refactor(P3b): move validator factories into the ModalityRegistry

tracebloc_ingestor/modalities/registry.py

@@ -13,79 +13,92 @@
 from typing import Dict
 
 from ..utils.constants import TaskCategory
+from . import validators as v
 from .spec import ModalitySpec
 
-# One entry per supported category. P3a populates the three behavior flags
-# that were hand-maintained frozensets in ingestors/base.py.
+# One entry per supported category — the single source of truth. P3a: the three
+# behavior flags (was three frozensets in ingestors/base.py). P3b: the
+# validator factory (was the map_validators if/elif arm).
 _SPECS = (
     # File-bearing categories (per-row sidecar files under SRC_PATH).
     ModalitySpec(
         TaskCategory.IMAGE_CLASSIFICATION,
         is_file_bearing=True,
         is_tabular_family=False,
         is_self_supervised=False,
+        build_validators=v.image_classification,
     ),
     ModalitySpec(
         TaskCategory.OBJECT_DETECTION,
         is_file_bearing=True,
         is_tabular_family=False,
         is_self_supervised=False,
+        build_validators=v.object_detection,
     ),
     ModalitySpec(
         TaskCategory.KEYPOINT_DETECTION,
         is_file_bearing=True,
         is_tabular_family=False,
         is_self_supervised=False,
+        build_validators=v.keypoint_detection,
     ),
     ModalitySpec(
         TaskCategory.SEMANTIC_SEGMENTATION,
         is_file_bearing=True,
         is_tabular_family=False,
         is_self_supervised=False,
+        build_validators=v.semantic_segmentation,
     ),
     ModalitySpec(
         TaskCategory.TEXT_CLASSIFICATION,
         is_file_bearing=True,
         is_tabular_family=False,
         is_self_supervised=False,
+        build_validators=v.text_classification,
     ),
     ModalitySpec(
         TaskCategory.TOKEN_CLASSIFICATION,
         is_file_bearing=True,
         is_tabular_family=False,
         is_self_supervised=False,
+        build_validators=v.token_classification,
     ),
     # masked_language_modeling is file-bearing AND self-supervised (no label).
     ModalitySpec(
         TaskCategory.MASKED_LANGUAGE_MODELING,
         is_file_bearing=True,
         is_tabular_family=False,
         is_self_supervised=True,
+        build_validators=v.masked_language_modeling,
     ),
     # Tabular family (structured feature tables; no sidecar files).
     ModalitySpec(
         TaskCategory.TABULAR_CLASSIFICATION,
         is_file_bearing=False,
         is_tabular_family=True,
         is_self_supervised=False,
+        build_validators=v.tabular_classification,
     ),
     ModalitySpec(
         TaskCategory.TABULAR_REGRESSION,
         is_file_bearing=False,
         is_tabular_family=True,
         is_self_supervised=False,
+        build_validators=v.tabular_regression,
     ),
     ModalitySpec(
         TaskCategory.TIME_SERIES_FORECASTING,
         is_file_bearing=False,
         is_tabular_family=True,
         is_self_supervised=False,
+        build_validators=v.time_series_forecasting,
     ),
     ModalitySpec(
         TaskCategory.TIME_TO_EVENT_PREDICTION,
         is_file_bearing=False,
         is_tabular_family=True,
         is_self_supervised=False,
+        build_validators=v.time_to_event_prediction,
     ),
 )
 

tracebloc_ingestor/modalities/spec.py

@@ -25,14 +25,19 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
+from typing import Any, Callable, Dict, List
 
 
 @dataclass(frozen=True)
 class ModalitySpec:
-    """Per-category behavior flags (P3a). More fields land in P3b–P3d.
+    """Per-category behavior (built up over P3a–P3d). More fields land in P3c/P3d.
 
     Attributes:
         category: the ``TaskCategory`` value this spec describes.
+        build_validators: ``(file_options) -> [validators]`` — the validator
+            set this category runs (P3b). Replaces the corresponding
+            ``map_validators`` if/elif arm; the factory bodies live in
+            ``modalities/validators.py``.
         is_file_bearing: every record references sidecar files (images,
             annotations, masks, texts, sequences) under ``SRC_PATH`` that must
             be copied to ``DEST_PATH``. Drives both the SRC_PATH preflight and
@@ -50,3 +55,4 @@ class ModalitySpec:
     is_file_bearing: bool
     is_tabular_family: bool
     is_self_supervised: bool
+    build_validators: Callable[[Dict[str, Any]], List]

