bootstrap kfold

configs/mlp_bootstrap.yaml

@@ -0,0 +1,81 @@
+# Example configuration for MLP bootstrap ensemble training with toy data.
+
+seed: 2025
+device: cpu
+
+datasource:
+  datasource_type: pmgjson
+  json_path: ./data/toy_data/
+  spectrum_key: "XANES"
+
+dataset:
+  # general:
+  dataset_type: descriptor
+  root: ./data/processed/toy_data_mlp_bootstrap/  # Where should the processed data be stored
+  preload: True                    # Preload dataset into RAM if True; otherwise load on-the-fly
+  skip_prepare: False
+  split_ratios: [0.8, 0.2]
+  # params:
+  # descriptors:
+  descriptors:
+    - descriptor_type: wacsf
+      r_min: 1.0
+      r_max: 6.0
+      n_g2: 16
+      n_g4: 32
+
+encodings:
+  - encoding_type: identity
+
+model:
+  # general:
+  model_type: mlp
+  # params:
+  in_size: auto
+  out_size: auto
+  hidden_size: 256
+  dropout: 0.1
+  num_hidden_layers: 3
+  shrink_rate: 0.5
+  activation: prelu
+
+trainer:
+  # general:
+  trainer_type: basic
+  batch_size: 4
+  shuffle: True
+  drop_last: False
+  num_workers: 0
+  # params:
+  epochs: 20
+  learning_rate: 0.001
+  optimizer: Adam
+  max_norm: null
+  validation_interval: 10
+  lr_warmup: True
+  warmup_steps: 500
+  loss:
+    - loss_type: mse
+  regularizer:
+    regularizer_type: none
+  lr_scheduler:
+    lr_scheduler_type: linear
+    start_factor: 1.0
+    end_factor: 0.1
+    total_iters: 40
+  early_stopper:
+    early_stopper_type: basic
+    patience: 25
+    min_delta: 0.001
+    restore_best: True
+
+strategy:
+  # general:
+  strategy_type: bootstrap
+  weight_init: xavier_uniform
+  bias_init: zeros
+  checkpoint_interval: 25
+  # params:
+  n_models: 5
+  sample_fraction: 1.0
+  seeds: [101, 202, 303, 404, 505]

configs/mlp_bootstrap_infer.yaml

@@ -0,0 +1,23 @@
+# Example configuration for MLP bootstrap ensemble inference with toy data.
+
+seed: 2025
+device: cuda
+
+datasource:
+  datasource_type: pmgjson
+  json_path: ./data/toy_data/
+  spectrum_key: "XANES"
+
+dataset:
+  root: ./data/processed/toy_data_mlp_bootstrap/  # Where should the processed data be stored
+  preload: True                    # Preload dataset into RAM if True; otherwise load on-the-fly
+  skip_prepare: False
+
+inferencer:
+  inferencer_type: ensemble
+  batch_size: 4
+  shuffle: False
+  drop_last: False
+  num_workers: 0
+  buffer_size: 1000
+  model_device_policy: sequential

configs/mlp_kfold.yaml

@@ -0,0 +1,80 @@
+# Example configuration for MLP k-fold cross-validation training with toy data.
+
+seed: 2025
+device: cpu
+
+datasource:
+  datasource_type: pmgjson
+  json_path: ./data/toy_data/
+  spectrum_key: "XANES"
+
+dataset:
+  # general:
+  dataset_type: descriptor
+  root: ./data/processed/toy_data_mlp_kfold/  # Where should the processed data be stored
+  preload: True                    # Preload dataset into RAM if True; otherwise load on-the-fly
+  skip_prepare: False
+  # params:
+  # descriptors:
+  descriptors:
+    - descriptor_type: wacsf
+      r_min: 1.0
+      r_max: 6.0
+      n_g2: 16
+      n_g4: 32
+
+encodings:
+  - encoding_type: identity
+
+model:
+  # general:
+  model_type: mlp
+  # params:
+  in_size: auto
+  out_size: auto
+  hidden_size: 256
+  dropout: 0.1
+  num_hidden_layers: 3
+  shrink_rate: 0.5
+  activation: prelu
+
+trainer:
+  # general:
+  trainer_type: basic
+  batch_size: 4
+  shuffle: True
+  drop_last: False
+  num_workers: 0
+  # params:
+  epochs: 20
+  learning_rate: 0.001
+  optimizer: Adam
+  max_norm: null
+  validation_interval: 10
+  lr_warmup: True
+  warmup_steps: 500
+  loss:
+    - loss_type: mse
+  regularizer:
+    regularizer_type: none
+  lr_scheduler:
+    lr_scheduler_type: linear
+    start_factor: 1.0
+    end_factor: 0.1
+    total_iters: 40
+  early_stopper:
+    early_stopper_type: basic
+    patience: 25
+    min_delta: 0.001
+    restore_best: True
+
+strategy:
+  # general:
+  strategy_type: kfold
+  weight_init: xavier_uniform
+  bias_init: zeros
+  checkpoint_interval: 25
+  # params:
+  n_splits: 3
+  n_repeats: 1
+  seed: 2025

xanesnet/schemas/strategies/bootstrap.schema.yaml

@@ -1,7 +1,6 @@
 $schema: https://json-schema.org/draft/2020-12/schema
 $id: bootstrap.schema.yaml
 title: 'XANESNET strategy schema: bootstrap'
-$comment: Registered in the strategy schemas, but the implementation currently raises NotImplementedError.
 $defs:
   strategyBootstrap:
     type: object
@@ -33,5 +32,20 @@ $defs:
       strategy_type:
         const: bootstrap
         description: Strategy identifier.
-    description: Bootstrap ensemble strategy.
+      n_models:
+        type: integer
+        minimum: 1
+        default: 5
+        description: Number of bootstrap ensemble members.
+      sample_fraction:
+        type: number
+        exclusiveMinimum: 0
+        default: 1.0
+        description: Fraction of the training subset resampled for each bootstrap member.
+      seeds:
+        type: array
+        items:
+          type: integer
+        description: Optional per-model random seeds for resampling and weight initialization.
+    description: Sequential bootstrap-ensemble training and aggregate inference strategy.
 description: Strategy schemas backed by ``StrategyRegistry`` classes.

