Add ClavaDDPM to SDGym multi table synthesizers by sarahmish · Pull Request #630 · sdv-dev/SDGym

sarahmish · 2026-07-22T22:12:51Z

original implementation is https://github.com/weipang142857/ClavaDDPM

The important class to review is ClavaDDPM and ClavaDDPMSynthesizer as these are the wrappers I wrote. I've written other code pieces, if it's taken from the source code directly, I mention "from" in the docstrings.

codecov · 2026-07-22T22:16:08Z

Codecov Report

❌ Patch coverage is 85.47401% with 95 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.03%. Comparing base (410fef0) to head (d1340b2).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
sdgym/synthesizers/clavaddpm.py	85.33%	95 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #630      +/-   ##
==========================================
+ Coverage   86.86%   87.03%   +0.17%     
==========================================
  Files          41       42       +1     
  Lines        4409     5061     +652     
==========================================
+ Hits         3830     4405     +575     
- Misses        579      656      +77

Flag	Coverage Δ
integration	`56.11% <83.33%> (+4.42%)`	⬆️
unit	`76.64% <32.72%> (-6.51%)`	⬇️

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

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

pvk-developer

When fitting I also see the steps (they come from TabDDPM), if verbose is set to False do we need to display them?

pvk-developer · 2026-07-31T16:54:05Z

During fit there is a point where MLoss and GLoss start to produce nan values, should we stop there for that table?

Step 9000/25000 MLoss: 2.7555 GLoss: 0.9242 Sum: 3.6797
Step 9500/25000 MLoss: nan GLoss: nan Sum: nan
...
Step 24500/25000 MLoss: nan GLoss: nan Sum: nan
Step 25000/25000 MLoss: nan GLoss: nan Sum: nan

sarahmish · 2026-08-03T13:01:31Z

During fit there is a point where MLoss and GLoss start to produce nan values, should we stop there for that table?

I lean towards keeping it as faithful to the original implementation as possible. Sometimes nan indicates that this model doesn't work for a particular dataset on this synthesizer.

R-Palazzo · 2026-08-10T14:36:24Z

+
+    LOGGER = logging.getLogger(__name__)
+    _MODEL_KWARGS = None
+    _MODALITY_FLAG = 'multi_table'


You could remove this line since this is already set in the MultiTableBaselineSynthesizer:

SDGym/sdgym/synthesizers/base.py

Line 142 in 410fef0

_MODALITY_FLAG = 'multi_table'

R-Palazzo

This is looking good!

I’m still reviewing the logic and checking that it aligns with their implementation. In the meantime, I’m thinking about two things:

For the structure, we could either:

Define a clavaddpm folder where we keep files such as pipeline_modules.py and preprocess_utils.py, along with our SDGym logic.
Keep everything in one file, but be more explicit about what comes from their implementation versus what we added. It could look something like:

######################### SDGym extra logic #########################

def get_group_data_dict(np_data, group_id_attrs=[0]):
    """Grouping dictionary from pipeline_utils.py."""
    group_data_dict = {}
    data_len = len(np_data)
    for i in range(data_len):
        row_id = tuple(np_data[i, group_id_attrs])
        if row_id not in group_data_dict:
            group_data_dict[row_id] = []
        group_data_dict[row_id].append(np_data[i])
    return group_data_dict


######################## From preprocess_utils.py ###################

def get_domain(df, id_cols, discrete_cols):
    """Build the ``{col: {'size', 'type'}}`` domain of a table from preprocess_utils.py."""
    domain = {}
    for col in df.columns:
        if col in discrete_cols:
            domain[col] = {'size': len(df[col].unique()), 'type': 'discrete'}
        elif col not in id_cols:
            domain[col] = {'size': len(df[col].unique()), 'type': 'continuous'}

    return domain
.
.
.
######################## From pipeline_modules.py ###################
.
.
.

Would it be possible to have an integration test that runs the SDGym ClavaDDPM implementation against their original implementation and checks that the results are the same or similar? It would be nice to have, although I’m not sure how feasible it is. It might require adding their repo as a test dependency.

R-Palazzo · 2026-08-11T12:43:54Z

+
+
+def match_tables(A, B, n_clusters=25, unique_matching=True, batch_size=100):
+    """Nearest-neighbour match of every row of ``A`` to a row of ``B``.


We could add from pipeline_utils.py

R-Palazzo · 2026-08-11T12:44:31Z

+    batch_size=100,
+    no_matching=False,
+):
+    """Reconcile a child generated once per parent into a single table."""


Maybe also add from pipeline_utils.py

R-Palazzo · 2026-08-11T13:38:37Z

+        meta = metadata if isinstance(metadata, dict) else metadata.to_dict()
+
+        self._table_names = list(meta['tables'])
+        self._primary_key = {}


Also define in l804

R-Palazzo · 2026-08-12T14:41:37Z

+        self.matching_batch_size = matching_batch_size
+        self.unique_matching = unique_matching
+        self.no_matching = no_matching
+        self.max_categories = max_categories


This seems to never be used, should we get rid of it?

R-Palazzo · 2026-08-12T14:48:26Z

+                chosen, chosen_distance = int(candidate), float(distance)
+                break
+        if chosen is None:
+            chosen = next(j for j in range(len(B)) if j not in used)


Should we also update chosen_distance here?

R-Palazzo · 2026-08-12T14:49:12Z

+
+    Mirrors ``GaussianMultinomialDiffusion._sample`` but takes explicit
+    per-row labels instead of drawing them from the empirical distribution.
+    Similar to the original classifier-guided ``conditional_sample`` implemtation.


