Add hamming, jensen-shannon, kl-divergence, correlation and russellrao distance metrics #4155

mdoijade · 2021-08-11T15:08:57Z

-- This PR depends on RAFT PR - rapidsai/raft#306
-- Adds cpp & python interfaces for these distance metrics with pytest support for each of them.
-- also remove redundant commented code in canberra distance metric

…ivergence and russell-rao distance prims

…n headers

…rking except some bug in kldivergence

mdoijade · 2021-08-11T18:06:50Z

@dantegd the cuml build looks to be broken due to this error below, is there a PR which is working on fixing this?
This blocks my PR.
@teju85 @cjnolet for help with reviewing this cuML side PR.

https://gpuci.gpuopenanalytics.com/job/rapidsai/job/gpuci/job/cuml/job/prb/job/cuml-cpu-cuda-build/CUDA=11.2/1452/console
3:31:26 ../bench/prims/gram_matrix.cu(61): error: identifier "CUBLAS_CHECK_NO_THROW" is undefined
23:31:26 detected during:
23:31:26 instantiation of "MLCommon::Bench::Matrix::GramMatrix::GramMatrix(const std::string &, const MLCommon::Bench::Matrix::GramTestParams &) [with T=float]"
23:31:26 ../bench/common/ml_benchmark.hpp(197): here
23:31:26 instantiation of "MLCommon::Bench::internal::Registrar<Params, Class>::Registrar(const std::vector<Params, std::allocator> &, const std::string &, const std::string &) [with Params=MLCommon::Bench::Matrix::GramTestParams, Class=MLCommon::Bench::Matrix::GramMatrix]"
23:31:26 (137): here
23:31:26
23:31:26 1 error detected in the compilation of "../bench/prims/gram_matrix.cu".

mdoijade · 2021-08-11T18:41:44Z

seems that this PR introduced this build failure - e977f3e
@harrism

harrism · 2021-08-12T00:33:35Z

@mdoijade CUBLAS_CHECK_NO_THROW was enabled by rapidsai/raft#311, on which the PR linked depended. That PR had to be merged before e977f3e would pass CI.

mdoijade · 2021-08-12T08:02:04Z

@mdoijade CUBLAS_CHECK_NO_THROW was enabled by rapidsai/raft#311, on which the PR linked depended. That PR had to be merged before e977f3e would pass CI.

I see. thanks! merged it in my RAFT PR now.

mdoijade · 2021-08-12T18:13:56Z

@dantegd all the test runs are failing with some CI error - https://gpuci.gpuopenanalytics.com/job/rapidsai/job/gpuci/job/cuml/job/prb/job/cuml-cpu-python-build/CUDA=11.2,PYTHON=3.7/1239/console

Also can we merge the raft PR now, as I believe it shouldn't break cuml builds as here all builds looks fine ? - rapidsai/raft#306

GPUtester · 2021-08-23T15:28:08Z

Can one of the admins verify this patch?

mdoijade · 2021-08-24T08:31:25Z

@dantegd all the test runs are failing with some CI error, is it due to using custom raft fork ?

https://gpuci.gpuopenanalytics.com/job/rapidsai/job/gpuci/job/cuml/job/prb/job/cuml-gpu-test/CUDA=11.0,GPU_LABEL=gpu-a100,LINUX_VER=centos7,PYTHON=3.7/1254/console

