Refactor data layers to avoid duplication of data transformation code

[arntanguy]

As discussed in #941, I started refactoring the data layers to avoid the current code duplication of data
transformations.

• I ceated a new protobuf message TransformerParameter that contains the transformation parameters
  (crop, mirror, mean file...)
• I created a new class DataTransformer that takes care of applying the transformation defined in the protobuf file to the data
• I refactored DataLayer to use this new class.
• I updated the code to update the old network definition files to the new one
• I updated the unit tests

Could @shelhamer or someone else could have a look at it before I go about refactoring the other
data layers?

[sguada]

@GeenuX Thanks for your initial commit, could you also update the ImageDataLayer to use the same DataTransformer?

[bhack]

@sguada Is this superseding #569?

[jeffdonahue]

Thanks for working on this @GeenuX, it will be great to have the madness that is the current set of DataLayers cleaned up.

Since commenting on @sguada's #569 I've changed my mind a bit...I think each of the transformations actually could be done as layers with little or no extra cost in terms of memory/time. The key would be to use @longjon's Blob::set_data to make the input blob point directly to the data field of the BlobProto, thereby avoiding any additional explicit copying. set_data would need to be called for each input datum, but it's cheap as it just modifies a pointer.

The current design looks good and would definitely be an improvement over the current state as well, but is slightly less flexible than having each transformation be a layer. We could modify this to use layers later but might end up having to make proto-breaking changes, which isn't ideal. I'll leave it up to @GeenuX to decide if/how to take this suggestion.

[arntanguy]

I've refactored the ImageDataLayer as well.

I agree, it would be nice to have the transformations defined as separate layers. I'm not sure
I'll have the time to look into it though. In the meanwhile, I think this ought to be merged, as it gets rid of all the duplicated code between Data and Image layers.

[shelhamer]

@GeenuX I agree this is a worthwhile improvement in itself. Thanks for your further work to improve the data layers and make data processing more sane.

[shelhamer]

@sguada I will let you decide how to carry on in #569. While the DataTransformer is a step in the right direction, it would be better still to have transformation layers along the lines of #954 (comment).

I will follow-up by refactoring WindowDataLayer and MemoryDataLayer; in the latter's case I will add data transformations and expose its use for deployment nets.

[shelhamer]

Don't panic, but this breaks model definitions in dev as pointed out by @jeffdonahue.

I will follow up shortly with a fix that #963 fixes this by

1. restoring the prototxt field numbers: these should never change!
2. adding a data layer parameter -> transform parameter upgrade path to Net::Init()
3. updating the model definitions to the new format

Sorry for merging before correcting these issues.

[arntanguy]

Sorry about that! Thanks for the fix!