# strategies

Python API: `luxonis_train.strategies`

Training strategies that own the optimization schedule.

A strategy adds parameter-group rules to the partition of the model parameters, and can change its own groups on every step. The
package ships
[TripleLRSGDStrategy](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/strategies/triple_lr_sgd.md).
It runs SGD with a warmup and a cosine or linear decay of the learning rate.

 * `Writing a custom strategy:`: Subclass
   [BaseTrainingStrategy](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/strategies/base_strategy.md)
   and implement these methods: * `rules()` returns the rules in order. Each
   [StrategyRule](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/lightning/training_plan.md)
   has a `tag`, a `selector` over `(module, module_name, parameter, parameter_name)`, an `OptimizerConfig`, and an optional
   `SchedulerConfig`. A rule without a scheduler uses the base scheduler.
    * `get_base_configs()` returns the base optimizer config and the base scheduler config. The `finetuning` entries of the nodes
      merge their overrides into them, and the default rule uses them.
    * `update_parameters()` is optional.
      [TrainingManager](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/callbacks/training_manager.md)
      calls it after each backward pass, before the optimizers step. Reach a group through a handle that `attach()` stores for the
      rule tag: `self.runtime.group(handle)["lr"] = ...`.
      [resolve_training_plan](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/lightning/training_plan.md)
      evaluates the strategy rules after the `finetuning` entries of every node and before the default rule. A node entry
      therefore wins, and every parameter that the strategy does not claim still gets an optimizer. Rules with the same optimizer
      name and the same scheduler config share one inner optimizer, with one parameter group for each rule. With more than one
      inner optimizer, one
      [CompositeOptimizer](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/optimizers/composite_optimizer.md)
      drives them, so the training stays in the automatic optimization of Lightning.

A strategy of the previous `configure_optimizers()` API still works until the next minor release.
[luxonis_train.lightning.utils.build_training_strategy](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/lightning/utils.md)
logs a deprecation warning and wraps such a strategy in
[LegacyStrategyAdapter](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/strategies/legacy.md).
To port a strategy, express its parameter split as `rules()`, and move its per-step logic to `update_parameters()`.

## Child Pages

 * [base_strategy](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/strategies/base_strategy.md)
 * [legacy](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/strategies/legacy.md)
 * [triple_lr_sgd](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/strategies/triple_lr_sgd.md)