xanesnet/schemas/strategies/kfold.schema.yaml

@@ -0,0 +1,49 @@
+$schema: https://json-schema.org/draft/2020-12/schema
+$id: kfold.schema.yaml
+title: 'XANESNET strategy schema: kfold'
+$defs:
+  strategyKFold:
+    type: object
+    additionalProperties: false
+    required:
+    - strategy_type
+    properties:
+      weight_init:
+        $ref: ../components/component_types.schema.yaml#/$defs/weightInitName
+        default: default
+        description: Weight initialization scheme name.
+      weight_init_params:
+        type: object
+        default: {}
+        additionalProperties: true
+        $comment: Forwarded as kwargs to the selected torch.nn.init weight initializer. Examples include {a,b} for uniform_,
+          {mean,std} for normal_, {gain} for xavier_*, and {a,mode,nonlinearity} for kaiming_*.
+        description: Additional weight-initializer parameters.
+      bias_init:
+        $ref: ../components/component_types.schema.yaml#/$defs/biasInitName
+        default: zeros
+        description: Bias initialization scheme name.
+      checkpoint_interval:
+        type:
+        - integer
+        - 'null'
+        default: null
+        description: Epoch interval between checkpoints, or ``None``.
+      strategy_type:
+        const: kfold
+        description: Strategy identifier.
+      n_splits:
+        type: integer
+        minimum: 2
+        default: 3
+        description: Number of folds per repeat.
+      n_repeats:
+        type: integer
+        minimum: 1
+        default: 1
+        description: Number of repeated k-fold shuffles.
+      seed:
+        type: integer
+        description: Random seed used to shuffle samples before splitting.
+    description: Repeated k-fold cross-validation strategy returning the best fold model.
+description: Strategy schemas backed by ``StrategyRegistry`` classes.

xanesnet/schemas/strategies/strategies.schema.yaml

@@ -7,6 +7,7 @@ $defs:
     - $ref: single.schema.yaml#/$defs/strategySingle
     - $ref: deep_ensemble.schema.yaml#/$defs/strategyDeepEnsemble
     - $ref: bootstrap.schema.yaml#/$defs/strategyBootstrap
+    - $ref: kfold.schema.yaml#/$defs/strategyKFold
     - $ref: snapshot_ensemble.schema.yaml#/$defs/strategySnapshotEnsemble
     description: Union of strategy configuration objects instantiated by ``StrategyRegistry``.
 description: Strategy schemas backed by ``StrategyRegistry`` classes.

xanesnet/schemas/strategies/strategy_types.schema.yaml

@@ -8,6 +8,7 @@ $defs:
     - single
     - deep_ensemble
     - bootstrap
+    - kfold
     - snapshot_ensemble
     description: Strategy registry key accepted by ``StrategyRegistry.get``.
 description: Strategy schemas backed by ``StrategyRegistry`` classes.

xanesnet/serialization/runtime_contracts.py

@@ -110,20 +110,20 @@ def _require_concrete_inference_encodings(config: ConfigRaw) -> None:
         )
 
 
-def _require_ensemble_inferencer_for_deep_ensemble(config: ConfigRaw) -> None:
-    """Validate deep-ensemble inference runner selection.
+def _require_ensemble_inferencer_for_ensemble_strategies(config: ConfigRaw) -> None:
+    """Validate ensemble-strategy inference runner selection.
 
     Args:
         config: Schema-valid merged inference configuration.
 
     Raises:
-        ConfigError: If a deep-ensemble strategy is paired with a non-ensemble
+        ConfigError: If an ensemble strategy is paired with a non-ensemble
             inferencer.
     """
     strategy_type = _section_value(config, "strategy", "strategy_type")
     inferencer_type = _section_value(config, "inferencer", "inferencer_type")
-    if strategy_type == "deep_ensemble" and inferencer_type != "ensemble":
-        raise ConfigError("Inference strategy 'deep_ensemble' requires inferencer 'ensemble'.")
+    if strategy_type in {"deep_ensemble", "bootstrap"} and inferencer_type != "ensemble":
+        raise ConfigError(f"Inference strategy '{strategy_type}' requires inferencer 'ensemble'.")
 
 
 def _section_value(config: ConfigRaw, section: str, key: str) -> Any:
@@ -169,7 +169,7 @@ def _auto_token_paths(value: Any, path: tuple[str, ...]) -> Iterator[str]:
         _require_registered_batch_processor,
         _require_concrete_inference_model,
         _require_concrete_inference_encodings,
-        _require_ensemble_inferencer_for_deep_ensemble,
+        _require_ensemble_inferencer_for_ensemble_strategies,
     ),
     "analyze": (),
 }

xanesnet/strategies/__init__.py

@@ -23,6 +23,7 @@
 from .base import Strategy
 from .bootstrap import Bootstrap
 from .deep_ensemble import DeepEnsemble
+from .kfold import KFold
 from .registry import StrategyRegistry
 from .single import Single
 from .snapshot_ensemble import SnapshotEnsemble
@@ -31,6 +32,7 @@
     "Strategy",
     "Bootstrap",
     "DeepEnsemble",
+    "KFold",
     "SnapshotEnsemble",
     "Single",
     "StrategyRegistry",