fixes for batched strategies

[konstpieper]

• heteroscedastic y_err data with batched strategies
• transforms like normalize and log with Kriging believer
• default local host fix (could go to separate MR)
• remove gpax decorators from tests that do not work with gpax or do not use a model (indexed startegies)

fixes #55

[vreshniak]

self.dataset_y property returns self._dataset_y. What is the point to explicitly use 'hidden' variable? Can self.dataset_y change in future?

[konstpieper]

the setter for dataset_y calls this getter. I was afraid of a circular dependency, which may not be a problem. Is there a problem using the private variable here?

[Lance-Drane]

It generally shouldn't matter whether or not we use the private variable in this instance, since we don't have any side effects in the getter functions. (I am strongly opposed to side effects in getter functions but we don't have any.)

[vreshniak]

Not with current type options, but pydantic can support callables, e.g.

from collections.abc import Callable
from pydantic import BaseModel

class WorkerModel(BaseModel):
    # Pydantic validates that 'callback' is a function/callable
    callback: Callable[[int, str], bool]

# This works fine
def my_function(x: int, s: str) -> bool:
    return True

model = WorkerModel(callback=my_function)

I prefer having an option for user defined callbacks

[konstpieper]

@Lance-Drane this is for you...

[Lance-Drane]

Unfortunately, Callables will not work as a Dataclass type over INTERSECT. Under the hood INTERSECT needs to use the JSON conversion table (not the Python conversion table), and Callables are only valid under the Python conversion table

https://pydantic.dev/docs/validation/latest/concepts/conversion_table/

so we don't want Callable types in any of the dial_dataclass Pydantic BaseModels

[Copilot]

🟡 Changes recommended

Critical batching defects and moderate data/backend issues remain unresolved.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR updates batched strategies for transformed and heteroscedastic outputs, adds backend model-update support, improves validation, and restricts Docker services to localhost.

Changes:

• Reworks liar/believer pseudo-observation handling.
• Updates dataset transformations and dimension handling.
• Expands tests and adjusts Docker port bindings.

File summaries

File	Review comments
tests/unit/test_internals.py	Add regression coverage for transformed, standardized, and heteroscedastic SABLE batching beyond shape assertions. Nit, 1 vote (line 1556).
src/dial_service/utilities/strategies.py	Initialize planning_model for non-indexed liar batching (critical, 3 votes, line 394); normalize scalar/list liar values into complete heteroscedastic rows (critical, 3 votes, line 321); add transformed and heteroscedastic liar/believer regression tests (nit, 1 vote, line 398); fix configureable typo (nit, 1 vote, line 359).
src/dial_service/serverside_data.py	Recompute or invalidate cached training data after bounds overrides (moderate, 3 votes, line 112); handle scalar negative Normal.scale values without indexing errors (moderate, 1 vote, line 98).
src/dial_service/backends/sklearn_backend.py	No review comments.
src/dial_service/backends/sable_backend.py	No review comments.
src/dial_service/backends/gpax_backend.py	No review comments.
src/dial_service/backends/__init__.py	Clarify or implement the abstract update_model fallback (moderate, 2 votes, line 38); fix learing typo (nit, 3 votes, line 42).
src/dial_dataclass/dial_dataclass.py	No review comments.
docker-compose.yml	No review comments.
docker-compose-dev.yml	No review comments.

Review details

Suppressed comments (4)

src/dial_service/serverside_data.py:100

• For a constant negative Normal.scale, _yerr_train_raw is a 0-D array. This branch indexes it with idxs[0], which raises an indexing error instead of the intended validation ValueError; handle scalar errors separately or normalize the value to a 1-D array before indexing.

        if np.any(_yerr_train_raw < 0):
            idxs = np.where(_yerr_train_raw < 0)
            msg = f'yerr values in statistics_y.scale must be non-negative, found {_yerr_train_raw[idxs[0]]} at {idxs[0]}.'

src/dial_service/utilities/strategies.py:359

• Typo: configureable should be configurable.

            # could make this configureable, default to median right now

src/dial_service/utilities/strategies.py:405

• This change adds a regression-prone pseudo-observation path for transformed and heteroscedastic targets, but the batch tests still use scalar outputs; the heteroscedastic test only covers prediction. Add liar and believer cases with Normal(loc='y', scale='yerr') and log/standardize preprocessing, including assertions that the original raw dataset is restored.

            next_y = predictor(next_x).reshape(1, -1)

            data.dataset_x = np.concatenate([data.dataset_x, next_x])
            data.dataset_y = np.concatenate([data.dataset_y, next_y])

            if data.strategy not in INDEXED_STRATEGIES:
                # update_model updates the planning model, not modifying the original
                planning_model = backend_module.update_model(planning_model, data)

tests/unit/test_internals.py:1559

• The SABLE branch only checks the returned shape, so it would pass even if the new inverse-transform logic or heteroscedastic pseudo-observations were incorrect. Since this PR targets preprocess_log, standardization, and Normal(..., scale='yerr') with batching, add a regression case that exercises those settings and asserts the raw-data behavior rather than shape alone.

    else:
        # if we do not use sklearn based GP, there is no truth solution.
        # Check only the shape of the return value
        assert output.shape == (5, dim)

• Files reviewed: 10/10 changed files
• Comments generated: 5
• Review effort level: Lite

---

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

[vreshniak]

Looks good for me

[Lance-Drane]

LGTM, going to merge now