[translate] Add translate ranker model loader.

components/translate/core/browser/proto/translate_ranker_model.proto

Index: components/translate/core/browser/proto/translate_ranker_model.proto
diff --git a/components/translate/core/browser/proto/translate_ranker_model.proto b/components/translate/core/browser/proto/translate_ranker_model.proto
index 4ac58b4d5c7b983e344d9c364b252fd429a701af..23904c07990aa6df88bbcf257f0edcf59749213a 100644
--- a/components/translate/core/browser/proto/translate_ranker_model.proto
+++ b/components/translate/core/browser/proto/translate_ranker_model.proto
@@ -10,7 +10,7 @@ option optimize_for = LITE_RUNTIME;
 
 package chrome_intelligence;
 
-// Defines a Chrome Ranker Translate model.
+// Defines a Chrome Ranker Translate Model

[hamelphi, 2017/01/20 18:44:18]

'.'

[Roger McFarlane (Chromium), 2017/02/08 23:08:09]

Done.

 message TranslateRankerModel {

[hamelphi, 2017/01/20 18:44:18]

IIUC, the structure of this proto will be required to be the same for any
FooRankerModel, except for the list of protos in the |model_iteration|. Am I
correct?

[Roger McFarlane (Chromium), 2017/02/08 23:08:09]

There no technical requirement that other FooRankerModel's follow this pattern.

The expected convention would be that they would either have a oneof list of
contained model implementations or a set of optional model implementation and
some way to discriminate between them.

I would ultimately expect these to all look something like:

message FooRankerModel {
  oneof {
    // Initial release of the FooRankerModel. Supports
    // features X and Y. Deprecated as of DD/MM/YYYY.
    GenericNeuralNetModelProto v1 = 1;

    // Second release of the FooRankerModel. Adds features
    // A and B and renames feature Y to Z.
    GenericNeuralNetModelProto v2 = 1;
  }
}

Or something similar.

   // A number that identifies the version of this model.
   optional uint32 version = 1;
@@ -29,5 +29,6 @@ message TranslateRankerModel {
     map<string, float> country_weight = 7;
     map<string, float> locale_weight = 8;
   }
-  optional LogisticRegressionModel logistic_regression_model = 2;
+
+  oneof model_iteration { LogisticRegressionModel v1 = 2; }

[hamelphi, 2017/01/20 18:44:18]

optional: model_revision maybe?

[hamelphi, 2017/01/20 18:44:18]

This is not backwards compatible. I'd keep the logistic_regression_model until
it is rolled out.

[Roger McFarlane (Chromium), 2017/02/08 23:08:09]

Done.

[Roger McFarlane (Chromium), 2017/02/08 23:08:09]

This whole this isn't backwards compatible now that there's an envelope around
the model.

The current model has to stay where it currently lives and the new model can be
brought up going forward.

 }