Auto_lr_find and auto scale batch size error

[minwang-ai]

Hi,

I used auto_lr_find and auto_scale_batch_size before. The batch size is not correct. (For one 32GB GPU, it returned the batch size over 10000…)

Now, it does not work due to the following error:

Traceback (most recent call last):
I google the key words and update lightning version according to https://github.com/PyTorchLightning/pytorch-lightning/discussions/7092, but the issue is still open. @https://github.com/THasthika
  File "/home/mwa7rng/mwa7rng/run_repos1/FFB6D/ffb6d/main.py", line 119, in main
    lr_finder = trainer.tuner.lr_find(model)
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/tuner/tuning.py", line 197, in lr_find
    "update_attr": update_attr,
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/trainer/trainer.py", line 1092, in tune
    result = self.tuner._tune(model, scale_batch_size_kwargs=scale_batch_size_kwargs, lr_find_kwargs=lr_find_kwargs)
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/tuner/tuning.py", line 53, in _tune
    result["lr_find"] = lr_find(self.trainer, model, **lr_find_kwargs)
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/tuner/lr_finder.py", line 238, in lr_find
    trainer.tuner._run(model)
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/tuner/tuning.py", line 63, in _run
    self.trainer._run(*args, **kwargs)
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/trainer/trainer.py", line 1115, in _run
    verify_loop_configurations(self, model)
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/trainer/configuration_validator.py", line 32, in verify_loop_configurations
    __verify_train_val_loop_configuration(trainer, model)
  File "/home/mwa7rng/.local/lib/python3.6/site-packages/pytorch_lightning/trainer/configuration_validator.py", line 71, in __verify_train_val_loop_configuration
    "No `train_dataloader()` method defined. Lightning `Trainer` expects as minimum a"
pytorch_lightning.utilities.exceptions.MisconfigurationException: No `train_dataloader()` method defined. Lightning `Trainer` expects as minimum a `training_step()`, `train_dataloader()` and `configure_optimizers()` to be defined.

Process finished with exit code 1

My setting is as follows:

dm = YcbDataModule()
    dm.setup()
    #Todo if args.dataset == "linemod":


    # init model
    model = FFB6D(
                    n_classes=config.n_objects, n_pts=config.n_sample_points, rndla_cfg=rndla_cfg,
                    n_kps=config.n_keypoints, task=args.task, gpus=args.gpus, num_train=len(dm.dataset_train))

    # init trainer
    trainer = pl.Trainer(
                         auto_lr_find=True,
                         #auto_scale_batch_size="binsearch",
                         #fast_dev_run=True,                              # flat for debug
                         callbacks=[EarlyStopping(monitor="val_loss/dataloader_idx_1"), checkpoint_callback, lr_monitor, TQDMProgressBar(refresh_rate=20)],
                         logger=tb_logger, profiler="simple",
                         #deterministic=True,  # ensure full reproducibility
                         max_epochs=25,
                         gpus=-1,
                        #strategy=DDPPlugin(find_unused_parameters=False),
                            sync_batchnorm=True)
                        # use all available GPUs

    if args.trainer == "fit":  # train your model

        # tune to find the learning rate
        # Run learning rate finder
        trainer.tune(model)

        # fix model
        trainer.fit(model, datamodule=dm)

[tchaton]

Dear @MinWang1997.
Without YcbDataModule and FFB6D, we can't help you out !

You are missing one of the following methods as the error is sharing:

No train_dataloader() method defined. Lightning Trainer expects as minimum a training_step(), train_dataloader() and configure_optimizers() to be defined.