Add default values for args to fix Attribute Error

Signed-off-by: Abhishree <[email protected]>

Author: athitten

nemo/core/classes/modelPT.py

@@ -849,8 +849,11 @@ def test_dataloader(self):
         if self._test_dl is not None:
             return self._test_dl
 
+    #TODO: Confirm if outputs default vals are required
     def on_validation_epoch_end(
-        self, outputs: Union[List[Dict[str, torch.Tensor]], List[List[Dict[str, torch.Tensor]]]]
+        self, outputs: Union[List[Dict[str, torch.Tensor]], List[List[Dict[str, torch.Tensor]]]] = [{'val_loss':torch.tensor(0.), 'val_wer_num':torch.tensor(0.), 'val_wer_denom':torch.tensor(0.),
+                                    'val_final_loss': torch.tensor(0.), 'val_inter_ctc_loss_l2': torch.tensor(0.), 'val_inter_ctc_loss_l4': torch.tensor(0.), 'val_inter_wer_l2': torch.tensor(0.)}]
+        #[[torch.tensor(0)], [torch.tensor(0)]]
     ) -> Optional[Dict[str, Dict[str, torch.Tensor]]]:
         """
         Default DataLoader for Validation set which automatically supports multiple data loaders

nemo/utils/exp_manager.py

@@ -197,10 +197,10 @@ def on_train_batch_start(self, trainer, pl_module, batch, batch_idx):
     def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
         self._on_batch_end("train_step_timing", pl_module)
 
-    def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx):
+    def on_validation_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx=0):
         self._on_batch_start("validation_step_timing")
 
-    def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):
+    def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx=0):
         self._on_batch_end("validation_step_timing", pl_module)
 
     def on_test_batch_start(self, trainer, pl_module, batch, batch_idx, dataloader_idx):
@@ -453,7 +453,6 @@ def exp_manager(trainer: 'pytorch_lightning.Trainer', cfg: Optional[Union[DictCo
     if cfg.disable_validation_on_resume:
         # extend training loop to skip initial validation when resuming from checkpoint
         configure_no_restart_validation_training_loop(trainer)
-
     # Setup a stateless timer for use on clusters.
     if cfg.max_time_per_run is not None:
         found_ptl_timer = False
@@ -937,8 +936,8 @@ def configure_no_restart_validation_training_loop(trainer: pytorch_lightning.Tra
     if type(trainer.fit_loop.epoch_loop) != _TrainingEpochLoop:
         warnings.warn("Detected custom epoch loop. Skipping no validation on restart support.", UserWarning)
         return
-    loop = SkipResumeTrainingValidationLoop(trainer.min_steps, trainer.max_steps)
-    loop.trainer = trainer
+    ## Pass trainer object to avoid trainer getting overwritten as None
+    loop = SkipResumeTrainingValidationLoop(trainer, trainer.min_steps, trainer.max_steps)
     trainer.fit_loop.epoch_loop = loop
 
 

tests/collections/common/test_ema.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import os.path
-from typing import Any, Dict, Union
+from typing import Any, Dict, Union, Optional
 
 import pytest
 import pytorch_lightning as pl
@@ -105,7 +105,7 @@ def setup_validation_data(self, val_data_config: Union[DictConfig, Dict]):
     def setup_test_data(self, val_data_config: Union[DictConfig, Dict]):
         pass
 
-    def on_validation_epoch_end(self, loss):
+    def on_validation_epoch_end(self, loss: torch.tensor = [torch.tensor([0.0])]):
         self.log("val_loss", torch.stack(loss).mean())