Epoch counting is one-off in multiple instances

[AAnoosheh]

🐛 Bug

Two issues occur:

1. The final epoch does not save a checkpoint during training.
2. Resuming from a checkpoint N will start the epochs at N+2.

Expected behavior

1. Final checkpoint should save a .ckpt file, as usual.
2. Should resume from epoch N+1.

Environment

* CUDA:
	- GPU:
		- Tesla V100-DGXS-16GB
		- Tesla V100-DGXS-16GB
		- Tesla V100-DGXS-16GB
		- Tesla V100-DGXS-16GB
	- available:         True
	- version:           10.1
* Packages:
	- numpy:             1.18.1
	- pyTorch_debug:     False
	- pyTorch_version:   1.6.0
	- pytorch-lightning: 0.9.0rc12
	- tensorboard:       2.2.1
	- tqdm:              4.46.1
* System:
	- OS:                Linux
	- architecture:
		- 64bit
		-
	- processor:         x86_64
	- python:            3.7.7
	- version:           #113-Ubuntu SMP Thu Jul 9 23:41:39 UTC 2020

[awaelchli]

@AAnoosheh Honestly I do not understand how the PR you linked relates to the bug your report. Did you mean to link another issue?

The final epoch does not save a checkpoint during training.

I don't experience this. The epoch number is 0-indexed, and by default it only saves best checkpoints. Could one of these reasons be why you may think this is a bug?

How can I reproduce the second issue?

[AAnoosheh]

Sorry I should have clarified I use the following to save every epoch:

pl.callbacks.ModelCheckpoint(save_top_k=-1, verbose=True)

The second is done via Trainer(resume_from_checkpoint=some_ckpt_file)

I assume some change was made to move epochs to 0-index, when previously they were 1-indexed, and there's a mismatch now.

EDIT:
I also have no idea how a PR was linked in my comment. Those numbers came out of nowhere from the auto-generated issue template.

[ananyahjha93]

@AAnoosheh so when you run pl.callbacks.ModelCheckpoint(save_top_k=-1, verbose=True) all the checkpoints are saved, however we do not save the last one as 'last.ckpt'. Also, the checkpoints are numbered from 0, so if you run for 4 epochs, the last checkpoint saved will be 'epoch=3.ckpt' and when you resume, it resumes from the expected 5th epoch.

Updating tests and code for this