sed4onnx

Simple ONNX constant encoder/decoder.

PINTO0309/simple-onnx-processing-tools

Key concept

• Since the constant values in the JSON files generated by onnx2json are Base64-encoded values, ASCII <-> Base64 conversion is required when rewriting JSON constant values.
• After writing the converted Base64 strings to JSON using this tool, json2onnx can be used to regenerate the constant-modified ONNX file.

1. Setup

1-1. HostPC

### option
$ echo export PATH="~/.local/bin:$PATH" >> ~/.bashrc \
&& source ~/.bashrc

### run
$ pip install -U sed4onnx

1-2. Docker

PINTO0309/simple-onnx-processing-tools

2. CLI Usage

$ sed4onnx -h

usage:
    sed4onnx [-h]
    -cs CONSTANT_STRING
    [-d {float16,float32,float64,uint8,int8,int16,int32,int64,string}]
    [-m {encode,decode}]

optional arguments:
  -h, --help
        show this help message and exit.

  -cs CONSTANT_STRING, --constant_string CONSTANT_STRING
        Strings to be encoded and decoded for ONNX constants.

  -d {float16,float32,float64,uint8,int8,int16,int32,int64,string}, \
    --dtype {float16,float32,float64,uint8,int8,int16,int32,int64,string}
        Data type.

  -m {encode,decode}, --mode {encode,decode}
        encode: Converts the string specified in constant_string to a Base64 format string
                that can be embedded in ONNX constants.
        decode: Converts a Base64 string specified in constant_string to ASCII like
                Numpy string or pure string.

3. In-script Usage

>>> from sed4onnx import encode
>>> from sed4onnx import decode
>>> help(encode)

Help on function encode in module sed4onnx.onnx_constant_encoder_decoder:

encode(constant_string: str) -> str

    Parameters
    ----------
    constant_string: str
        ASCII string to be encoded.

    dtype: str
        'float16' or 'float32' or 'float64' or 'uint8'
        or 'int8' or 'int16' or 'int32' or 'int64' or 'string'

    Returns
    -------
    encoded_string: str
        Base64-encoded ASCII string.


>>> help(decode)
Help on function decode in module sed4onnx.onnx_constant_encoder_decoder:

decode(constant_string: str, dtype: str) -> numpy.ndarray

    Parameters
    ----------
    constant_string: str
        Base64 string to be decoded.

    dtype: str
        'float16' or 'float32' or 'float64' or 'uint8'
        or 'int8' or 'int16' or 'int32' or 'int64' or 'string'

    Returns
    -------
    decoded_ndarray: np.ndarray
        Base64-decoded numpy.ndarray.

4. CLI Execution

$ sed4onnx \
--constant_string [-1,3,224,224] \
--dtype int64 \
--mode encode

$ sed4onnx \
--constant_string '//////////8DAAAAAAAAAOAAAAAAAAAA4AAAAAAAAAA=' \
--dtype int64 \
--mode decode

5. In-script Execution

from sed4onnx import encode
from sed4onnx import decode

base64_string = encode(
  constant_string='[-1,3,224,224]',
  dtype='int64',
)

numpy_ndarray = decode(
  constant_string='//////////8DAAAAAAAAAOAAAAAAAAAA4AAAAAAAAAA=',
  dtype='int64',
)

6. Sample

$ sed4onnx \
--constant_string [-1,3,224,224] \
--dtype int64 \
--mode encode

//////////8DAAAAAAAAAOAAAAAAAAAA4AAAAAAAAAA=


$ sed4onnx \
--constant_string '//////////8DAAAAAAAAAOAAAAAAAAAA4AAAAAAAAAA=' \
--dtype int64 \
--mode decode

[-1,3,224,224]

7. Reference

1. https://github.com/onnx/onnx/blob/main/docs/Operators.md
2. https://docs.nvidia.com/deeplearning/tensorrt/onnx-graphsurgeon/docs/index.html
3. https://github.com/NVIDIA/TensorRT/tree/main/tools/onnx-graphsurgeon
4. PINTO0309/simple-onnx-processing-tools
5. PINTO0309/PINTO_model_zoo

8. Issues

https://github.com/PINTO0309/simple-onnx-processing-tools/issues