Is your feature request related to a problem? Please describe.
Currently, there is no clean way, to continue training past its initial epochs due to the fact that the state of the trained is being loaded by CheckpointLoader.

Describe the solution you'd like
A quick solution would be to assume that if this case appears the code of the user will account for those things (aka just gamma decay, etc) and the CheckpointLoader should only modify the max_epochs after loading the checkpoint data and issue a Warning through a logger.

Describe alternatives you've considered
None

Additional context
None