Fixes bug where scaled values were saved in hyp.yaml, which would cause continuity issues with --resume

save the model.hyp etc to checkpoint when use multi GPU training

Set save_dir = log_dir in train.py

Add example of hyp yaml

In torch==1.5, the import of the API has changed. Although it does not interrupt the operation of the program, it seems to me to be an implicit error and may throw an exception in later versions.