[ENH] Parallelize SAX and PAA transformers #2980

aadya940 · 2025-07-31T09:22:28Z

Related to #2972

aeon-actions-bot · 2025-07-31T09:22:50Z

Thank you for contributing to `aeon`

I have added the following labels to this PR based on the title: [ enhancement ].
I have added the following labels to this PR based on the changes made: [ transformations ]. Feel free to change these if they do not properly represent the PR.

The Checks tab will show the status of our automated tests. You can click on individual test runs in the tab or "Details" in the panel below to see more information if there is a failure.

If our pre-commit code quality check fails, any trivial fixes will automatically be pushed to your PR unless it is a draft.

Don't hesitate to ask questions on the aeon Slack channel if you have any.

PR CI actions

These checkboxes will add labels to enable/disable CI functionality for this PR. This may not take effect immediately, and a new commit may be required to run the new configuration.

Run pre-commit checks for all files
Run mypy typecheck tests
Run all pytest tests and configurations
Run all notebook example tests
Run numba-disabled codecov tests
Stop automatic pre-commit fixes (always disabled for drafts)
Disable numba cache loading
Push an empty commit to re-run CI checks

hadifawaz1999 · 2025-07-31T11:43:31Z

Nice addition @aadya940 !

Could you run examples comparing the old and new code in terms of output accuracy ? to make sure the algorithm didnt change
Also can we benchmark the time gain ? like see how faster it becomes as plot in function of n samples length of the series etc.

aadya940 · 2025-07-31T12:11:49Z

@hadifawaz1999 Yes sure

patrickzib

Thank you for your additions. Some comments of mine.

patrickzib · 2025-07-31T12:39:20Z

aeon/transformations/collection/dictionary_based/_paa.py

+    return X_paa
+
+
+@njit(parallel=True, fastmath=True)


please add cache=True

patrickzib · 2025-07-31T12:39:33Z

aeon/transformations/collection/dictionary_based/_paa.py

@@ -143,3 +146,45 @@ def _get_test_params(cls, parameter_set="default"):
        """
        params = {"n_segments": 10}
        return params
+
+
+@njit(parallel=True, fastmath=True)


please add cache=True

patrickzib · 2025-07-31T12:41:11Z

aeon/transformations/collection/dictionary_based/_paa.py

+        for i in range(n_samples):
+            for j in range(n_channels):
+                acc = 0.0
+                for k in range(seg_len):


Why not use the .mean() here as in the original version??

numba doesn't have an implementation for mean

Not sure, I understand. Both seem to work fine (I might be missing something)

for i in range(n_samples): for j in range(n_channels): acc = X[i, j, segment].mean()

for i in range(n_samples): for j in range(n_channels): acc = np.mean(X[i, j, segment])

I meant we used axis=-1 in the original implementation, however, numba doesn't support optional arguments. Hence, implemented it this way.
:))

Thanks. I understand..

Yet, why not use one of the two alternatives presented above that do not require axis=-1?

patrickzib · 2025-07-31T12:41:20Z

aeon/transformations/collection/dictionary_based/_paa.py

+            continue  # skip empty segment
+
+        for i in range(n_samples):
+            for j in range(n_channels):


Wouldn't that spawn too many threads, given that the outer loop is with a prange?

You have used three nested prange loops in _parallel_inverse_paa_transform ?

But, in fact numba seems to ignore nested loops. I did not know this before:

Loop serialization
Loop serialization occurs when any number of prange driven loops are present inside another prange driven loop. In this case the outermost of all the prange loops executes in parallel and any inner prange loops (nested or otherwise) are treated as standard range based loops. Essentially, nested parallelism does not occur.

https://numba.pydata.org/numba-doc/dev/user/parallel.html

Yep, missed it. However, I'd also like to point out, the previously implemented numba functions in SAX use multiple nested pranges as well which is essentially dead code. I can remove it in this PR?

