Superseded by #85: Investigate HMM backend surface by AKuederle · Pull Request #79 · mad-lab-fau/gaitmap

AKuederle · 2026-03-09T11:39:59Z

No description provided.

AKuederle · 2026-03-09T11:41:43Z

+    """Configuration of a single trainable HMM submodule."""
+
+    name: str
+    role: Literal["transition", "stride", "other"]


why "other" should be Literal["transition", "stride"] | str

AKuederle · 2026-03-09T11:42:26Z

+
+    def __init__(
+        self,
+        modules: dict[str, HmmSubModelConfig] = cf(_default_modules()),


Dicts are bad for tpcp. Check the docs list of tuples is better

AKuederle · 2026-03-09T11:44:05Z

-    To rephrase it again: We want to create a fully labeled dataset with already optimal hidden-state labels, but as
-    these lables are hidden, we need to predict them with our already trained models...
-    """
+    """Create fully labeled gait sequences from typed regions and trained submodels."""


Expand doccstring with more info again.

AKuederle · 2026-03-09T11:46:09Z

-            Sequence of gaitmap stride lists.
-            The number of stride lists must match the number of sensordata objects (i.e. they must belong together).
+        region_list_sequence
+            Sequence of typed region lists.


Explain the expected format

AKuederle · 2026-03-09T11:48:05Z

-                architecture="left-right-strict",
-                name="stride_model",
+    @classmethod
+    def _from_json_dict(cls, json_dict: dict) -> Self:


This is done for backwards compatibility right? Add a todo note that we should remove this again and rather convert the old model to the correct format

AKuederle · 2026-03-09T11:48:46Z

+        return sum(module.n_states for module in self.model_config.modules.values())
+
+    @property
+    def module_offsets(self) -> dict[str, int]:


Not relevant for user facing -> _module_offsets

AKuederle · 2026-03-09T11:50:17Z


        # predict hidden state labels for complete walking bouts
+        module_offsets = self.module_offsets
        labels_train_sequence = create_fully_labeled_gait_sequences(


Should we rename create_fully_labeled_gait_sequences?

AKuederle · 2026-03-09T11:52:42Z

+    return normalized_region_list
+
+
+def convert_region_list_to_transition_list(


Can we find a more performant version for this? I think we have a couple of helper funcs in utisl that can help here.

AKuederle · 2026-03-09T11:53:54Z

+
+def _stride_list_to_region_list(stride_list: pd.DataFrame, region_type: str = "stride") -> pd.DataFrame:
+    region_list = stride_list[["start", "end"]].copy()
+    region_list.insert(0, "roi_id", np.arange(len(region_list)))


Just rename stride_id (the index to roi id)

AKuederle · 2026-03-09T19:37:49Z

+    return log_emissions
+
+
+def _viterbi_decode(model: HMMState, log_emissions: np.ndarray) -> np.ndarray:


Can we use numba to speed this and other funcs up here?

AKuederle · 2026-03-09T19:38:38Z

+    return np.argmax(posterior, axis=1)
+
+
+class ScipyHmmInferenceBackend(BaseHmmBackend):


Move the scipy backend in its own file, so that we can import it even when pomegrante isnot installed

AKuederle · 2026-03-09T19:40:03Z

+region_list_sequence = []
+for sensor in ["left_sensor", "right_sensor"]:
+    region_list = stride_list[sensor][["start", "end"]].copy()
+    region_list.insert(0, "roi_id", np.arange(len(region_list)))


use assign and reuse the stride index/stride id

AKuederle · 2026-03-09T19:41:08Z

+    return tuple(transitions)
+
+
+class PomegranateHmmBackend(BaseHmmBackend):


Move to dedicated file to allow imports of base and scipy backend when pomegrante not installed

AKuederle · 2026-03-09T19:42:50Z

+        model = fix_model_names(model)
+        model.bake()
+        model.freeze_distributions()
+        return _clone_model(model, assert_correct=False)


Do we need to clone, given that we just created the model here?

AKuederle · 2026-03-09T19:46:11Z

+        return tuple(module.name for module in self.modules if module.role == "stride")
+
+    @property
+    def custom_model_names(self) -> tuple[str, ...]:


AKuederle · 2026-03-09T19:47:20Z

+        model_config: CompositeHmmConfig = cf(CompositeHmmConfig()),
+        feature_transform: RothHmmFeatureTransformer = cf(RothHmmFeatureTransformer()),
+        *,
+        algo_predict: Literal["viterbi", "map"] = "viterbi",


Wouldn't it make more sense to have the training params on teh composite config instead of here?

