Added forward transfer

[AndreaCossu]

This builds on #493 to implement the forward transfer.

The problem is mainly about computing the model accuracy at random initialization on all experiences. This is now done by adding to the BaseStrategy a periodic_eval call before training. The user is expected to:

1. Enable periodic eval
2. Pass the entire test stream to the eval_streams parameter of the strategy.train() call. This is the recommended behavior in any case, because otherwise the evaluation plugin will raise a warning/error, depending on its configuration.

This is not the only way to implement forward transfer. For example, the forward transfer metric could automatically call an eval on the entire test stream to avoid relying too much on the user calls. We can discuss other ways.

In case you think of a better way to proceed, I will wait some days before moving on to fully test this implementation.

This closes #212

[coveralls]

Pull Request Test Coverage Report for Build 1090448588

• 170 of 187 (90.91%) changed or added relevant lines in 7 files are covered.
• 3 unchanged lines in 2 files lost coverage.
• Overall coverage increased (+0.2%) to 82.328%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/training/plugins/evaluation.py	5	10	50.0%
avalanche/evaluation/metrics/forward_transfer.py	137	149	91.95%

Files with Coverage Reduction	New Missed Lines	%
avalanche/training/strategies/base_strategy.py	1	96.8%
avalanche/evaluation/metrics/forgetting_bwt.py	2	91.26%

Totals
Change from base Build 1075424779:	0.2%
Covered Lines:	10468
Relevant Lines:	12715

---

💛 - Coveralls

[vlomonaco]

Nice @AndreaCossu! Can you add also an example about this? thanks :)

[AndreaCossu]

I will add the forward transfer in the example about the evaluation plugin and I will provide a usage example directly in the API doc once we refine it.

[AndreaCossu]

@AntonioCarta what do you think about the periodic_eval call also before training starts? I think this is the main change related to the training module.

[AntonioCarta]

@AntonioCarta what do you think about the periodic_eval call also before training starts? I think this is the main change related to the training module.

I think it's ok for now, but it's an ugly hack. Everyone gets a performance hit because of this. I also don't think it completely solves the problem with metrics. Different metrics will need a different evaluation loop and we will have to adjust it again.