Better message when DataLoader is wrong #1131
Description
On the verge between bug and improvement.
There was a bug in my Validation DataLoader and was returning irrelevant staff. Accidentally the length was 0. Probably an edge case combination. The error I was getting during the validation sanity check was quite cryptic:
Traceback (most recent call last):
File "UNet_WaveProp.py", line 174, in <module>
trainer.fit(model)
File "/mnt/RDS/home/code/pytorch-lightning/pytorch_lightning/trainer/trainer.py", line 629, in fit
self.run_pretrain_routine(model)
File "/mnt/RDS/home/code/pytorch-lightning/pytorch_lightning/trainer/trainer.py", line 809, in run_pretrain_routine
False)
File "/mnt/RDS/home/code/pytorch-lightning/pytorch_lightning/trainer/evaluation_loop.py", line 300, in evaluate
eval_results = model.validation_epoch_end(outputs)
File "UNet_WaveProp.py", line 138, in validation_epoch_end
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
RuntimeError: stack expects a non-empty TensorList
I had to go through the code of pytorch-lightning for few hours to understand what was happening.
Maybe a more informative message would make more sense?
One thing would be to check if the DataLoader's size is 0.
What do you think? I could take a stab at a PR.