"""Model classes - one file per model.

Traditional ML (sklearn-backed, TF-IDF input):
- ``naive_bayes.NaiveBayesModel``
- ``svm_classifier.SVMModel``     (LinearSVC + CalibratedClassifierCV)
- ``random_forest.RandomForestModel``

Hybrid DL (Keras, FastText input):
- ``cnn_bigru.CNNBiGRUModel``
- ``cnn_lstm.CNNLSTMModel``
- ``cnn_svm.CNNSVMModel``         (Keras CNN feature extractor + LinearSVC head)

Transformer fine-tuning (PyTorch + HuggingFace):
- ``indobert.IndoBERTModel``      (`indobenchmark/indobert-base-p1`)
- ``indobertweet.IBTModel``       (`indolem/indobertweet-base-uncased`)
- ``xlm_roberta.XLMRobertaModel`` (`xlm-roberta-base`)
- ``mdeberta.MDeBERTaModel``      (`microsoft/mdeberta-v3-base`)

Hybrid Transformer - encoder + custom head, all use ``bert_heads`` modules:
- ``indobert_hybrid``, ``mdeberta_hybrid``, ``xlm_roberta_hybrid``
  (each exposes a model class for CNN / BiLSTM / CNN-BiLSTM head)
- IBT-specific (separate due to active tuning scope):
  ``indobertweet_cnn.IBTCNNModel``,
  ``indobertweet_bilstm.IBTBiLSTMModel``,
  ``indobertweet_cnn_bilstm.IBTCNNBiLSTMModel``

Heads:
- ``bert_heads.BertCNNHead`` - Conv1d → ReLU → Dropout → MaxPool → Linear
- ``bert_heads.BertBiLSTMHead`` - BiLSTM(1 layer) → Dropout → MaxPool → Linear
- ``bert_heads.BertCNNBiLSTMHead`` - Conv1d → ReLU → BiLSTM → Dropout → MaxPool → Linear

DL/transformer classes are not re-exported here to avoid importing torch/keras
at package-import time. Import them directly from their submodule.
"""

# Traditional-ML classes pull scikit-learn/joblib. Those are optional for the
# slim demo image (which only needs the transformer path), so guard the
# re-export: when the deps are installed this behaves exactly as before; when
# they're absent (deploy env) the package still imports and the transformer
# submodules remain usable via direct import.
try:
    from .naive_bayes import NaiveBayesModel
    from .random_forest import RandomForestModel
    from .svm_classifier import SVMModel
except ModuleNotFoundError:
    pass