Suggested change

Similar to the original classifier-guided ``conditional_sample`` implemtation.

Similar to the original classifier-guided ``conditional_sample`` implementation.

R-Palazzo · 2026-08-12T14:59:30Z

+    encoder's fitted range get fresh values: numeric ids continue after the
+    largest original id, other ids become new ``'{column}_{code}'`` strings.


Do we need to handle the numerical case here?

In l875 and l883 we're skipping the encoder for numerical ids

R-Palazzo · 2026-08-12T15:01:12Z

+                parent_num_cols.append((col_index, col))
+
+    parent_primary_key_index = original_parent_cols.index(parent_primary_key)
+    foreing_key_index = original_child_cols.index(foreign_key)


Suggested change

foreing_key_index = original_child_cols.index(foreign_key)

foreign_key_index = original_child_cols.index(foreign_key)

R-Palazzo · 2026-08-12T15:06:58Z

+            parent_scale * cat_one_hot[:, joint_cat_matrix_p_index:]
+        )
+
+    # Perform quantile normalization using QuantileTransformer


We removed num_quantile because it was not used?
https://github.com/weipang142857/ClavaDDPM/blob/fa6ef4d7c6d2c584b45a398c4471bb51e0bb5e17/pipeline_modules.py#L243

Maybe we should update the comment

R-Palazzo · 2026-08-12T15:14:54Z

+    num_clusters = min(num_clusters, len(cluster_data))
+
+    init_param = 'k-means++'
+    if SKLEARN_VERSION.release[:2] < (1, 1):


n_init='auto' is also impacted by the sklearn version. It was introduced in scikit-learn 1.2:

https://scikit-learn.org/stable/whats_new/v1.2.html

At some point maybe we could just bump the minimum version for sklearn

sarahmish added 7 commits July 16, 2026 12:43

wip

21bb7e8

add tabddpm for single-table synthesis

4be6278

update synthesizer list test

f18a5ca

update code

431e053

add clavaddpm and tests

68feb7f

add benchmark_multi_table test

a185358

fix lint

f71f40a

sarahmish requested a review from a team as a code owner July 22, 2026 22:12

sarahmish requested review from R-Palazzo and amontanez24 and removed request for a team July 22, 2026 22:12

sarahmish self-assigned this Jul 22, 2026

sarahmish added 3 commits July 27, 2026 10:48

fix sklearn onhotencoding version

21d0094

change init_param based on sklearn version

6ef5fa4

kmeans

e3513c2

sarahmish requested a review from frances-h July 27, 2026 17:18

add space

f312993

sarahmish requested a review from pvk-developer July 29, 2026 18:21

create primary key if none exist

8ce08a5

pvk-developer reviewed Jul 31, 2026

View reviewed changes

Comment thread sdgym/synthesizers/clavaddpm.py Outdated

Comment thread sdgym/synthesizers/clavaddpm.py Outdated

Comment thread sdgym/synthesizers/clavaddpm.py Outdated

Comment thread sdgym/synthesizers/clavaddpm.py Outdated

Comment thread sdgym/synthesizers/clavaddpm.py Outdated

add id column to discrete

bc076e2

sarahmish removed the request for review from frances-h August 3, 2026 11:31

address comments

288b95b

sarahmish requested a review from pvk-developer August 3, 2026 13:01

Base automatically changed from tabddpm to main August 3, 2026 13:55

Merge branch 'main' into clavaddpm

8224ad1

pvk-developer approved these changes Aug 3, 2026

View reviewed changes

amontanez24 reviewed Aug 3, 2026

View reviewed changes

Comment thread sdgym/synthesizers/clavaddpm.py

Comment thread sdgym/synthesizers/clavaddpm.py Outdated

sarahmish added 2 commits August 5, 2026 01:10

change to in-house datetime guesser

0f05bf7

import guess_datetime_format depending on version

72ecee0

sarahmish requested a review from amontanez24 August 5, 2026 13:07

pvk-developer reviewed Aug 5, 2026

View reviewed changes

Comment thread sdgym/synthesizers/clavaddpm.py

amontanez24 approved these changes Aug 5, 2026

View reviewed changes

sarahmish added 2 commits August 6, 2026 11:32

cap categorical columns to 100

29fee25

cast mixed formats to be uniform

b201719

pvk-developer approved these changes Aug 6, 2026

View reviewed changes

R-Palazzo approved these changes Aug 10, 2026

View reviewed changes

move preprocessing to fit only

8048265

R-Palazzo reviewed Aug 12, 2026

View reviewed changes

skip validation on tabddpm if run from clavaddpm

d1340b2

R-Palazzo self-requested a review August 17, 2026 12:47



		def match_tables(A, B, n_clusters=25, unique_matching=True, batch_size=100):
		"""Nearest-neighbour match of every row of ``A`` to a row of ``B``.

	Similar to the original classifier-guided ``conditional_sample`` implemtation.
	Similar to the original classifier-guided ``conditional_sample`` implementation.

		encoder's fitted range get fresh values: numeric ids continue after the
		largest original id, other ids become new ``'{column}_{code}'`` strings.

	foreing_key_index = original_child_cols.index(foreign_key)
	foreign_key_index = original_child_cols.index(foreign_key)

Uh oh!

Conversation

sarahmish commented Jul 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Jul 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

pvk-developer left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pvk-developer commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

sarahmish commented Aug 3, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

R-Palazzo left a comment • 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!

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

sarahmish commented Jul 22, 2026 •

edited

Loading

codecov Bot commented Jul 22, 2026 •

edited

Loading

pvk-developer commented Jul 31, 2026 •

edited

Loading

R-Palazzo left a comment •

edited

Loading