How do I prevent initial validation run in Trainer 1.9.0?

Boris_K · January 24, 2023, 6:43am

In 1.9, the Trainer insists on running a validation step which I don’t know how to disable!
I tried setting num_sanit_val_steps=0 in trainer initialization - but to no avail. In 1.7.7 it works the way I want. Thanks.

justusschock · January 24, 2023, 11:13am

Hi,

I don’t know why this shouldn’t be available. We did not remove this feature:

github.com

Lightning-AI/lightning/blob/master/src/pytorch_lightning/trainer/trainer.py#L151


      
          limit_val_batches: Optional[Union[int, float]] = None,
          limit_test_batches: Optional[Union[int, float]] = None,
          limit_predict_batches: Optional[Union[int, float]] = None,
          val_check_interval: Optional[Union[int, float]] = None,
          log_every_n_steps: int = 50,
          accelerator: Optional[Union[str, Accelerator]] = None,
          strategy: Optional[Union[str, Strategy]] = None,
          sync_batchnorm: bool = False,
          precision: _PRECISION_INPUT = 32,
          enable_model_summary: bool = True,
          num_sanity_val_steps: int = 2,
          profiler: Optional[Union[Profiler, str]] = None,
          benchmark: Optional[bool] = None,
          deterministic: Optional[Union[bool, _LITERAL_WARN]] = None,
          reload_dataloaders_every_n_epochs: int = 0,
          auto_lr_find: Union[bool, str] = False,
          replace_sampler_ddp: bool = True,
          detect_anomaly: bool = False,
          auto_scale_batch_size: Union[str, bool] = False,
          plugins: Optional[Union[PLUGIN_INPUT, List[PLUGIN_INPUT]]] = None,
          multiple_trainloader_mode: str = "max_size_cycle",