scs4onnx

A very simple tool that compresses the overall size of the ONNX model by aggregating duplicate constant values as much as possible. Simple Constant value Shrink for ONNX.

Key concept

• If the same constant tensor is found by scanning the entire graph for Constant values, it is aggregated into a single constant tensor.
• Ignore scalar values.
• Ignore variables.
• Finally, create a Fork of onnx-simplifier and merge this process just before the onnx file output process -> Temporarily abandoned because it turned out that the onnx-simplifier specification needed to be changed in a major way.
• Implementation of a specification for separating the weight of a specified OP name to an external file.
• Implementation of a specification for separating the weight of a specified Constant name to an external file.
• Added option to downcast from Float64 to Float32 and INT64 to INT32 to attempt size compression.
• Post an issue of onnx-simplifier. Excessive bloating of ONNX files due to over-efficient conversion of "Tile" to constants (Protocol Buffers .onnx > 2GB) #178
• Add sample onnx models.

1. Setup

1-1. HostPC

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

### run
$ pip install -U onnx \
&& python3 -m pip install -U onnx_graphsurgeon --index-url https://pypi.ngc.nvidia.com \
&& pip install -U scs4onnx

1-2. Docker

### docker pull
$ docker pull pinto0309/scs4onnx:latest

### docker build
$ docker build -t pinto0309/scs4onnx:latest .

### docker run
$ docker run --rm -it -v `pwd`:/workdir pinto0309/scs4onnx:latest
$ cd /workdir

2. CLI Usage

$ scs4onnx -h

usage:
  scs4onnx [-h]
  [--mode {shrink,npy}]
  [--forced_extraction_op_names FORCED_EXTRACTION_OP_NAMES]
  [--forced_extraction_constant_names FORCED_EXTRACTION_CONSTANT_NAMES]
  [--disable_auto_downcast]
  [--non_verbose]
  input_onnx_file_path output_onnx_file_path


positional arguments:
  input_onnx_file_path
                        Input onnx file path.
  output_onnx_file_path
                        Output onnx file path.

optional arguments:
  -h, --help
                        show this help message and exit
  --mode {shrink,npy}
                        Constant Value Compression Mode.
                        shrink: Share constant values inside the model as much as possible.
                                The model size is slightly larger because
                                some shared constant values remain inside the model,
                                but performance is maximized.
                        npy:    Outputs constant values used repeatedly in the model to an
                                external file .npy. Instead of the smallest model body size,
                                the file loading overhead is greater.
                        Default: shrink
  --forced_extraction_op_names FORCED_EXTRACTION_OP_NAMES
                        Extracts the constant value of the specified OP name to .npy
                        regardless of the mode specified.
                        Specify the name of the OP, separated by commas.
                        Cannot be used with --forced_extraction_constant_names at the same time.
                        e.g. --forced_extraction_op_names aaa,bbb,ccc
  --forced_extraction_constant_names FORCED_EXTRACTION_CONSTANT_NAMES
                        Extracts the constant value of the specified Constant name to .npy
                        regardless of the mode specified.
                        Specify the name of the Constant, separated by commas.
                        Cannot be used with --forced_extraction_op_names at the same time.
                        e.g. --forced_extraction_constant_names aaa,bbb,ccc
  --disable_auto_downcast
                        Disables automatic downcast processing from Float64 to Float32 and INT64
                        to INT32. Try enabling it and re-running it if you encounter type-related
                        errors.
  --non_verbose
                        Do not show all information logs. Only error logs are displayed.

3. In-script Usage

$ python
>>> from scs4onnx import shrinking
>>> help(shrinking)

Help on function shrinking in module scs4onnx.onnx_shrink_constant:

shrinking(
  input_onnx_file_path: Union[str, NoneType] = '',
  output_onnx_file_path: Union[str, NoneType] = '',
  onnx_graph: Union[onnx.onnx_ml_pb2.ModelProto, NoneType] = None,
  mode: Union[str, NoneType] = 'shrink',
  forced_extraction_op_names: List[str] = [],
  forced_extraction_constant_names: List[str] = [],
  disable_auto_downcast: Union[bool, NoneType] = False
  non_verbose: Union[bool, NoneType] = False
) -> Tuple[onnx.onnx_ml_pb2.ModelProto, str]

    Parameters
    ----------
    input_onnx_file_path: Optional[str]
        Input onnx file path.
        Either input_onnx_file_path or onnx_graph must be specified.

    output_onnx_file_path: Optional[str]
        Output onnx file path.
        If output_onnx_file_path is not specified, no .onnx file is output.

