GitLab

* solving bug where for small epochs and large gradient_accumulation_steps we never train

* black formatting

* no need to change these files