jaganseshadri / stfsclient Goto Github PK

The protobuf wrapper class mimic the structure of the message defined inside the .proto files.
Pros:

• Provides namespace organization synonymous with the protobuf definition.
  Cons:
• Loses out on namespace abstraction that makes it easier for people who don't know either gRPC or protobuf or both to work with it, i.e. it's still laborious to know which message needs what attributes to be set.

Proposed solution:

1. More documentation would definitely help.
2. Flatten out the protobuf definition, i.e. bring out some nested attributes to the parent level.

The library doesn't work well with TensorFlow or Keras or with any package that can import TensorFlow into a common namespace. The reason is because of protobufs that are used in this library that are common with tensorflow.

The way the solution to this is implemented to get around this was to default to the TensorFlow's implementation if available otherwise use the internal implementation.

However, it's not currently possible to programmatically force the TensorFlow backend.