Training

20 public types in this namespace, organized by kind.

Models & Types (7)

Type Summary
DpSgdFusedStep<T> Fused per-example DP-SGD (Abadi et al.
MemorySavingsEstimate Memory savings estimate from optimization techniques.
Trainer<T> Default trainer that delegates to the model’s built-in Train() method each epoch.
TrainingMemoryManager<T> Manages memory optimization during neural network training including gradient checkpointing, activation pooling, and model sharding.
TrainingObjective<T> One optimizer step a model declares instead of writing its own training loop: the step’s data and the loss built from it.
TrainingResult<T> Contains the results of a training run, including the trained model, loss history, and metadata.
WganGpFusedStep<T> Fused WGAN-GP critic training step (Gulrajani et al.

Base Classes (1)

Type Summary
TrainerBase<T> Abstract base class for all trainers, providing shared infrastructure for configuration-driven training pipelines.

Options & Configuration (7)

Type Summary
DatasetConfig Configuration for the dataset section of a training recipe.
LossFunctionConfig Configuration for the loss function section of a training recipe.
ModelConfig Configuration for the model section of a training recipe.
OptimizerConfig Configuration for the optimizer section of a training recipe.
TrainerSettings Configuration for the trainer behavior section of a training recipe.
TrainingMemoryConfig Configuration for training memory management including gradient checkpointing, activation pooling, and model sharding.
TrainingRecipeConfig Root configuration object for a complete training recipe defined in YAML.

Helpers & Utilities (5)

Type Summary
CompiledStepHostValue<T> A tensor a training forward computes on the host from other tensors of the same step (a data-dependent choice: the strongest autocorrelation lags, the most active queries), declared so that a compiled training plan recomputes it on every re…
CompiledStepRandom<T> Random tensors a training step draws fresh on every step (reparameterization noise, a GAN’s latent batch, a diffusion step’s noise), declared so that a compiled training plan redraws them on every replay.
CompiledTapeTrainingStep<T> Compiled training step — auto-compiles the forward + backward pass on the first step, then replays the compiled plan on subsequent steps for near-zero overhead training.
LossFunctionFactory<T> Factory for creating loss function instances from LossType enum values.
TapeTrainingStep<T> Provides PyTorch-style training step using tape-based automatic differentiation, with two-level caching for parameter collection that outperforms PyTorch’s model.parameters() which rebuilds the full list on every call.