Consistency in ModelCheckpoint verbose mode logs

[khazit]

What does this PR do?

When using ModelCheckpoint in verbose mode, there are two different logs that you can get :

• If your model did not improve, you get Epoch 1, step 100: val_loss was not in top 1
• If your model did improve, you get Epoch 1, global step 100: val_loss was not in top 1

In the first case, the global keyword is missing, and it can be quite confusing.
I was stuck for some time as I was trying to figure out why I had the following output:

Epoch 1, step 48: val_loss was not in top 1
Epoch 2, step 96: val_loss was not in top 1
Epoch 3, step 144: val_loss was not in top 1

I believed that each epoch the callback was looking at a specific step for the validation loss.

Adding the word global makes it consistent with the message displayed if the model has improved.

I did not open an issue, as it's a simple fix.

Did you have fun?

Make sure you had fun coding 🙃

[codecov]

Codecov Report

Merging #7244 (351c11f) into master (94fcaaf) will decrease coverage by 4%.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master   #7244    +/-   ##
=======================================
- Coverage      91%     87%    -4%     
=======================================
  Files         199     199            
  Lines       12783   12783            
=======================================
- Hits        11662   11150   -512     
- Misses       1121    1633   +512