Using print_nan_grads in the Trainer results in an error

[iamsimha]

Describe the bug
When using

print_nan_grads=True

in the Trainer, I am getting the error below.

trainer.fit(lstm_model)
File "/Users/anaconda3/envs/snorkel/lib/python3.6/site-packages/pytorch_lightning/trainer/trainer.py", line 364, in fit
self.run_pretrain_routine(model)
File "/Users/anaconda3/envs/snorkel/lib/python3.6/site-packages/pytorch_lightning/trainer/trainer.py", line 471, in run_pretrain_routine
self.train()
File "/Users/anaconda3/envs/snorkel/lib/python3.6/site-packages/pytorch_lightning/trainer/train_loop_mixin.py", line 60, in train
self.run_training_epoch()
File "/Users/anaconda3/envs/snorkel/lib/python3.6/site-packages/pytorch_lightning/trainer/train_loop_mixin.py", line 99, in run_training_epoch
output = self.run_training_batch(batch, batch_nb)
File "/Users/anaconda3/envs/snorkel/lib/python3.6/site-packages/pytorch_lightning/trainer/train_loop_mixin.py", line 219, in run_training_batch
self.print_nan_gradients()
File "/Users/anaconda3/envs/snorkel/lib/python3.6/site-packages/pytorch_lightning/trainer/training_tricks_mixin.py", line 16, in print_nan_gradients
if torch.isnan(param.grad.float()).any():
AttributeError: 'NoneType' object has no attribute 'float'

To Reproduce
Steps to reproduce the behavior:
If some param object, does not have .grad, then that object should not be checked for nans

[awaelchli]

sounds like an easy fix: iterate only over params where grad is not None. Would that solve the issue?

[williamFalcon]

exactly. Anyone want to submit the PR?