[Fix] Isolate RNG seed generator changes and clarify documentation by ChiragSW · Pull Request #22709 · keras-team/keras

ChiragSW · 2026-04-18T19:40:20Z

Fixes part 3 of #22595

Description

Contributor Agreement

Please check all boxes below before submitting your PR for review:

I am a human, and not a bot.
I will be responsible for responding to review comments in a timely manner.
I will work with the maintainers to push this PR forward until submission.

Note: Failing to adhere to this agreement may result in your future PRs no longer being reviewed.

Isolate random seed generator and rng utility updates into a separate branch to clarify scope and rationale.

Document why default initializer seeds use a deterministic counter after set_random_seed() to address reviewer questions about context and intent.

gemini-code-assist

Code Review

This pull request introduces a mechanism to ensure that default initializer seeds are deterministic and replayable after calling keras.utils.set_random_seed(). It achieves this by maintaining a global counter for initializer seeds within rng_utils.py. The review feedback suggests simplifying the logic in make_default_seed by delegating more responsibility to the utility function and refactoring consume_default_initializer_seed to remove redundant checks and handle unseeded cases more consistently.

gemini-code-assist · 2026-04-18T19:42:15Z

+    if rng_utils.get_random_seed() is not None:
+        # Keep implicit initializer seeds deterministic after
+        # `keras.utils.set_random_seed()` calls.
+        return rng_utils.consume_default_initializer_seed()
+    return int(np.random.randint(1, int(1e9)))


The logic here can be simplified by delegating the decision to rng_utils.consume_default_initializer_seed(). This avoids duplicating the fallback logic and makes the code more maintainable while preserving the intended behavior for both seeded and unseeded runs.

# Keep implicit initializer seeds deterministic after # `keras.utils.set_random_seed()` calls. return rng_utils.consume_default_initializer_seed()

gemini-code-assist · 2026-04-18T19:42:15Z

+    seed = global_state.get_global_attribute(GLOBAL_DEFAULT_INITIALIZER_SEED)
+    if seed is None:
+        seed = get_random_seed()
+    if seed is None:
+        seed = int(np.random.randint(1, int(1e9)))
+    global_state.set_global_attribute(
+        GLOBAL_DEFAULT_INITIALIZER_SEED, int(seed) + 1
+    )
+    return int(seed)


The current implementation contains redundant checks and dead code (lines 96-97) given the current caller's guard in seed_generator.py. By unifying the logic here, we can support both seeded and unseeded cases consistently. Returning early for the unseeded case (when seed is None) preserves the 'draw every time' behavior for unseeded runs, as the global counter won't be initialized, while still providing a robust internal utility.

Suggested change

seed = global_state.get_global_attribute(GLOBAL_DEFAULT_INITIALIZER_SEED)

if seed is None:

seed = get_random_seed()

if seed is None:

seed = int(np.random.randint(1, int(1e9)))

global_state.set_global_attribute(

GLOBAL_DEFAULT_INITIALIZER_SEED, int(seed) + 1

)

return int(seed)

seed = global_state.get_global_attribute(GLOBAL_DEFAULT_INITIALIZER_SEED)

if seed is None:

seed = get_random_seed()

if seed is not None:

global_state.set_global_attribute(

GLOBAL_DEFAULT_INITIALIZER_SEED, seed + 1

)

return seed

return int(np.random.randint(1, 1_000_000_001))

codecov-commenter · 2026-04-18T20:02:46Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.03%. Comparing base (86d7dea) to head (fcebb94).
⚠️ Report is 66 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #22709      +/-   ##
==========================================
+ Coverage   75.25%   83.03%   +7.78%     
==========================================
  Files         596      596              
  Lines       68741    69911    +1170     
  Branches    10749    10900     +151     
==========================================
+ Hits        51729    58052    +6323     
+ Misses      14473     9007    -5466     
- Partials     2539     2852     +313

Flag	Coverage Δ
keras	`82.85% <100.00%> (+7.78%)`	⬆️
keras-jax	`58.35% <100.00%> (-0.93%)`	⬇️
keras-numpy	`54.23% <100.00%> (-0.72%)`	⬇️
keras-openvino	`59.77% <100.00%> (?)`
keras-tensorflow	`59.89% <100.00%> (-0.80%)`	⬇️
keras-torch	`58.71% <100.00%> (-0.76%)`	⬇️

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

☔ View full report in Codecov by Sentry.
📢 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.

ChiragSW · 2026-04-19T14:15:23Z

/gemini review

gemini-code-assist

Code Review

This pull request introduces a mechanism to ensure that default initializer seeds are deterministic and replayable following a call to keras.utils.set_random_seed(). It adds a global state tracker for these seeds and a consumption function that increments the seed value to maintain uniqueness across calls. Feedback was provided regarding the fallback logic in consume_default_initializer_seed, specifically that using np.random.randint when no Keras seed is set could introduce unwanted side effects on the global NumPy RNG state, potentially impacting users who manage their own NumPy seeds.

gemini-code-assist · 2026-04-19T14:20:16Z

+            GLOBAL_DEFAULT_INITIALIZER_SEED, seed + 1
+        )
+        return seed
+    return int(np.random.randint(1, 1_000_000_001))


The switch from python_random.randint to np.random.randint for the fallback case introduces a side effect on the global NumPy RNG state when no Keras seed is set. If a user seeds NumPy manually but not Keras, layer initialization will now consume values from their NumPy RNG, potentially breaking their own reproducibility. It is safer to use a dedicated np.random.Generator or stick to python_random to avoid interfering with the user's numerical RNG state.

