Support scalar-like analytics values

[YuanTingHsieh]

What changed

• Normalize scalar-like analytics values when constructing AnalyticsData.
• Accept Python int/float values as before, plus scalar-like objects with .item() such as numpy.float32 and 0-D numpy arrays.
• Normalize numeric values inside SCALARS and METRICS dictionaries as well.
• Add unit coverage for generic scalar-like objects and numpy scalar/0-D array values.

Why

Client tracking writers currently reject values such as numpy.float32 for single scalar metrics because analytics validation only accepts built-in Python float/int. This can happen with code like:

running_loss += cost.cpu().detach().numpy() / images.size()[0]

Centralizing the normalization in AnalyticsData keeps SummaryWriter, MLflowWriter, WandBWriter, and lower-level analytics logging paths consistent without adding a numpy dependency to core API code.

[Copilot]

Pull request overview

This PR improves analytics logging ergonomics by normalizing “scalar-like” numeric values (e.g., objects with .item() such as NumPy scalars and 0-D arrays) when constructing AnalyticsData, so downstream writers receive built-in Python int/float consistently.

Changes:

• Normalize scalar-like values for SCALAR/METRIC analytics types during AnalyticsData validation.
• Normalize numeric values inside SCALARS/METRICS dict payloads.
• Add unit tests covering generic scalar-like objects and NumPy scalar / 0-D array inputs.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
nvflare/apis/analytix.py	Adds numeric scalar normalization logic during validation and applies it to scalar and dict-based analytics payloads.
tests/unit_test/apis/analytix_test.py	Adds test coverage for scalar-like normalization, including NumPy scalars and 0-D arrays (skipped if NumPy is unavailable).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[greptile-apps]

Greptile Summary

This PR centralizes normalization of scalar-like analytics values inside AnalyticsData, accepting not just Python float/int but also any object exposing a scalar .item() method (e.g. numpy.float32, 0-D numpy arrays) for SCALAR, METRIC, SCALARS, and METRICS data types.

• A new _normalize_numeric_scalar static method uses duck-typing (.item() + .shape == ()) to extract a Python float/int from scalar-like objects without importing numpy.
• _validate_data_types now returns the (possibly normalized) value, which the constructor assigns to self.value, keeping all writers consistent.
• Unit tests cover a generic ScalarLike stub, numpy scalars, and 0-D numpy arrays for both single-scalar and dict-of-scalars cases.

Confidence Score: 4/5

Safe to merge; the normalization logic is sound and the existing test suite is largely preserved.

The core normalization path is correct and well-tested for the primary use cases. The main gap is that non-normalizable values inside a SCALARS/METRICS dict are silently kept rather than rejected, so a mistyped dict entry would slip past _validate_data_types and only fail later inside a writer. Additionally, the new item() path means numpy booleans are now accepted as scalars where they were previously rejected — a minor but undocumented behavioral expansion.

Pay closest attention to the elif data_type in [AnalyticsDataType.METRICS, AnalyticsDataType.SCALARS] normalization loop in nvflare/apis/analytix.py (lines 197-202) and the corresponding test coverage gap for AnalyticsDataType.METRICS in the test file.

Important Files Changed

Filename	Overview
nvflare/apis/analytix.py	Adds _normalize_numeric_scalar static method and rewires _validate_data_types to normalize scalar-like values (numpy scalars, 0-D arrays) for SCALAR/METRIC/SCALARS/METRICS types; returns the (possibly normalized) value instead of void.
tests/unit_test/apis/analytix_test.py	Adds three new test cases: generic scalar-like normalization, dict value normalization for SCALARS, and numpy scalar/0-D array normalization; METRICS dict normalization is not covered.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[AnalyticsData.__init__] --> B[_validate_data_types]
    B --> C{data_type?}
    C -->|SCALAR or METRIC| D[_normalize_numeric_scalar]
    D --> E{isinstance float or int?}
    E -->|Yes| F[return True plus value]
    E -->|No| G{has callable item?}
    G -->|No| H[return False plus value]
    G -->|Yes| I{shape attribute exists?}
    I -->|Yes and shape not empty tuple| H
    I -->|No or shape is empty tuple| J[call item]
    J --> K{result is float or int?}
    K -->|Yes| L[return True plus scalar]
    K -->|No| H
    F --> M[value normalized stored in self.value]
    L --> M
    H --> N[raise TypeError for SCALAR or METRIC]
    C -->|METRICS or PARAMETERS or SCALARS| O{isinstance dict?}
    O -->|No| P[raise TypeError]
    O -->|Yes| Q{METRICS or SCALARS?}
    Q -->|Yes| R[loop normalize each dict entry]
    R --> S{entry normalizable?}
    S -->|Yes| T[replace with Python float or int]
    S -->|No| U[keep original value silently]
    T --> V[value = normalized_dict]
    U --> V
    Q -->|No - PARAMETERS| W[value unchanged]
    C -->|TEXT| X{isinstance str?}
    X -->|No| Y[raise TypeError]
    C -->|TAGS| Z{isinstance dict?}
    Z -->|No| AA[raise TypeError]
    M --> BB[return value to init]
    V --> BB
    W --> BB

Loading

_{Reviews (1): Last reviewed commit: "Support scalar-like analytics values" | Re-trigger Greptile}

[YuanTingHsieh]

Superseded by replacement PR #4683, which was recreated from latest main after this PR glitched.