    onnx_graph: Optional[onnx.ModelProto]
        onnx.ModelProto.
        Either input_onnx_file_path or onnx_graph must be specified.
        onnx_graph If specified, ignore input_onnx_file_path and process onnx_graph.

    mode: Optional[str]
        Constant Value Compression Mode.
        'shrink': Share constant values inside the model as much as possible.
            The model size is slightly larger because some shared constant values remain
            inside the model, but performance is maximized.
        'npy': Outputs constant values used repeatedly in the model to an external file .npy.
            Instead of the smallest model body size, the file loading overhead is greater.
        Default: shrink

    forced_extraction_op_names: List[str]
        Extracts the constant value of the specified OP name to .npy
        regardless of the mode specified.
        Cannot be used with --forced_extraction_constant_names at the same time.
        e.g. ['aaa','bbb','ccc']

    forced_extraction_constant_names: List[str]
        Extracts the constant value of the specified Constant name to .npy
        regardless of the mode specified.
        Cannot be used with --forced_extraction_op_names at the same time.
        e.g. ['aaa','bbb','ccc']

    disable_auto_downcast: Optional[bool]
        Disables automatic downcast processing from Float64 to Float32 and INT64 to INT32.
        Try enabling it and re-running it if you encounter type-related errors.
        Default: False

    non_verbose: Optional[bool]
        Do not show all information logs. Only error logs are displayed.
        Default: False

    Returns
    -------
    shrunken_graph: onnx.ModelProto
        Shrunken onnx ModelProto

    npy_file_paths: List[str]
        List of paths to externally output .npy files.
        An empty list is always returned when in 'shrink' mode.

3. CLI Execution

$ scs4onnx input.onnx output.onnx --mode shrink

4. In-script Execution

4-1. When an onnx file is used as input

If output_onnx_file_path is not specified, no .onnx file is output.

from scs4onnx import shrinking

shrunk_graph, npy_file_paths = shrinking(
  input_onnx_file_path='input.onnx',
  output_onnx_file_path='output.onnx',
  mode='npy',
  non_verbose=False
)

4-2. When entering the onnx.ModelProto

onnx_graph If specified, ignore input_onnx_file_path and process onnx_graph.

from scs4onnx import shrinking

shrunk_graph, npy_file_paths = shrinking(
  onnx_graph=graph,
  mode='npy',
  non_verbose=True
)

5. Sample

5-1. shrink mode sample

• 297.8MB -> 67.4MB (.onnx)

  $ scs4onnx gmflow_sintel_480x640.onnx gmflow_sintel_480x640_opt.onnx
  

  

• 1.8GB -> 886.8MB (.onnx)

  $ scs4onnx hitnet_sf_finalpass_720x960.onnx hitnet_sf_finalpass_720x960_opt.onnx
  

  

• 1.8GB -> 2.1MB (.onnx) + 884.7MB (.npy)

  $ scs4onnx \
  hitnet_sf_finalpass_720x960.onnx \
  hitnet_sf_finalpass_720x960_opt.onnx \
  --forced_extraction_op_names GatherElements_660
  

  

• 297.8MB -> 21.3MB (.onnx) + 46.1MB (.npy)

  $ scs4onnx \
  gmflow_sintel_480x640.onnx \
  gmflow_sintel_480x640_opt.onnx \
  --forced_extraction_constant_names 1646
  

  

5-2. npy mode sample

• 297.8MB -> 21.3MB (.onnx)

  

5-3. .npy file view

$ python
>>> import numpy as np
>>> param = np.load('gmflow_sintel_480x640_shrunken_exported_1646.npy')
>>> param.shape
(8, 1200, 1200)
>>> param
array([[[   0.,    0.,    0., ...,    0.,    0.,    0.],
        [   0.,    0.,    0., ...,    0.,    0.,    0.],
        [   0.,    0.,    0., ...,    0.,    0.,    0.],
        ...,
        [-100., -100., -100., ...,    0.,    0.,    0.],
        [-100., -100., -100., ...,    0.,    0.,    0.],
        [-100., -100., -100., ...,    0.,    0.,    0.]]], dtype=float32)

6. Sample ONNX models

1. gmflow_sintel_480x640.onnx - Optical flow calculation
2. hitnet_sf_finalpass_720x960.onnx - Stereo depth estimation

7. Reference

1. https://docs.nvidia.com/deeplearning/tensorrt/onnx-graphsurgeon/docs/index.html
2. https://github.com/NVIDIA/TensorRT/tree/main/tools/onnx-graphsurgeon
3. https://github.com/PINTO0309/sne4onnx
4. https://github.com/PINTO0309/snd4onnx