test: add unit tests for EmbeddingModel, GAN, metrics, and label processors

[sacredvoid]

Summary

Add 6 new test files with 73 test cases covering modules that previously had no unit tests: EmbeddingModel, GAN, binary_metrics_fn, multiclass_metrics_fn, regression_metrics_fn, BinaryLabelProcessor, and MultiClassLabelProcessor.

All tests use synthetic data, complete in under 5 seconds, and follow the established patterns from test_transformer.py / test_retain.py.

test_embedding_model.py (14 tests)

Test	What it verifies
test_initialization	Correct embedding_dim, layers created for each feature
test_embedding_layers_are_correct_type	nn.Embedding for sequences, nn.Linear for tensors/multi_hot
test_forward_output_shapes	Embedded output has correct batch and embedding dimensions
test_forward_with_output_mask	output_mask=True returns masks dict alongside embeddings
test_gradients_flow	Gradients propagate through embedding layers
Tensor/MultiHot/NestedSequence variants	Same checks across all 4 processor types
test_mixed_forward	Mixed sequence + tensor inputs work together
test_custom_embedding_dim	Non-default embedding_dim produces correct output sizes

test_gan.py (14 tests)

Test	What it verifies
test_initialization	Model attributes and submodules created correctly
test_discriminator_output_shape	Discriminator output is (batch, 1) for 32/64/128 inputs
test_discriminator_output_range	Scores in [0, 1] (sigmoid output)
test_generate_fake_shape	Generator produces correct spatial dimensions
test_generate_fake_pixel_range	Pixel values in [0, 1]
test_sampling_shape	Latent noise has shape (n, hidden_dim, 1, 1)
test_discriminator_backward / test_generator_backward	Gradients flow through both networks
test_end_to_end / test_multichannel_end_to_end	Generate then discriminate pipeline works

test_binary_metrics.py (10 tests)

Tests binary_metrics_fn with all 9 supported metrics (pr_auc, roc_auc, accuracy, balanced_accuracy, f1, precision, recall, cohen_kappa, jaccard), plus default behavior, custom thresholds, value range validation, perfect predictions, and error handling. One test is skipped documenting a bug in ece_confidence_binary (calibration.py:150) where 1D arrays are indexed as 2D.

test_multiclass_metrics.py (11 tests)

Tests multiclass_metrics_fn with ROC-AUC variants (macro/weighted, ovo/ovr), F1 variants (micro/macro/weighted), Jaccard variants, calibration metrics (ECE, brier_top1, cwECEt), cohen_kappa, hits@n, and mean_rank.

test_regression_metrics.py (9 tests)

Tests regression_metrics_fn with MSE, MAE, KL divergence, perfect reconstruction (zero error), shape mismatch error handling, and 2D array flattening.

test_label_processor.py (15 tests)

Tests BinaryLabelProcessor (fitting with int/bool/string labels, non-binary raises error, process output, size, schema) and MultiClassLabelProcessor (fitting, correct index mapping, size, schema).

Pattern

Follows the established test patterns:

• Synthetic data via create_sample_dataset() or small numpy arrays
• torch.manual_seed(42) / np.random.seed(42) for determinism
• Standard unittest.TestCase structure
• Google-style docstrings on every method

How to test

python -m pytest tests/core/test_embedding_model.py tests/core/test_gan.py tests/core/test_binary_metrics.py tests/core/test_multiclass_metrics.py tests/core/test_regression_metrics.py tests/core/test_label_processor.py -v

Ref #425