Trait TrainingControl 

pub trait TrainingControl:
    Debug
    + Send
    + Sync {
    // Required methods
    fn reset_gradients(&mut self) -> Result<()>;
    fn reduce_learning_rate(&mut self, factor: f64) -> Result<()>;
    fn clip_gradients(&mut self, max_norm: f64) -> Result<()>;
    fn restart_optimizer(&mut self) -> Result<()>;
    fn skip_batch(&mut self) -> Result<()>;
    fn reset_weights(&mut self, layer_name: &str) -> Result<()>;
    fn apply_weight_decay(&mut self, rate: f64) -> Result<()>;
    fn emergency_stop(&mut self) -> Result<()>;
}

A live handle into the training loop that AnomalyDetector can drive automatic recovery actions through.

Implementations perform the real mutation (clip gradients, reduce the learning rate, restart the optimizer, …) against whatever optimizer or model state the caller owns. Without a TrainingControl attached via AnomalyDetector::set_training_control, AnomalyDetector never claims a recovery action succeeded: every action honestly reports “not performed” instead of the old unconditional Ok(true).

Each method returns Err (propagated into the recorded RecoveryAttempt::error_message, never swallowed) if the underlying mutation could not be carried out.

Required Methods

fn reset_gradients(&mut self) -> Result<()>

Zero out (or otherwise reset) the current gradients.

fn reduce_learning_rate(&mut self, factor: f64) -> Result<()>

Multiply the optimizer’s learning rate by factor.

fn clip_gradients(&mut self, max_norm: f64) -> Result<()>

Clip gradients to max_norm (e.g. global-norm clipping).

fn restart_optimizer(&mut self) -> Result<()>

Reset the optimizer’s internal state (momentum buffers, etc.).

fn skip_batch(&mut self) -> Result<()>

Skip the batch currently being processed.

fn reset_weights(&mut self, layer_name: &str) -> Result<()>

Reset a specific layer’s weights to their initialization.

fn apply_weight_decay(&mut self, rate: f64) -> Result<()>

Apply weight decay at rate to the current weights.

fn emergency_stop(&mut self) -> Result<()>

Immediately halt training.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors