Hey there,
My name is Roy Sadaka, and I used to design and develop a DL model with TensorFlow 
, few months ago I moved to PyTorch 
and love it so far.
I’ve recently open-sourced my little project - lpd
It’s a structured trainer with some common callbacks support, I use it for production needs so it’s fully maintained.
I’m well aware that there are other, more mature alternatives out there, I’ve created this project to give myself the highest level of control over the training process, now you can use it too 
lpd provides a structured trainer where you can train/validate/test/predict your model, it also collects statistics during training and prints them at the end of every epoch (you have full control when/if to print them). you can save/load the whole trainer and continue training at a later time.
It also exposes some common callbacks such as ModelCheckPoint, EarlyStopping, SchedulerStep, and Tensorboard, and you can create custom callbacks with ease while still maintain full control.
lpd is currently available on pip or you can visit the source code
Feel free to contact me if you think something is missing in lpd
Roy