Undefined name 'loss' in the creation of 'training_loss_dict'

[cclauss]

Description

Undefined name loss caused by an invalid comprehension in the code:

training_loss_dict = average_sharded_losses([{
              "training": l
} for l in loss for loss in sharded_losses.values()])

https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/utils/t2t_model.py#L341-L343

The above code is an invalid comprehension and will fail with a NameError just like

python -c '[{"training": l} for l in loss for loss in "this is a test".split()]'

Environment information

OS: https://github.com/tensorflow/tensor2tensor

$ python -V
# 3.7.0

For bugs: reproduction and error logs

Steps to reproduce:

$ flake8 . --count --select=E901,E999,F821,F822,F823 --show-source --statistics

Error logs:

flake8 testing of https://github.com/tensorflow/tensor2tensor

$ flake8 . --count --select=E901,E999,F821,F822,F823 --show-source --statistics

./tensor2tensor/utils/t2t_model.py:227:22: F821 undefined name 'loss'
          } for l in loss for loss in sharded_losses.values()])
                     ^
1    F821 undefined name 'loss'
1