00:29:19 Cloning into '_external_repositories/raft'...
00:29:20 error: pathspec 'branch-21.10' did not match any file(s) known to git
00:29:20 -- CUDA_HOME detected with value: /usr/local/cuda
00:29:20 -- CUML_BUILD_PATH environment variable not set.
00:29:20 -- RAFT_PATH environment variable not set.
00:29:20 -- Third party repositories have not been found so they will be cloned. To avoid this set the environment variable CUML_BUILD_PATH, containing the absolute path to the build folder where libcuml++ was built.
00:29:20 Cloning repository raft into _external_repositories/raft
00:29:20 Traceback (most recent call last):
00:29:20 File "setup.py", line 108, in
00:29:20 raft_include_dir = use_raft_package(raft_path, libcuml_path)
00:29:20 File "/workspace/python/setuputils.py", line 100, in use_raft_package
00:29:20 git_info_file=git_info_file)
00:29:20 File "/workspace/python/setuputils.py", line 132, in clone_repo_if_needed
00:29:20 git_info_file=git_info_file)
00:29:20 File "/workspace/python/setuputils.py", line 193, in get_submodule_dependency
00:29:20 clone_repo(repo, repo_info[repo][0], repo_info[repo][1])
00:29:20 File "/workspace/python/setuputils.py", line 230, in clone_repo
00:29:20 GIT_TAG])
00:29:20 File "/opt/conda/envs/rapids/lib/python3.7/subprocess.py", line 363, in check_call
00:29:20 raise CalledProcessError(retcode, cmd)
00:29:20 subprocess.CalledProcessError: Command '['git', 'checkout', 'branch-21.10']' returned non-zero exit status 1.
00:29:20 Build step 'Execute shell' marked build as failure
00:29:21 Recording test results
00:29:24 [Checks API] No suitable checks publisher found.

cjnolet

I have a question and a reminder to change the raft branch back, otherwise this looks good.

cjnolet · 2021-08-25T22:57:44Z

cpp/cmake/thirdparty/get_raft.cmake

-            GIT_REPOSITORY https://github.com/${PKG_FORK}/raft.git
-            GIT_TAG        ${PKG_PINNED_TAG}
+            GIT_REPOSITORY https://github.com/mdoijade/raft.git
+            GIT_TAG         additionalDistPrims


Reminder to update this now that the raft side has been merged

cjnolet · 2021-08-25T22:59:24Z

cpp/src/metrics/pairwise_distance_correlation.cu

+
+  // Call the distance function
+  switch (metric) {
+    case raft::distance::DistanceType::CorrelationExpanded:


Is the switch state propagated into this in case there is also an unexpanded version in the future? (Asking for the other distances as well)

No there is no unexpanded version planned for this or other distance metrics. should I replace switch() with an if condition ?

These functions are named after the associated distance measure so I would call a function named pairwise_distance_correlation expecting it to compute the correlation distance. If it's not possible that this will ever compute a distance other than correlation, why are we passing the DistanceType into these functions at all? Rather than using an if condition, I would propose removing the metric argument altogether and just having the function compute the distance for which its name implies. What do you think?

Good point! I think the metric arg got propagated to all distances due to euclidean dist having multiple sub-implementations.
I've removed this redundant check & arg now for all non-euclidean metrics which doesn't have multiple implementations.

mdoijade · 2021-08-26T09:44:25Z

@teju85 can I get a rerun of tests here?
PR is hitting some CI errors below,

json.decoder.JSONDecodeError: Unterminated string starting at: line 2407838 column 18 (char 72712781)

Traceback (most recent call last):
File "/opt/conda/envs/rapids/lib/python3.7/site-packages/conda/core/subdir_data.py", line 705, in fetch_repodata_remote_request
resp.raise_for_status()
File "/opt/conda/envs/rapids/lib/python3.7/site-packages/requests/models.py", line 953, in raise_for_status
raise HTTPError(http_error_msg, response=self)
requests.exceptions.HTTPError: 504 Server Error: Gateway Time-out for url: https://conda.anaconda.org/rapidsai/linux-64/repodata.json

teju85 · 2021-08-27T08:52:24Z

rerun tests

…contains single distance metric implementation

cjnolet

Thanks, Majesh. The changes look great! Just pending CI now.

mdoijade · 2021-09-01T15:23:35Z

@dantegd before merging branch-21.10 recent changes in my PR all pytests related to new distance metrics were passing,
but since inclusion of I guess this change 903d576
It has started failing in CI.
locally I am unable to reproduce these failures using recent raft branch-21.10 and changes from this cuML PR.
I see that in CI results all the values from raft kernel is getting reported as 0's so I guess the gpu kernel is not getting called, but I am not sure why.
@viclafargue to help check the changes done to pairwise_distance_correlation.cu in this PR if it is as expected.