ChiragSW · 2026-04-19T14:33:47Z

/gemini review

gemini-code-assist

Code Review

This pull request introduces a mechanism to ensure that default initializer seeds are deterministic and replayable after calling keras.utils.set_random_seed(). It achieves this by implementing consume_default_initializer_seed in rng_utils.py, which tracks a separate global state for these seeds, preventing them from being affected by other random draws such as those from NumPy. Feedback suggests using 2**31 - 1 as the upper bound for random seeds to maximize entropy and ensure 32-bit compatibility. Additionally, it is recommended to initialize the global seed state even in unseeded workflows to maintain consistency across subsequent calls.

gemini-code-assist · 2026-04-19T14:38:45Z

+            GLOBAL_DEFAULT_INITIALIZER_SEED, seed + 1
+        )
+        return seed
+    return random.randint(1, int(1e9))


While int(1e9) is functional, using a plain integer literal like 1_000_000_000 is more idiomatic in Python and avoids the unnecessary float conversion. Additionally, since Keras often deals with 32-bit integer seeds in backends, using 2**31 - 1 (the maximum value for a signed 32-bit integer) as the upper bound would maximize the entropy of the generated seeds while remaining safe for all backends.

Suggested change

return random.randint(1, int(1e9))

return random.randint(1, 2**31 - 1)

gemini-code-assist · 2026-04-19T14:38:45Z

+    seed = global_state.get_global_attribute(GLOBAL_DEFAULT_INITIALIZER_SEED)
+    if seed is None:
+        seed = get_random_seed()
+    if seed is not None:
+        global_state.set_global_attribute(
+            GLOBAL_DEFAULT_INITIALIZER_SEED, seed + 1
+        )
+        return seed
+    return random.randint(1, int(1e9))


The current implementation of consume_default_initializer_seed does not update the GLOBAL_DEFAULT_INITIALIZER_SEED attribute when it falls back to random.randint(1, int(1e9)). This means that in an unseeded environment, every call to this function will perform a new draw from the global Python random state.

To improve isolation and consistency even in unseeded workflows, consider initializing the global attribute with the first draw. This ensures that subsequent initializer seeds are deterministic relative to the first one, even if the global Python random state is modified elsewhere later.

Suggested change

seed = global_state.get_global_attribute(GLOBAL_DEFAULT_INITIALIZER_SEED)

if seed is None:

seed = get_random_seed()

if seed is not None:

global_state.set_global_attribute(

GLOBAL_DEFAULT_INITIALIZER_SEED, seed + 1

)

return seed

return random.randint(1, int(1e9))

seed = global_state.get_global_attribute(GLOBAL_DEFAULT_INITIALIZER_SEED)

if seed is None:

seed = get_random_seed()

if seed is None:

# Initialize the counter if no global seed is set.

seed = random.randint(1, 2**31 - 1)

global_state.set_global_attribute(

GLOBAL_DEFAULT_INITIALIZER_SEED, seed + 1

)

return seed

ChiragSW · 2026-04-19T16:23:03Z

@hertschuh please review

hertschuh

What is the context of this? What is the issue this is addressing?

ChiragSW · 2026-04-25T14:15:40Z

What is the context of this? What is the issue this is addressing?

Context: this PR was split out from #22595 after review feedback asked that the mixed PR be separated into smaller PRs. This is the “RNG seed generator changes” slice.

It is not part of the original #22472 symbolic validation bug. The purpose is a separate reproducibility cleanup: after keras.utils.set_random_seed(seed), implicit seeds assigned to random initializers should come from Keras global seed state rather than from Python’s global random stream. This makes the sequence of default initializer seeds replayable after resetting the Keras seed, and avoids coupling initializer seed allocation to unrelated RNG draws.

set_random_seed(seed) now initializes a separate default-initializer seed counter.
seed_generator.make_default_seed() consumes that counter.
unseeded behavior still falls back to a random integer.
tests cover replaying the default seed sequence after set_random_seed() and avoiding NumPy RNG advancement.

ChiragSW added 2 commits April 16, 2026 23:16

Split RNG seed generator behavior changes.

a513615

Isolate random seed generator and rng utility updates into a separate branch to clarify scope and rationale.

Clarify RNG default-seed rationale in code comments.

6f0f452

Document why default initializer seeds use a deterministic counter after set_random_seed() to address reviewer questions about context and intent.

google-ml-butler Bot added the size:S label Apr 18, 2026

google-ml-butler Bot assigned gbaned Apr 18, 2026

ChiragSW closed this Apr 18, 2026

gemini-code-assist Bot reviewed Apr 18, 2026

View reviewed changes

ChiragSW reopened this Apr 19, 2026

gemini suggested changes

8b86edc

gemini-code-assist Bot reviewed Apr 19, 2026

View reviewed changes

used random.randint

13201cb

gemini-code-assist Bot reviewed Apr 19, 2026

View reviewed changes

hertschuh reviewed Apr 23, 2026

View reviewed changes

hertschuh added the stat:awaiting response from contributor label Apr 23, 2026

google-ml-butler Bot removed the stat:awaiting response from contributor label Apr 25, 2026

added more robust test, used safe range

fcebb94

ChiragSW requested a review from hertschuh April 28, 2026 19:49

google-ml-butler Bot added the awaiting review label Apr 28, 2026

	return random.randint(1, int(1e9))
	return random.randint(1, 2**31 - 1)

Conversation

ChiragSW commented Apr 18, 2026

Description

Contributor Agreement

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

codecov-commenter commented Apr 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

ChiragSW commented Apr 19, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

ChiragSW commented Apr 19, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

ChiragSW commented Apr 19, 2026

Uh oh!

hertschuh left a comment

Choose a reason for hiding this comment

Uh oh!

ChiragSW commented Apr 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

codecov-commenter commented Apr 18, 2026 •

edited

Loading