Skip to content

Latest commit

 

History

History
executable file
·
36 lines (25 loc) · 1.37 KB

paddle_frontend_architecture.md

File metadata and controls

executable file
·
36 lines (25 loc) · 1.37 KB
Raw

OpenVINO™ Paddle Frontend Architecture

The diagram below shows the Paddle Frontend architecture and its interaction with OpenVINO Frontend API and Core APIs.

flowchart TB
    fw_model[(Paddle Model)]
    style fw_model fill:#427cb0
    
    protobuf([protobuf])
    subgraph frontend [ov::frontend::paddle::FrontEnd]
        load_impl["load_impl()"]
    end
    fw_model--as stream-->load_impl
    load_impl--load stream-->protobuf
    protobuf--parsed object-->load_impl
    
    
    subgraph input_model [ov::frontend::paddle::InputModel]
        convert["convert()"]
    end
    
    load_impl--create-->input_model
    
    ov_model[ov::Model]
    
    convert--recursively parse all operations from the model-->ov_model
Loading
  • The input to the Paddle Frontend is a PaddlePaddle protobuf model, and the output is the ov::Model which is semantically equivalent to the input.
  • Paddle Frontend is an implementation of the OpenVINO Frontend, which implements two main interfaces. The first interface is load, which reads a Paddle protobuf model and represents it using InputModel with OpPlaces and TensorPlaces. The second interface is convert, which represents the InputModel with ov::Model by semantically mapping these Places to the OpenVINO opset.

See also