudoinvert fix add debugging output indexing seems to be off by 1 in some places comment out prints rank scheme redundance--simplify check if res is 0 and if so break add a lot of debugging output for troubleshooting kernel basis fix only mpi rank 0 should print rearrange some code minor also need to only print for one k move code around some more output oops nbasis fix typo another typo? exclude redundant gaussian int2 corrections try pseudoinverse? multiply by ginv after cross too change res new criterion sump p implementation small fix typo in keyword registering initialize sump height forgot to write vb to disk some debugging output oops print tt see if nonzero ranks are trying to read prematurely try adding barrier right after ttsumwrite move variable declarations around lets try something need to append to h5 file something is wrong with pluseq--debug match site indices minor add some more useful output about ranks oops forgot to account for vshift oops try diff aca version debugging output more debugging oops more detailed output this should fix it minor revert to old aca minor output info from ala3 vbias, sketch and mtd include fstream some more fixes comment out metad code check quality of ginv correct matrix mult fix output oops try increasing kernel basis bandwidth remove bandwidth increase clean up vshift got excluded from sump vbias for some reason comment out some vbias printing code fix bf output for sump minor fix add vbias printing for 4d comment out mtd code slight tweak in cross convergence slightly more revert cross change seems to be problematic try compiling ttmetad fix issues one more small thing some more small changes oops some debugging output fix oops cant print vectors try again minor another possible fix even more debugging output minor typo test with delta oops test something go back to gaussians hard code rank try fourier basis clean up use gaussians instead of dirac delta for sump typo fix output can still check empirical stats try printing error info minor test with dirac again go back to gaussian try smoothing with tiny bandwidth separate bandwidth for gaussian sum sigma needs to be a vector minor fix this should likely fix indexing oops debugging output oops segfault fix one more fix test hard code false fix ttmetad small fixes now force true change back small fix try analytical formula for conv typo compare two methods oops missing factor typo never mind damn integer division remove all numerical integration minor fix freeze bias potential after some time typo small changes minor minor add vbias outputting for 3d bias check if vbias is the same across walkers undo oops some more useful output undo again output metad vbias also rct clean up minor comment out code instead minor lots of metad testing clean up frozen param redundant oops clean up once again oops again clean up actually keep modifications a bit longer set up non mw clean again multithreading added minor minor this is better debugging output typo more test oops more detailed test even more detail this should fix it multithreading for ttcross fix minor minor minor comment out vbias output output for 6d bias minor comment out temporarily print metad bias for each walker every now and then barrier one more barrier one more thing comment out 6d vbias output for mtd change output temporarily undo improved gaussian sketch fixes minor slight restructuring typo another fix some more debugging output yet another fix minor clean up minor vbias output for 6d oops clean up bug with kernel width add restart option temporarily comment out metad output add back oops change pace better randomization corrections minor more small changes seed rng better seeding clean up clean up major clean up add stopwatch minor try kernel basis with ttmetad oops oops again should not sketch vb with ginv temporarily uncomment vbias printing comment again improvements in kernel sketching print vbias clean up improvement in aca minor improved quadrature oops vbias printing again revert nbins improvements in sample choice for aca add pivot I/O should fix issue minor oops clean minor improvement to ttsketch minor another tweak minor some restart related tweaks minor add option to use diff basis for aca than sketch oops typo major changes to ttsketch minor fixes oops initialize files earlier try closing file before opening again oops reset each count flush and close? some debugging minor more detailed debugging output rearrange code some clean up flush and close more changes minor several other changes minor fix tiny fix more improvements weird error this should fix it improvements to restart part one more thing in restart minor revert ttcross need to initialize value ptrs debugging output empty traj if restarting from step 0 never mind stopwatch fix fix clean up file I/O for TTMetaD typo minor kernel smoothing option for TTMetaD fixes only rank 0 process should write to file minor try rewind file this should fix? this should be better oops minor roll back fix ttread debugging ttcross some more debugging minor fix better version lots of debugging output some more output oops try rewind? slight improvement further improvements to file io minor fix clean up + nbasis for ttcross minor minor fix file io debugging output oops more detailed output same oops ugh fix revert this is best i think minor enable vbias printing for ala4 option to specify kernel basis width minor fix vbias printing for aladip oops minor oops silly error disable output stopwatch stuff minor stopwatch fix experiment with no gaussians in getbias temporarily uncomment fix A0 oops indexing issue revert prepare for aladip enable stopwatch initial ttopes implementation fixes minor output vbias during opes add include vbias output for ttopes need to register outputs debugging output diff output more refined intbasis output some traj output fix? clean up print density fix add grad output minor minor nonintrusive version implemented fix segfault? debugging output matching index fix attempt more debugging output this should do it? wrong function names fix A matrices too 1 indexed? vbias printing for ala4 non negative least squares attempt namespace clash minor fixes oops minor debugging output more debugging fix more output refined debugging another fix check matrices oops output extra info minor small fix print nnls info oops typo more detailed output try positive sketch try ommitting ginv try least squares without positivity constraint declaration issue oops least squares using full matrices fixes minor try this? should be good now ok now hmm small revert clean up minor regularize gram matrix added clipping pseudoinverse of ginv try regularizing ginv adjust lambda try smaller lambda now revert to 1e-6 revert to pinv for ginv and solve ls for first core instead oops minor try fourier revert oops try kernel without gram revert again try slightly different scheme try fourier again correct bias based on gradient magnitude change grad threshold 1e3 try 100 revert bias needs to be something else at the start 1e4 1e3 revert again output some bias and gradient info step too typo new p threshold 1e-3 oops minor set up ttmetad branch itensor config file minor started cleanup comments added copyright stuff ttmetad regtest ran regtest minor Update src/ttsketch/TTMetaD.cpp assign instead of resize Co-authored-by: Daniele <Iximiel@users.noreply.github.com>
src/ttsketch/BasisFunc.cpp
Outdated
| } else if(pos % 2 == 0) { | ||
| return -pow(1 / this->L_, 1.5) * M_PI * (pos / 2) * sin(M_PI * (x - this->shift_) * (pos / 2) / this->L_); | ||
| } else { | ||
| return pow(1 / this->L_, 1.5) * M_PI * (pos / 2) * cos(M_PI * (x - this->shift_) * (pos / 2) / this->L_); |
There was a problem hiding this comment.
