OptimConfig

Contents

OptimConfig#

class stable_ssl.config.OptimConfig(optimizer: str = 'LARS', lr: float = 1.0, batch_size: int = 256, epochs: int = 1000, max_steps: int = -1, weight_decay: float = 0, momentum: float | None = None, nesterov: bool | None = None, betas: Tuple[float, float] | None = None, grad_max_norm: float | None = None)[source]#

Bases: object

Configuration for the ‘optimizer’ parameters.

Parameters:

  • optimizer (str) – Type of optimizer to use (e.g., “AdamW”, “RMSprop”, “SGD”, “LARS”). Default is “LARS”.

  • lr (float) – Learning rate for the optimizer. Default is 1e0.

  • batch_size (int, optional) – Batch size for training. Default is 256.

  • epochs (int, optional) – Number of epochs to train the model. Default is 10.

  • max_steps (int, optional) – Maximum number of steps per epoch. Default is -1.

  • weight_decay (float) – Weight decay for the optimizer. Default is 1e-6.

  • momentum (float) – Momentum for the optimizer. Default is None.

  • nesterov (bool) – Whether to use Nesterov momentum. Default is False.

  • betas (Tuple[float, float], optional) – Betas for the AdamW optimizer. Default is (0.9, 0.999).

  • grad_max_norm (float, optional) – Maximum norm for gradient clipping. Default is None.

Contents