Onnxsim

Context

Add the proptotype of an onnx simplifier, a framework to simplify ONNX graph.

This tool can be used with a CLI:

aidge_sim shuffle_layer.onnx shuffle_layer_sim.onnx

┏━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━┓
┃            ┃ Original Model ┃ Simplified Model ┃
┡━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━┩
│ Concat     │              2 │                0 │
│ Constant   │             15 │                0 │
│ Div        │              1 │                0 │
│ Gather     │              4 │                0 │
│ Identity   │              1 │                0 │
│ Reshape    │              2 │                2 │
│ Shape      │              4 │                0 │
│ Transpose  │              1 │                1 │
│ Unsqueeze  │              9 │                0 │
│ Model Size │         3.6KiB │           492.0B │
└────────────┴────────────────┴──────────────────┘

Here are the arguments supported by this tool:

aidge_sim --help
usage: aidge_sim [-h] [-v] [--show_recipes] [--input-shape INPUT_SHAPE [INPUT_SHAPE ...]]
                 [input_model] [output_model]

positional arguments:
  input_model           Path to the input ONNX model.
  output_model          Path to save the simplified ONNX model.

options:
  -h, --help            show this help message and exit
  -v, --verbose         Set the verbosity level of the console output. Use -v to increase
                        verbosity, with the following levels in ascending order: default:
                        WARN - Only warnings and higher (WARN, ERROR, FATAL) are displayed.
                        -v: NOTICE - Notices and higher (NOTICE, WARN, ERROR, FATAL) are
                        displayed. -vv INFO - Informational messages and higher (INFO,
                        NOTICE, WARN, ERROR, FATAL) are displayed. -vvv: DEBUG - All
                        messages, including debug information, are displayed. Available
                        levels in descending order of severity: DEBUG < INFO < NOTICE < WARN
                        < ERROR < FATAL.
  --show_recipes        Show available recipes and exit
  --input-shape INPUT_SHAPE [INPUT_SHAPE ...]
                        Overwrite the input shape. The format is
                        "input_name:dim0,dim1,...,dimN" or simply "dim0,dim1,...,dimN" when
                        there is only one input. For example, "data:1,3,224,224" or
                        "1,3,224,224". input_name correspond to the name of the output of
                        the input node. This convention is due to ONNX name being not
                        mandatory. If the output name begin by a number, you need to prepend
                        it with "data_". If one node take multiple input you can add it
                        multiple time, the order is determined by the order of the arguments
                        provided. For example, "--input-shape data:1,3 data:1,10". Here the
                        first input will be of size "1,3" and the second "1,10". Note: you
                        might want to use some visualization tools like netron to make sure
                        what the input name and dimension ordering (NCHW or NHWC) is.

TODO:

• Constant and Shape folding
• Add a better UI to see what has been changed
• Add pre-existing aidge recipes:
  • Fuse Batchnorm
• Add custom recipies:
  • Remove Identity node
• Package python module so that you can call it as a script (aidge_sim --help)

In future MR

• Allow to skip transformation using the CLI #57 (closed)
• Add a fuse to LayerNorm recipe (tailored to fuse LayerNorm from PyTorch and Keras) aidge_core#248 (closed)
• Create a tutorial to showcase the functionnality aidge#249

Changes

• Add support for ONNX export of Div operator
• Fix the export of the following operators:
  • Gather: export failed when indices were attribute instead of producers (+attributes are now scalar for better readabilitiy in netron)
  • Squeeze: Typo in attribute name (+attributes are now scalar for better readabilitiy in netron)
  • Unsqueeze: Typo in attribute name (+attributes are now scalar for better readabilitiy in netron)
  • Shape: end attribute was set by default at -1 which did not result in the intended behavior. Corrected this issue by letting it to default (undefined) which is the behavior of aidge value -1 (see comment in code 116aab01 if this explanation is too confusing).
• For ONNX export, ONNX output type is now no longer always FLOAT32 but instead derived from the dtype of the Aidge::Tensor
• Fix small typo in doc...