experiment

experiment

Configuration for experiment tracking and setup.

ExperimentConfig dataclass

ExperimentConfig(name=MISSING, seed=42, checkpoint_folder=MISSING, resume_from=None, use_wandb=True, wandb_project=None, wandb_entity=None, device='cuda', distributed=False, precision=value, float32_matmul_precision=value, checkpoint_every=100, save_checkpoints=True, val_every=1, validate_loss_keys=True)

Experiment tracking and setup configuration.

Attributes:

Name	Type	Description
name	str	Human-readable name.
seed	int	Random seed for reproducibility.
checkpoint_folder	str	Root directory receiving run checkpoints.
resume_from	str | None	Checkpoint path training resumes from, or null.
use_wandb	bool	Whether metrics are logged to Weights & Biases.
wandb_project	str | None	Weights & Biases project name.
wandb_entity	str | None	Weights & Biases entity, or null for the default.
device	str	Torch device for the module.
distributed	bool	Whether distributed training is enabled.
precision	str	Lightning precision setting for training.
float32_matmul_precision	str | None	torch.set_float32_matmul_precision mode.
checkpoint_every	int	Epoch interval between periodic checkpoints.
save_checkpoints	bool	Whether checkpoints are written at all.
val_every	int	Epoch interval between validation runs.
validate_loss_keys	bool	Whether loss modules are validated against decoder output keys.