AutoDL overview¶
The AutoDL entrypoint is autodl/train.py. It reuses prepare_training_data() and then runs the models listed in config.AUTODL_MODELS:
Shared PyTorch trainer¶
MLP, BLSTM, CNN1D, and Transformer use train_torch_binary_classifier().
The trainer:
- converts the feature matrix to
float32 - creates a stratified train/validation split
- builds
DataLoaderobjects - trains with
BCEWithLogitsLoss - updates parameters with
AdamW - evaluates F1, ROC AUC, and Accuracy after each epoch
- keeps the best validation AUC state
- stops early when AUC stops improving
Main settings:
AUTODL_BATCH_SIZE = 512
AUTODL_EPOCHS = 12
AUTODL_LEARNING_RATE = 1e-3
AUTODL_WEIGHT_DECAY = 1e-4
AUTODL_HIDDEN_DIM = 128
AUTODL_DROPOUT = 0.2
AUTODL_EARLY_STOPPING_PATIENCE = 3
TabNet uses pytorch-tabnet and its own fit() method, while still using the project’s split and metric utilities.