@teju85 for visibility.

viclafargue · 2021-09-06T11:55:01Z

@dantegd before merging branch-21.10 recent changes in my PR all pytests related to new distance metrics were passing,
but since inclusion of I guess this change 903d576
It has started failing in CI.
locally I am unable to reproduce these failures using recent raft branch-21.10 and changes from this cuML PR.
I see that in CI results all the values from raft kernel is getting reported as 0's so I guess the gpu kernel is not getting called, but I am not sure why.
@viclafargue to help check the changes done to pairwise_distance_correlation.cu in this PR if it is as expected.

@teju85 for visibility.

Sorry for the late reply, I was on vacation. Thank you for working on this. I took a look at this PR and the RAFT changes. It's hard to tell at the moment which change might cause the issue. However, I could reproduce the issue locally. For this, you may need to either update your environment or to remove RAFT package from your conda env (get_rmm.cmake file won't be used for cpp if RAFT is present in env). ~~I also noticed that the final commit in rapidsai/raft#306 containing the merge doesn't appear to be fully CI approved. Has it been tested some other way?~~ Sorry, that's the RAFT CI issue.

viclafargue · 2021-09-07T16:46:43Z

Seems like the issue comes from the libcumlprims package that is shipped with cmake files that refer to an older version of RAFT. The changes in RAFT that go hand-in-hand with this PR are absent. cc @mdoijade @dantegd

dantegd · 2021-09-08T15:00:33Z

rerun tests

codecov-commenter · 2021-09-08T20:19:16Z

Codecov Report

❗ No coverage uploaded for pull request base (branch-21.10@0e770fa). Click here to learn what that means.
The diff coverage is n/a.

@@               Coverage Diff               @@
##             branch-21.10    #4155   +/-   ##
===============================================
  Coverage                ?   85.98%           
===============================================
  Files                   ?      231           
  Lines                   ?    18515           
  Branches                ?        0           
===============================================
  Hits                    ?    15920           
  Misses                  ?     2595           
  Partials                ?        0

Flag	Coverage Δ
dask	`47.31% <0.00%> (?)`
non-dask	`78.59% <0.00%> (?)`

Flags with carried forward coverage won't be shown. Click here to find out more.

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0e770fa...6427584. Read the comment docs.

mdoijade · 2021-09-14T08:07:22Z

can this PR get merged now?

dantegd · 2021-09-14T13:16:56Z

@gpucibot merge

…o distance metrics (rapidsai#4155) -- This PR depends on RAFT PR - rapidsai/raft#306 -- Adds cpp & python interfaces for these distance metrics with pytest support for each of them. -- also remove redundant commented code in canberra distance metric Authors: - Mahesh Doijade (https://github.com/mdoijade) Approvers: - Corey J. Nolet (https://github.com/cjnolet) URL: rapidsai#4155

mdoijade added 7 commits August 7, 2021 00:04

Add interfaces in cuml for hamming, correlation, jensen-shannon, kl-d…

dfc28a4

…ivergence and russell-rao distance prims

fix clang formatting issues

ad67053

add all new distances to main API and fix function name in correlatio…

38729dc

…n headers

add python interfaces for all new dist metrics, with tests for all wo…

0a734fd

…rking except some bug in kldivergence

add test support for kl-divergence dist metric

aa072ac

Merge branch 'branch-21.10' into additionalDistPrims

38e1067

pin mdoijade raft fork for testing change

6dd14bf

mdoijade requested review from a team as code owners August 11, 2021 15:08

github-actions bot added CMake CUDA/C++ Cython / Python Cython or Python issue labels Aug 11, 2021