Not sure, who wrote these. @hadifawaz1999 ?

yes i wrote these, i dont the issue though from the discussion, SAX was working fine why is it dead code ? @patrickzib

aeon/transformations/collection/dictionary_based/_paa.py

hadifawaz1999 · 2025-07-31T16:23:28Z

nothing is dead code @aadya940 :) just probably inconsistent usage of prange, but code works fine

aadya940 · 2025-07-31T16:26:08Z

@hadifawaz1999 By dead code I meant it has no effect on the output and gives identical results in terms of speed and accuracy to if there was a single prange since numba ignores other nested prange :)

hadifawaz1999

Should be compared in terms of accuracy before runtime, and should see how worth the gain is compared to the new amount of code

hadifawaz1999 · 2025-07-31T16:27:24Z

aeon/transformations/collection/dictionary_based/_sax.py

@@ -292,3 +296,31 @@ def _invert_sax_symbols(sax_symbols, n_timepoints, breakpoints_mid):
                ]

    return sax_inverse
+
+
+@njit(fastmath=True, cache=True, parallel=True)


Not 100% sure what the gain would be of doing all this, if its not significant am not for doing this, it would mean instead of simply doing sax_symbols = np.digitize(x=X_paa, bins=self.breakpoints) we have 30+ new lines of code wiith nested loops and prange.

I have done a benchmark in the past, I think it performs little better after n_jobs > 2 and significantly better post 4 threads as compared to np.digitize

Having dealt with MONSTER datasets lately, not being able to parallelize stuff for such huge size datasets is a big downside.
It might not be significant for UCR/UEA, but can help a lot when dataset size grow.

I'm not a fan of the flatten and reshape, though, but I guess you did it to avoid the nested loop parallelism problem ?

You could add that breakspoints need to be sorted for the function to work for clarity.

You might want to check _sfa_fast.py

I use np.digitize within a parallel for-loop. This approach strikes a balance between using only explicit loops and relying solely on vectorized digitize operations.

for a in prange(dfts.shape[0]): for i in range(word_length): # range(dfts.shape[2]): words[a, : dfts.shape[1]] = ( words[a, : dfts.shape[1]] << letter_bits ) | np.digitize(dfts[a, :, i], breakpoints[i], right=True)

hadifawaz1999 · 2025-07-31T16:28:55Z

@hadifawaz1999 By dead code I meant it has no effect on the output and gives identical results in terms of speed and accuracy to if there was a single prange since numba ignores other nested prange :)

if am not mistaken numba considers a prange as a normal range when its nested, so its simply useless usage of prange nested yes

parallelize saa and pax

f1b04ce

aadya940 requested review from patrickzib, MatthewMiddlehurst and hadifawaz1999 as code owners July 31, 2025 09:22

aeon-actions-bot bot added enhancement New feature, improvement request or other non-bug code enhancement transformations Transformations package labels Jul 31, 2025

bug fix

cea39a5

aadya940 force-pushed the parallelise branch from 00eb4ac to cea39a5 Compare July 31, 2025 09:47

fix

f36f223

aadya940 force-pushed the parallelise branch from 9a95f58 to f36f223 Compare July 31, 2025 09:55

fix

6b8d269

aadya940 force-pushed the parallelise branch from b9aa22d to 6b8d269 Compare July 31, 2025 09:56

minor bug fix

6b15711

aadya940 force-pushed the parallelise branch from c565227 to 6b15711 Compare July 31, 2025 10:02

Automatic pre-commit fixes

1bf5b6b

patrickzib reviewed Jul 31, 2025

View reviewed changes

add cache=True to numba decorator

a6cc829

hadifawaz1999 requested changes Jul 31, 2025

View reviewed changes

[ENH] Parallelize SAX and PAA transformers #2980

Are you sure you want to change the base?

[ENH] Parallelize SAX and PAA transformers #2980

Uh oh!

Conversation

aadya940 commented Jul 31, 2025

Uh oh!

aeon-actions-bot bot commented Jul 31, 2025

Thank you for contributing to aeon

Uh oh!

hadifawaz1999 commented Jul 31, 2025

Uh oh!

aadya940 commented Jul 31, 2025

Uh oh!

patrickzib left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hadifawaz1999 Jul 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

hadifawaz1999 commented Jul 31, 2025

Uh oh!

aadya940 commented Jul 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hadifawaz1999 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hadifawaz1999 commented Jul 31, 2025

Uh oh!

Uh oh!

Thank you for contributing to `aeon`

hadifawaz1999 Jul 31, 2025 •

edited

Loading

aadya940 commented Jul 31, 2025 •

edited

Loading