pos/2 if odd should be floored or should be something.5?
in case it should be something.5 use *0.5 for speed (multiplication is generally faster that division) and correctness ((3/2)=1, (3/2.0)=1.5, 3*0.5=1.5)
ad the .5 multiplication can also be used in the sin expression (and in other places in the file, check if you want a floating point result or an integer result case per case)
consider also to precompute 1/L, and/or the whole pow constant pow(1 / this->L_, 1.5) * M_PI : you will save quite a lot time
|
thanks @strandn! While we wait for the workflow to be approved I would like to ask you some cosmetic changes, (we need to update this part in the developer manual...):
|
|
The ci will not work until we setup the installation process for ITensor. please run make astyle First of all:
You should start by adding in the .ci directory a script for installing ITensor (see the arrayfire one), maybe write an ad-hoc options.mk for that. Then you should make the linux runner of the ci call it in the - name: Configure ITensor
if: ${{ contains( matrix.variant, '-mpi-' ) }}
run: |
echo "LDFLAGS=$LDFLAGS -L/path/to/ITensor/lib" >> $GITHUB_ENV
echo "CPPFLAGS=$CPPFLAGS -I/path/to/ITensor -DITENSOR_USE_HDF5 -DITENSOR_USE_OMP -D__ITENSOR_LAPACK_WRAP_h" >> $GITHUB_ENV
#temporary, the following lines will be addresses in configure.ac
echo "CXXFLAGS=\"-O3 -fconcepts\"" >> $GITHUB_ENV
echo "LIBS=\"-litensor -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lhdf5 -lhdf5_hl\"" >> $GITHUB_ENV
# this will be uncommented after the configure.ac modifications
#echo "PLUMED_CONFIG=$PLUMED_CONFIG --enable-libitensor" >> $GITHUB_ENVProbably the mkl libraries will need to be removed from the CI, since those are configuration dependent. Tell me if you need any help to setup this Edit: |
strandn
force-pushed
the
ttsketch
branch
2 times, most recently
from
b8b213a to
6e64c97
Compare
DO NOT SQUASH YOUR CHANGES ON A MERGE COMMITif you need the changes on master: rebase on master, do not merge and then squash-rebase, or if you have merged, do an interactive rebase and squash your commits, leaving alone the the merge commit if you do not have a backup branch you can reset to, make a backup branch (for example use when you change the history (like doing a rebase or something similar) always create a backup branch to reset to in case of something goes wrong |
|
@Iximiel First of all, thank you very much for your messages and your assistance so far. I am very sorry for the inconvenience. This is my first time dealing with git rebasing. I think the culprit of the issue is the fact that I have two separate local "ttsketch" branches syncing to the same remote (GitHub) branch (one in my personal laptop and the other in my university's cluster). A few days ago, I successfully rebased (I thought) from one of my local branches but since then, blindly did |
|
@strandn do not worry, git is a difficult but necessary evil (see the famous readme of the first commit of git for the opinion of Linus). A good idea for moving things around could be to trying to use a TUI or GUI for git, I use lazygit, but there are good free addons for vscode and the usual gitk. Before rebasing/resetting or any other thing that change the history, remember to backup the branch you are going to work with by creating a new one, that will also help in case of missteps. Sorry for repeating this. |
3 participants
Description
This pull request involves an implementation of tensor train (TT)-metadynamics (TT-MetaD). Details of the method can be found in https://arxiv.org/abs/2603.13549. The relevant code is in the "ttsketch" module (src/ttsketch). Instructions for setup and compilation are written in at the top of src/ttsketch/TTMetaD.cpp.
Target release
Latest possible version (master/v2.11?).
Type of contribution
Copyright
I agree to transfer the copyright of the code I have written to the PLUMED developers or to the author of the code I am modifying.
the module I added or modified contains a
COPYRIGHTfile with the correct license information. Code should be released under an open source license. I also used the commandcd src && ./header.sh mymodulenamein order to make sure the headers of the module are correct.Tests