Scorer model loading

[graemenail]

Description

When creating scorers, there are two calls to loadItems:

1. getYamlFromModel loads the 'special' YAML from the model
2. When the scorer is created, a EncoderDecoder method results in a call to ExpressionGraph::load

This PR addresses these by calling io::loadItems in advance and passing the Item vector when creating scorers.

1. getYamlFromModel can now be called directly from an Item vector
2. ScorerWrapper can be constructed from an Item vector.

Amun and Nematus do some specific preprocessing before loading the graph, this PR moves that logic into a load from Items.

Closes #831

List of changes:

• Adds --model-mmap as a command line option (uses mio)
• Preloads model to Item vector before creating scorers
• try-catch when loading model YAML is replaced with if-else
• Extended IEncoderDecoder to load from Item vector
• Amun and Nematus models load overload for Item vector

Added dependencies: None

How to test

Used marian-conv to test mmap loading
Loading from Items passes regression tests

Checklist

• I have tested the code manually
• I have run regression tests
• I have read and followed CONTRIBUTING.md
• I have updated CHANGELOG.md

[snukky]

Looks good, only minor comments from me. Thanks! I will ask @emjotde to take a look too.

[snukky]

Please add an entry about the model-mmap option to CHANGELOG.

[snukky]

This regression test fails for me with this PR:

tests/training/restarting/test_restarting_finished.sh

It fails because logs no longer have the message saying 'Loading model' when trying to restart a training that has already finished. @graemenail, Do you think this is expected now? If so, I will just update the test.

This was tested after rebasing with the current master. That's the only regression test that fails.

[graemenail]

This regression test fails for me with this PR:

tests/training/restarting/test_restarting_finished.sh

It fails because logs no longer have the message saying 'Loading model' when trying to restart a training that has already finished. @graemenail, Do you think this is expected now? If so, I will just update the test.

This was tested after rebasing with the current master. That's the only regression test that fails.

Thanks for looking at this @snukky. I believe this is expected, the Loading model from {} log messages contained the filename of the model loaded. In this PR, the model is loaded beforehand, and so the filename is replaced with the vector<io:Items> from that file.

That being said, I think it's a useful log message to include. I'll see where is best to reinstate it, and update this PR.

[graemenail]

I have reinstated the log messages for Amun and Nematus, which should fix the regression.
I have also added log messages when loading model ahead of createScorer as well.