tracebloc_ingestor/modalities/validators.py

@@ -0,0 +1,254 @@
+"""Per-category validator factories (structural refactor — backend#796, P3b).
+
+One factory per task category, each ``(options) -> [validators]``. These are
+the bodies of the old ``utils.validators_mapping.map_validators`` if/elif arms,
+moved verbatim so the validator sets are byte-for-byte identical — now attached
+to each ModalitySpec (``build_validators``) instead of dispatched by a ladder.
+``map_validators`` is now a thin lookup over the registry.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Dict, List
+
+from ..utils.constants import FileExtension
+from ..validators.base import BaseValidator
+from ..validators.bio_label_validator import BIOLabelValidator
+from ..validators.data_validator import DataValidator
+from ..validators.duplicate_validator import DuplicateValidator
+from ..validators.file_pairing_validator import FilePairingValidator
+from ..validators.file_validator import FileTypeValidator
+from ..validators.image_validator import ImageResolutionValidator
+from ..validators.keypoint_annotation_validator import KeypointAnnotationValidator
+from ..validators.keypoint_visibility_validator import KeypointVisibilityValidator
+from ..validators.label_diversity_validator import LabelDiversityValidator
+from ..validators.numeric_columns_validator import NumericColumnsValidator
+from ..validators.table_name_validator import TableNameValidator
+from ..validators.time_before_today_validator import TimeBeforeTodayValidator
+from ..validators.time_format_validator import TimeFormatValidator
+from ..validators.time_ordered_validator import TimeOrderedValidator
+from ..validators.time_to_event_validator import TimeToEventValidator
+from ..validators.tokenizer_validator import TokenizerValidator
+from ..validators.xml_validator import PascalVOCXMLValidator
+
+
+def _label_diversity_validator(options: Dict[str, Any]) -> LabelDiversityValidator:
+    """Construct a LabelDiversityValidator using the user-configured label
+    column name (or the framework default ``label``). Centralised so every
+    classification-family factory wires the same instance shape.
+
+    Issue #251: a classification dataset with one distinct label value is
+    unlearnable and the backend rejects it at ``/global_meta/prepare/`` with
+    ``HTTP 400: "Please provide atleast 2 labels."``. Catching it at preflight
+    surfaces the actual cause (and lists the offending label value(s)) instead
+    of cascading to a misleading "Backend failed to prepare the dataset"
+    message after the rows have already landed in MySQL.
+    """
+    return LabelDiversityValidator(
+        label_column=options.get("label_column") or "label",
+        # Read the label column with the SAME NA / dtype rules CSVIngestor
+        # uses, or the distinct-label count disagrees with what's actually
+        # ingested (bugbot #252). Prefer ``full_schema`` (base.py passes the
+        # UNSTRIPPED schema here): ``schema``/``file_options["schema"]`` has the
+        # label column removed, so it can't carry the label's type — the very
+        # column this validator reads. Fall back to ``schema`` for direct
+        # callers / tests that pass an unstripped map.
+        schema=options.get("full_schema") or options.get("schema"),
+    )
+
+
+def image_classification(options: Dict[str, Any]) -> List[BaseValidator]:
+    return [
+        FileTypeValidator(allowed_extension=options["extension"], path="images"),
+        ImageResolutionValidator(expected_resolution=options["target_size"]),
+        _label_diversity_validator(options),
+        TableNameValidator(),
+        DuplicateValidator(),
+    ]
+
+
+def object_detection(options: Dict[str, Any]) -> List[BaseValidator]:
+    return [
+        FileTypeValidator(allowed_extension=options["extension"], path="images"),
+        FileTypeValidator(allowed_extension=".xml", path="annotations"),
+        PascalVOCXMLValidator(),
+        FilePairingValidator(
+            image_path="images",
+            sidecar_path="annotations",
+            sidecar_label="annotation",
+        ),
+        ImageResolutionValidator(expected_resolution=options["target_size"]),
+        _label_diversity_validator(options),
+        TableNameValidator(),
+        DuplicateValidator(),
+    ]
+
+
+def tabular_classification(options: Dict[str, Any]) -> List[BaseValidator]:
+    validators: List[BaseValidator] = []
+    # Add data validator if schema is provided
+    if options.get("schema"):
+        validators.append(DataValidator(schema=options["schema"]))
+    validators.append(_label_diversity_validator(options))
+    validators.append(TableNameValidator())
+    validators.append(DuplicateValidator())
+    return validators
+
+
+def text_classification(options: Dict[str, Any]) -> List[BaseValidator]:
+    validators: List[BaseValidator] = []
+    # Add text file validator
+    validators.append(
+        FileTypeValidator(
+            allowed_extension=options.get("extension", FileExtension.TXT),
+            path="texts",
+        ),
+    )
+    # Optional user-supplied tokenizer.json — warn (don't fail) if absent;
+    # if present, it must contain [PAD] (text classification pads batches).
+    validators.append(TokenizerValidator(required_tokens=("[PAD]",), optional=True))
+    # Add data validator if schema is provided
+    if options.get("schema"):
+        validators.append(DataValidator(schema=options["schema"]))
+    validators.append(_label_diversity_validator(options))
+    validators.append(TableNameValidator())
+    validators.append(DuplicateValidator())
+    return validators
+
+
+def token_classification(options: Dict[str, Any]) -> List[BaseValidator]:
