def test_full_training_loop_scalar(tmpdir): """ Checks train_step + training_step_end + training_epoch_end (all with scalar return from train_step) """ model = DeterministicModel() model.training_step = model.training_step__scalar_return model.training_step_end = model.training_step_end__scalar model.training_epoch_end = model.training_epoch_end__scalar model.val_dataloader = None trainer = Trainer( default_root_dir=tmpdir, max_epochs=1, weights_summary=None, ) trainer.fit(model) # make sure correct steps were called assert model.training_step_called assert model.training_step_end_called assert model.training_epoch_end_called # assert epoch end metrics were added assert len(trainer.logger_connector.callback_metrics) == 0 assert len(trainer.logger_connector.progress_bar_metrics) == 0 # make sure training outputs what is expected for batch_idx, batch in enumerate(model.train_dataloader()): break out = trainer.train_loop.run_training_batch(batch, batch_idx, 0) assert out.signal == 0 assert len(out.grad_norm_dic) == 0 and isinstance(out.grad_norm_dic, dict) train_step_out = out.training_step_output_for_epoch_end assert len(train_step_out) == 1 train_step_out = train_step_out[0][0] assert isinstance(train_step_out['minimize'], torch.Tensor) assert train_step_out['minimize'].item() == 171 # make sure the optimizer closure returns the correct things opt_closure_result = trainer.train_loop.training_step_and_backward( batch, batch_idx, 0, trainer.optimizers[0], trainer.hiddens ) assert opt_closure_result['loss'].item() == 171
def test_train_step_epoch_end(tmpdir): """ Checks train_step + training_epoch_end (NO training_step_end) """ model = DeterministicModel() model.training_step = model.training_step__dict_return model.training_step_end = None model.training_epoch_end = model.training_epoch_end__dict model.val_dataloader = None trainer = Trainer( default_root_dir=tmpdir, max_epochs=1, weights_summary=None, ) trainer.fit(model) # make sure correct steps were called assert model.training_step_called assert not model.training_step_end_called assert model.training_epoch_end_called # assert epoch end metrics were added assert trainer.logger_connector.callback_metrics['epoch_end_log_1'] == 178 assert trainer.logger_connector.progress_bar_metrics[ 'epoch_end_pbar_1'] == 234 # make sure training outputs what is expected batch_idx, batch = 0, next(iter(model.train_dataloader())) out = trainer.train_loop.run_training_batch(batch, batch_idx, 0) assert out.signal == 0 assert trainer.logger_connector.logged_metrics['log_acc1'] == 12.0 assert trainer.logger_connector.logged_metrics['log_acc2'] == 7.0 # outputs are for 1 optimizer and no tbptt train_step_end_out = out.training_step_output_for_epoch_end assert len(train_step_end_out) == 1 train_step_end_out = train_step_end_out[0][0] pbar_metrics = train_step_end_out['progress_bar'] assert pbar_metrics['pbar_acc1'] == 17.0 assert pbar_metrics['pbar_acc2'] == 19.0
def test_result_obj_lr_scheduler_step(tmpdir): """ test that the LR scheduler was called at the correct time with the correct metrics """ model = DeterministicModel() model.training_step = model.training_step__for_step_end_dict model.training_step_end = model.training_step_end__dict model.training_epoch_end = model.training_epoch_end__dict model.val_dataloader = None model.configure_optimizers = model.configure_optimizers__lr_on_plateau_step trainer = Trainer( default_root_dir=tmpdir, max_epochs=2, weights_summary=None, ) trainer.fit(model) assert len(trainer.dev_debugger.saved_lr_scheduler_updates) == 8