cl-onnx

CLI-Complete ONNX Graph Manipulator for Common Lisp (and more!).

Disclaimer

• cl-onnx is in the early stage of development, has not been fully tested, and is unstable.

• I do not intend to make this a stable product because I am not confident that I will continue to update the accompanying useful utilities and other features.

  • If you are going to use cl-onnx just for reading .onnx graph, it is safer for your product to use cl-protobufs

  • And here you can find how to unpack raw_data existing in TensorProto.

• Never use it for products as we cannot guarantee its quality.

• I think cl-onnx is dedicated to my personal projects.

Workload

• Protobuf <-> Common Lisp Objects

• Implement DAG Visualizer (like CLI Netron), which is available on Emacs/REPL (i.e.: ncurses is unavailable.).

• Find Opset Documentation on REPL

• C2FFI Based Binding for ONNX

• utils (and useful macros) for accessing onnx graph in common lisp

Features

CLI Netron

I am too lazy to open netron to check the small graphs :< since we are already using a wonderful tool called REPL.

So I made a tool to draw the DAG in the REPL. ncurses is not used, so it should help you with fast deploying even in Emacs/Lem.

CL-USER> (onnx:viewnode (onnx:graph (onnx:load-model "./dummy_graph/layernorm.onnx")))
                       ┌─────┐                       
                       │input│                       
                       └──┬──┘────────────────┐      
                          │                   │      
            ┌─────────────┴────────────┐      │      
            │ReduceMean (/ReduceMean)  │      │      
            ├──────────────────────────┤      │      
            │axes     ?                │      │      
            │keepdims 1                │      │      
            └─────────────┬────────────┘      │      
                          │                   │      
                  ┌───────┴────┐──────────────┘      
                  │Sub (/Sub)  │                     
                  └──────┬─────┘──────────────┐      
                         │                    │      
                  ┌──────┴─────┐              │      
                  │Pow (/Pow)  │              │      
                  └──────┬─────┘              │      
                         │                    │      
           ┌─────────────┴──────────────┐     │      
           │ReduceMean (/ReduceMean_1)  │     │      
           ├────────────────────────────┤     │      
           │axes     ?                  │     │      
           │keepdims 1                  │     │      
           └──────────────┬─────────────┘     │      
                          │                   │      
                  ┌───────┴────┐              │      
                  │Add (/Add)  │              │      
                  └──────┬─────┘              │      
                         │                    │      
                 ┌───────┴──────┐             │      
                 │Sqrt (/Sqrt)  │             │      
                 └───────┬──────┘             │      
                         │                    │      
                  ┌──────┴─────┐──────────────┘      
                  │Div (/Div)  │                     
                  └──────┬─────┘                     
                         │                           
                 ┌───────┴─────┐                     
                 │Mul (/Mul)   │                     
                 ├─────────────┤                     
                 │gamma (10)   │                     
                 └──────┬──────┘                    
                        │                            
                 ┌──────┴───────┐                    
                 │Add (/Add_1)  │                    
                 ├──────────────┤                    
                 │beta (10)     │                    
                 └───────┬──────┘                    
                         │                           
                      ┌──┴───┐                       
                      │output│                       
                      └──────┘

Requirements

• Roswell

• qlot

• protobuf

• I have confirmed cl-onnx works on macOS/Linux. I dunno about Windows😎.

Installing (TODO)

• do not forget to update all submodules in advance.

$ make build