+    validators: List[BaseValidator] = []
+    # Validate text file extensions (one .txt of whitespace-tokenized words
+    # per sample, same layout as text classification).
+    validators.append(
+        FileTypeValidator(
+            allowed_extension=options.get("extension", FileExtension.TXT),
+            path="texts",
+        ),
+    )
+    # Validate BIO labels: one tag per word, valid BIO/IOB2 format. Honor a
+    # custom label column name when one is configured in the YAML.
+    validators.append(
+        BIOLabelValidator(
+            texts_path="texts",
+            extension=options.get("extension", FileExtension.TXT),
+            label_column=options.get("label_column") or "label",
+        )
+    )
+    # Optional user-supplied tokenizer.json — warn (don't fail) if absent;
+    # if present, it must contain [PAD].
+    validators.append(TokenizerValidator(required_tokens=("[PAD]",), optional=True))
+    # Add data validator if schema is provided
+    if options.get("schema"):
+        validators.append(DataValidator(schema=options["schema"]))
+    validators.append(TableNameValidator())
+    validators.append(DuplicateValidator())
+    return validators
+
+
+def time_series_forecasting(options: Dict[str, Any]) -> List[BaseValidator]:
+    validators: List[BaseValidator] = []
+    schema = options.get("schema", {})
+    validators.append(TimeFormatValidator(schema=schema))
+    validators.append(TimeOrderedValidator())
+    validators.append(TimeBeforeTodayValidator())
+    validators.append(NumericColumnsValidator(schema=schema))
+    if options.get("schema"):
+        schema_without_timestamp = {
+            k: v for k, v in options["schema"].items() if k.lower() != "timestamp"
+        }
+        if schema_without_timestamp:
+            validators.append(DataValidator(schema=schema_without_timestamp))
+    validators.append(TableNameValidator())
+    validators.append(DuplicateValidator())
+    return validators
+
+
+def tabular_regression(options: Dict[str, Any]) -> List[BaseValidator]:
+    validators: List[BaseValidator] = []
+    # Add data validator if schema is provided
+    if options.get("schema"):
+        validators.append(DataValidator(schema=options["schema"]))
+    validators.append(TableNameValidator())
+    validators.append(DuplicateValidator())
+    return validators
+
+
+def time_to_event_prediction(options: Dict[str, Any]) -> List[BaseValidator]:
+    validators: List[BaseValidator] = []
+    # Add time to event validator with schema to identify time column
+    if options.get("schema"):
+        validators.append(
+            TimeToEventValidator(
+                schema=options["schema"],
+                time_column=options.get("time_column"),
+            )
+        )
+    else:
+        # If no schema, use default time column name
+        validators.append(
+            TimeToEventValidator(time_column=options.get("time_column", "time"))
+        )
+    # Add data validator if schema is provided
+    if options.get("schema"):
+        validators.append(DataValidator(schema=options["schema"]))
+    validators.append(TableNameValidator())
+    validators.append(DuplicateValidator())
+    return validators
+
+
+def semantic_segmentation(options: Dict[str, Any]) -> List[BaseValidator]:
+    return [
+        FileTypeValidator(allowed_extension=options["extension"], path="images"),
+        FileTypeValidator(allowed_extension=FileExtension.PNG, path="masks"),
+        FilePairingValidator(
+            image_path="images",
+            sidecar_path="masks",
+            sidecar_label="mask",
+            # Documented + shipped convention for semantic_segmentation masks
+            # is `<filename>_mask.png` (#196). Strip the suffix before matching
+            # so image_001.jpg pairs with image_001_mask.png. object_detection's
+            # pairing is plain stem (no suffix) — the default.
+            sidecar_suffix="_mask",
+        ),
+        ImageResolutionValidator(expected_resolution=options["target_size"]),
+        _label_diversity_validator(options),
+        TableNameValidator(),
+        DuplicateValidator(),
+    ]
+
+
+def keypoint_detection(options: Dict[str, Any]) -> List[BaseValidator]:
+    # ``number_of_keypoints`` is required by the ingest schema for
+    # keypoint_detection (see ``schema/ingest.v1.json``) and plumbed into
+    # ``file_options`` by ``cli/conventions.py``. Passing it to
+    # ``KeypointAnnotationValidator`` enables the per-row count check that
+    # rejects datasets whose annotations drift from the declared K.
+    return [
+        FileTypeValidator(allowed_extension=options["extension"], path="images"),
+        ImageResolutionValidator(expected_resolution=options["target_size"]),
+        KeypointAnnotationValidator(num_keypoints=options.get("number_of_keypoints")),
+        KeypointVisibilityValidator(),
+        _label_diversity_validator(options),
+        TableNameValidator(),
+        DuplicateValidator(),
+    ]
+
+
+def masked_language_modeling(options: Dict[str, Any]) -> List[BaseValidator]:
+    validators: List[BaseValidator] = []
+    # Validate text file extensions
+    validators.append(
+        FileTypeValidator(
+            allowed_extension=options.get("extension", FileExtension.TXT),
+            path="sequences",
+        ),
+    )
+    # Validate tokenizer.json has required special tokens ([MASK], [PAD])
+    validators.append(TokenizerValidator())
+    # Add data validator if schema is provided
+    if options.get("schema"):
+        validators.append(DataValidator(schema=options["schema"]))
+    validators.append(TableNameValidator())
+    validators.append(DuplicateValidator())
+    return validators