It's slightly odd to copy the caffe.proto source for distribution. Are you compiling the proto definition for java separately or do you have different bindings than those generated by the usual protoc? The java protobuf compilation could be added to the build.

On the other hand shipping caffe.proto seems harmless.

Yes, we process the proto files in java as well, mainly for the input data layer. For example, we get the source location from the memory data layer. The location is typically an HDFS URI and is the base for a Spark RDD.

@shelhamer In our use case, we need Java class generated from caffe.proto at this point. Python binding may be needed later. It is okay for us if Caffe distribute those bindings for caffe.proto. However, it will be easier to let users handle them directly.

Ok this is fine, but to quote my earlier comment:

The java protobuf compilation could be added to the build

in the same way the Python proto bindings are built.