[Logiquo]

Stacked on #1361.

Summary

BaseModel decided how to train on a label from a mode string: it found the mode by walking each processor class's MRO against registry names, then switched on it in get_loss_function and prepare_y_prob. Label semantics lived in the model rather than in the processor that defines the label.

This PR moves those semantics onto the label processors:

• LabelProcessor mixin (pyhealth.processors.LabelProcessor): mode, is_timeseries, supports_pos_weight, activation(logits) and loss(logits, y_true, pos_weight=None).

  Processor	mode	activation	loss	pos_weight
  BinaryLabelProcessor	binary	sigmoid	BCE with logits	✓
  MultiClassLabelProcessor	multiclass	softmax	cross-entropy
  MultiLabelProcessor	multilabel	sigmoid	BCE with logits	✓
  RegressionLabelProcessor	regression	identity	MSE
  TimeseriesBinaryLabelProcessor	timeseries_binary	as binary	binary, per label record
  TimeseriesRegressionLabelProcessor	timeseries_regression	as regression	regression, per label record

  The time-series labels delegate to their static counterpart (base_cls) and keep the per-record shape check and the zero loss for a batch without label records (formerly _label_record_loss). Their fit is FeatureProcessor's no-op: SequenceProcessorV2.fit read every sample and found no tokens or statistics to learn.

• BaseModel delegates. get_loss_function() returns the label processor's loss with the model's pos_weight, and prepare_y_prob() its activation. The results are unchanged. _resolve_mode reads LabelProcessor.mode instead of walking the registry. set_pos_weight checks supports_pos_weight. Without a fitted label processor (e.g. a dataset whose processors are not fitted), the processor registered under the mode is used.

• Optional labels on MultiClassLabelProcessor and MultiLabelProcessor: the label set in index order, or its size K for labels already 0..K-1. When given, the vocabulary (and size()) is fixed, and fit only checks that every label is in it. A class missing from the training split still gets its output. When omitted, the label set is learned as before, which code-prediction tasks such as drug recommendation and medical coding rely on. Pass it as ("multiclass", {"labels": [...]}).

• autoimprove resolves the label mode from LabelProcessor.mode too.

• Own module. The time-series label processors move from timeseries_processor.py (the legacy TimeseriesProcessor input processor) to timeseries_label_processor.py. They are unreleased, so no import path or pickled cache depends on the old location; from pyhealth.processors import ... is unchanged.

Breaking changes

• get_loss_function() returns a functools.partial of the processor's loss, so it has no __name__. The loss values are unchanged.

Tests

• New tests/core/test_label_processor.py covers:
  • activation and loss of all six processors against the torch functions they replace, including pos_weight, its rejection where unsupported, the per-record shape check, the zero loss on an empty batch, and the time-series fit not reading the samples;
  • only label processors are labels (MultiHotProcessor is not);
  • labels: fixed indices and size, K meaning range(K), unknown and invalid labels, omitted labels learned as before, pickling (including an instance pickled before labels existed), and labels passed through a task schema;
  • BaseModel delegation.
• test_legacy_mode_resolution compares loss values rather than function names.
• Ran locally (unittest): the test modules for models, processors, interpretability, calibration, trainer and autoimprove, plus doctests of the changed modules. All pass, except two that fail for environment reasons: test_survival_mdn imports pytest, which is not installed, and test_transformers_model.TestChunking.test_windows_need_a_fast_tokenizer also fails on master here.

[github-actions]

Test Coverage

Total: 75.27% (21568 / 28656 lines)

Modified file	Coverage	Lines
pyhealth/autoimprove/improver.py	95.62%	371 / 388
pyhealth/models/base_model.py	88.99%	97 / 109
pyhealth/processors/__init__.py	94.44%	34 / 36
pyhealth/processors/label_processor.py	92.11%	175 / 190
pyhealth/processors/timeseries_label_processor.py	100.00%	53 / 53
pyhealth/processors/timeseries_processor.py	84.42%	65 / 77

Workflow run · 0c28e2f