fix flake formating issues in test_metrics

911eef8

temp commit to trigger ci

ae9f7a2

mdoijade added 3 commits August 12, 2021 13:33

temp commit to trigger ci

5fc77f7

temp commit to trigger ci to check updated raft changes

90c8f16

temp commit to trigger ci to check updated raft changes

1d5c966

mdoijade added 2 commits August 13, 2021 09:18

Merge branch 'branch-21.10' into additionalDistPrims

8225ee0

temp commit to test new raft commits

3f6de71

Merge branch 'branch-21.10' into additionalDistPrims

19d44db

cjnolet requested changes Aug 25, 2021

View reviewed changes

revert raft mdoijade fork as raft PR is merged now

b71866a

Merge branch 'branch-21.10' into additionalDistPrims

3ef11bc

teju85 added improvement Improvement / enhancement to an existing function non-breaking Non-breaking change labels Aug 27, 2021

remove redundant metric arg and switch based on it on the APIs which …

1a1dd71

…contains single distance metric implementation

cjnolet approved these changes Aug 31, 2021

View reviewed changes

mdoijade added 4 commits August 31, 2021 21:52

merge branch-21.10

96c0240

Add udevice_vector changes to new distances

675b661

fix clang format issues

ac4f067

Merge branch 'branch-21.10' into additionalDistPrims

8e2f01d

Merge branch 'branch-21.10' into additionalDistPrims

6427584

dantegd added feature request New feature or request and removed improvement Improvement / enhancement to an existing function labels Sep 14, 2021

rapids-bot bot merged commit 55ddc8b into rapidsai:branch-21.10 Sep 14, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add hamming, jensen-shannon, kl-divergence, correlation and russellrao distance metrics #4155

Add hamming, jensen-shannon, kl-divergence, correlation and russellrao distance metrics #4155

mdoijade commented Aug 11, 2021

mdoijade commented Aug 11, 2021

mdoijade commented Aug 11, 2021

harrism commented Aug 12, 2021

mdoijade commented Aug 12, 2021

mdoijade commented Aug 12, 2021

GPUtester commented Aug 23, 2021

mdoijade commented Aug 24, 2021

cjnolet left a comment

cjnolet Aug 25, 2021

cjnolet Aug 25, 2021

mdoijade Aug 26, 2021

cjnolet Aug 27, 2021

mdoijade Aug 30, 2021

mdoijade commented Aug 26, 2021

teju85 commented Aug 27, 2021

cjnolet left a comment •

edited

Loading

mdoijade commented Sep 1, 2021

viclafargue commented Sep 6, 2021 •

edited

Loading

viclafargue commented Sep 7, 2021

dantegd commented Sep 8, 2021

codecov-commenter commented Sep 8, 2021

mdoijade commented Sep 14, 2021

dantegd commented Sep 14, 2021

Add hamming, jensen-shannon, kl-divergence, correlation and russellrao distance metrics #4155

Add hamming, jensen-shannon, kl-divergence, correlation and russellrao distance metrics #4155

Conversation

mdoijade commented Aug 11, 2021

mdoijade commented Aug 11, 2021

mdoijade commented Aug 11, 2021

harrism commented Aug 12, 2021

mdoijade commented Aug 12, 2021

mdoijade commented Aug 12, 2021

GPUtester commented Aug 23, 2021

mdoijade commented Aug 24, 2021

cjnolet left a comment

Choose a reason for hiding this comment

cjnolet Aug 25, 2021

Choose a reason for hiding this comment

cjnolet Aug 25, 2021

Choose a reason for hiding this comment

mdoijade Aug 26, 2021

Choose a reason for hiding this comment

cjnolet Aug 27, 2021

Choose a reason for hiding this comment

mdoijade Aug 30, 2021

Choose a reason for hiding this comment

mdoijade commented Aug 26, 2021

teju85 commented Aug 27, 2021

cjnolet left a comment • edited Loading

Choose a reason for hiding this comment

mdoijade commented Sep 1, 2021

viclafargue commented Sep 6, 2021 • edited Loading

viclafargue commented Sep 7, 2021

dantegd commented Sep 8, 2021

codecov-commenter commented Sep 8, 2021

Codecov Report

mdoijade commented Sep 14, 2021

dantegd commented Sep 14, 2021

cjnolet left a comment •

edited

Loading

viclafargue commented Sep 6, 2021 •

edited

Loading