Plots outliers interactions and Model Tracking Upgrade

[guillaume-byte]

Model signals: docs + example

Adds documentation and a worked example for the new per-step, per-layer model
signal tracking feature ("model signals"), covering how to use and interpret
these training-dynamics signals, updated API references, and a gallery entry.

UI Plots Zoom on Outliers

Allow the UI to request outlayers for a specific run and step.

Documentation and examples

• New example page model_signals.rst: tracking and interpreting per-layer
  training dynamics (gradient norms, weight norms, activation stats) on
  Fashion-MNIST, with explanations and code samples.
• Listed the example in the use-cases index and the gallery, with tags and a
  summary for discoverability.

API and usage docs

• Documented wl.save_model_signals and wl.track_model_signals, plus the new
  wl.watch_or_edit arguments for configuring model signal collection.
• Expanded the logger docs to distinguish per-sample, per-annotation, per-group,
  and per-step (model) signals, with guidance on correct usage and plotting
  behavior.

Concepts and best practices

• New sections in the model interaction guide and the conventions guide
  explaining the purpose, integration, and interpretation of model signals,
  including diagnostic patterns and model-architecture best practices.

Navigation

• Standardized section captions in the main index, added a MIGRATION section,
  and clarified where to find integrations and references.

[Copilot]

Pull request overview

This PR extends WeightsLab’s signal/plot pipeline to (1) carry per-step outlier context to the UI (including fetching the full batch behind a plotted point) and (2) add a new “per-step model signals” API (gradient/weight norms and activation stats) with end-to-end examples, docs, and tests.

Changes:

• Add signal outlier payloads (trend band, per-step min/max range, top-N outliers) and downsampling that preserves outlier-bearing points; introduce GetStepSamples RPC to fetch the full batch behind a plotted step.
• Introduce wl.save_model_signals and wl.track_model_signals plus an implementation module that installs hooks and flushes per-step metrics.
• Add extensive docs/examples/tests for both outliers and model signals.

Reviewed changes

Copilot reviewed 20 out of 21 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
weightslab/trainer/services/experiment_service.py	Forwards new outlier/trend/range fields to proto points, preserves outliers when downsampling, adds GetStepSamples.
weightslab/src.py	Adds public per-step model-signal write path + wiring for model-signal tracking via watch_or_edit.
weightslab/proto/experiment_service.proto	Adds SignalOutlier, additional LoggerDataPoint fields, plus GetStepSamples RPC/messages.
weightslab/proto/experiment_service_pb2.py	Regenerated protobuf Python stubs to include new messages/fields.
weightslab/proto/experiment_service_pb2_grpc.py	Regenerated gRPC service/client stubs to include GetStepSamples.
weightslab/examples/Usecases/wl-fashion-mnist-signals/main.py	New runnable example demonstrating model-signal tracking.
weightslab/examples/Usecases/wl-fashion-mnist-signals/config.yaml	Config for the Fashion-MNIST model-signals example.
weightslab/examples/Notebooks/Usecases/ws-segmentation-loss-shapes-classification.ipynb	New/updated notebook use case content (segmentation + decorated signal).
weightslab/examples/Notebooks/Local/wl-local-studio-quickstart.ipynb	Notebook updated, but currently contains unresolved merge conflict markers.
weightslab/components/model_signals.py	New implementation for per-step model-signal collection/flush via hooks.
weightslab/backend/logger.py	Adds outlier detection, trend tracking, schema migrations, range tracking, and step-sample-id queries.
weightslab/init.py	Re-exports save_model_signals / track_model_signals from the public API.
tests/general/test_model_signals.py	Unit tests for save_model_signals and hook-based track_model_signals.
tests/backend/test_signal_outliers.py	Tests for trend/outlier/range plumbing, persistence/migration, and service-level downsampling behavior.
docs/user_functions.rst	Documents new public APIs and model kwargs.
docs/model_interaction.rst	Adds model-signal guidance and best practices.
docs/logger.rst	Explains signal types and plot grouping (including model signals).
docs/examples/usecases/model_signals.rst	New docs page for the model-signals use case.
docs/examples/usecases/index.rst	Adds model-signals use case to the index.
docs/_static/examples-gallery.js	Adds model-signals card to the examples gallery.
AGENTS.md	Adds guidance on choosing the right save_* verb and model-signal tracking.

Files not reviewed (1)

• weightslab/proto/experiment_service_pb2.py: Generated file

Suppressed comments (2)

weightslab/examples/Notebooks/Local/wl-local-studio-quickstart.ipynb:72

• This cell has merge conflict markers around the source field; the resulting notebook JSON is invalid.

<<<<<<< HEAD
   "source": "import os\n\n# Imports weightslab\nimport weightslab as wl\n\n# Serve the weightslab app\nwl.serve()\n\n# Define root log dir for experiment\nroot_log_dir = os.environ.get(\"WEIGHTSLAB_ROOT_LOG_DIR\", None)"
=======
   "source": [
    "import os\n",

weightslab/examples/Notebooks/Local/wl-local-studio-quickstart.ipynb:102

• There are merge conflict markers at the end of the notebook as well, leaving duplicate/marked closing braces and invalid JSON.

<<<<<<< HEAD
}
=======
}
>>>>>>> origin/main

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.