AKuederle · 2026-03-09T19:48:57Z

+
+
+class RothHmmConfig(_BaseSerializable):
+    """Serializable configuration bundle for `RothSegmentationHmm`."""


Docstring should explain all parameters (Same for all other classes in teh config module)

AKuederle · 2026-03-09T19:53:32Z

+        self.backend = backend
+
+    @property
+    def model_config(self) -> CompositeHmmConfig:


Remove all the roperty mappings. user should just access everything through hmm_config

AKuederle · 2026-03-09T20:10:39Z

+    return model
+
+
+def pomegranate_model_to_hmm_state(


move to pomegrante_legacy_backend file

codecov-commenter · 2026-03-09T20:42:11Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.95%. Comparing base (3c6e508) to head (ae62c36).
⚠️ Report is 2 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #79      +/-   ##
==========================================
+ Coverage   89.93%   89.95%   +0.01%     
==========================================
  Files          61       61              
  Lines        3438     3424      -14     
==========================================
- Hits         3092     3080      -12     
+ Misses        346      344       -2

☔ 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.

…igate/hmm-backend-surface

AKuederle · 2026-07-17T11:51:27Z

Superseded by #85, which implements the backend-neutral HMM model, swappable inference/training backends, reusable topology composition, and legacy JSON loading. Closing this earlier approach in favor of the completed replacement.

AKuederle added 4 commits March 9, 2026 11:21

Move project skill into .agent

e0ff322

Document HMM backend refactor direction

4f0dcb6

Move HMM refactor note out of docs

8ca5d8c

Refactor HMM config input surface

def5317

AKuederle commented Mar 9, 2026

View reviewed changes

AKuederle added 6 commits March 9, 2026 13:03

Address HMM config review feedback

5bb0e61

Add HMM state and pomegranate backend abstraction

8fb574d

Bundle Roth HMM preprocessing into config

6101b6d

Drop legacy clone mixin from Roth HMM

db24183

Add SciPy inference backend for HMM states

db3985a

Fix Roth HMM example and migration checks

a51d326

AKuederle commented Mar 9, 2026

View reviewed changes

AKuederle added 17 commits March 11, 2026 08:48

Fix legacy fused HMM end probabilities

25909f3

Repair pretrained Roth HMM terminal probabilities

c8a49bc

Refactor HMM backends into dedicated packages

48a14f2

Limit HMM backend imports to backend packages

b7b75b8

Widen optional pomegranate dependency range

cdde465

Apply repo formatting and lint cleanup

d865b62

Reload HMM package with backend selection tests

95991b6

Keep legacy HMM annotations compatible with Python 3.9

fdaaa0a

Align HMM backend imports and inference regression tests

83fc01b

Stabilize HMM training example snapshots

775ebb7

Simplify HMM test coverage

70fe6f9

Fix HMM lint issues

501b3cf

Merge branch 'investigate/hmm-pomegranate-modern-backend' into invest…

a1474e3

…igate/hmm-backend-surface

Install HMM extras in CI test jobs

4231bb1

Install all extras in CI test jobs

d9e2796

Fix modern pomegranate compatibility

104778b

Remove raw legacy HMM serialization

446f7e0

AKuederle added 2 commits March 11, 2026 14:31

Tighten legacy HMM import boundaries

ae62c36

Proper uv layout

58f2e9e

AKuederle changed the title ~~Investigate/hmm backend surface~~ Superseded by #85: Investigate HMM backend surface Jul 17, 2026

AKuederle closed this Jul 17, 2026

		return normalized_region_list


		def convert_region_list_to_transition_list(

		return log_emissions


		def _viterbi_decode(model: HMMState, log_emissions: np.ndarray) -> np.ndarray:

		return np.argmax(posterior, axis=1)


		class ScipyHmmInferenceBackend(BaseHmmBackend):

		return tuple(transitions)


		class PomegranateHmmBackend(BaseHmmBackend):



		class RothHmmConfig(_BaseSerializable):
		"""Serializable configuration bundle for `RothSegmentationHmm`."""

Uh oh!

Conversation

AKuederle commented Mar 9, 2026

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!

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!

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!

codecov-commenter commented Mar 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

AKuederle commented Jul 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

codecov-commenter commented Mar 9, 2026 •

edited

Loading