defaults.yaml

# Name of the model. Will be the stem of the input model if undefined.
name: ~

# List of stages for multistage models. Doesn't have to be provided
# for single-stage models.
stages:

  # Name of the stage. Required.
  stage_name:

    # Path to an ONNX or IR model. IR is supported only for RVC2 and RVC3 conversions.
    # It can be either a local file or an S3 url. Required.
    input_model: ~

    # Specify the current input and desired output encoding for image data inputs.
    encoding:
      from: RGB
      to: BGR

    # Configuration for the calibration dataset.
    # Required for Hailo and RVC4, optional for RVC3, ignored for RVC2.
    # If specified, then the `calibration.path` must be provided.
    calibration:
      # Path to the calibration dataset. It can be either a local directory,
      # an S3 url, or an LDF identificator in a form <dataset_name>:<split>.
      # Example:
      # path: models/coco128
      # path: test_name:val
      path: ~

      # Include only the specified number of images from the calibration dataset.
      # Optional.
      max_images: -1

      # What method to use for resizing. Options are:
      #  - CROP: center crop the image to the specified size.
      #  - PAD: resize and pad the image to the specified size.
      #  - RESIZE: resize the image to the specified size, ignoring aspect ratio.
      resize_method: RESIZE

      # For multi-stage models, the calibration can be an output of a previous stage.
      # The `link` section specifies how to use the outputs from the previous stage.
      # link:

        # Name of the stage to link to. Required.
        # stage: ~

        # Name of the output to use. Either `output` or `script` must be provided.
        # output: ~

        # A custom python script to use for post-processing the linked output.
        # It can be either a path or url to a python file, or a string with the script.
        # The script must contain a function
        # `run_script(outputs: dict[str, np.ndarray]) -> np.ndarray` which takes the
        # dictionary of linked outputs and produces a single array.
        # Either `output` or `script` must be provided.
        # script: ~

    # Scale values to be used for the input image per channel.
    # Can be either:
    #  - a single value to be used for all channels.
    #  - a list of values to be used for each channel.
    #  - a "imagenet" string to use the ImageNet values.
    #    Equivalent to [ 58.395, 57.12, 57.375 ]
    scale_values: ~

    # Mean values to be used for the input image per channel.
    # Same options as for scale_values. "imagenet" equivalent to [ 123.675, 116.28, 103.53 ].
    mean_values: ~

    # The input shape of the network. If not provided,
    # it will be inferred from the model if possible.
    shape: ~

    # Lettercode representation of the input layout. e.g. NCHW
    layout: ~

    # Remote path where to upload the compiled model. Optional.
    output_remote_url: ~

    # Keep the intermediate files created during the compilation.
    keep_intermediate_outputs: true

    # Do not run ONNX simplifier on the provided model.
    disable_onnx_simplification: false

    # List of input names with shapes,
    # data types, values for freezing and input modifiers.
    # Overrides the top-level input modifiers.
    # Example:
    # inputs:
    #  - name: input.0  # a node name for hailo, a tensor name of other platforms
    #    shape: [ 1, 3, 256, 256 ]
    #    data_type: float32 (default)
    #    scale_values: [ 255., 255., 255. ]
    #    encoding:
    #      from: RGB
    #      to: BGR
    #  - name: is_training
    #    freeze_value: false
    #  - name: sequence_len
    #    freeze_value: 250
    #
    # Examples of valid data types: 'float32', 'float32', 'uint8'
    inputs: []

    # List of the output names.
    # Example:
    # outputs:
    #   - name: output.0
    #   - name: output.1
    outputs: []

    # --- Hailo-Specific Arguments ---
    hailo:
      # Specifies the optimization level. A number between 0 and 3.
      optimization_level: 2

      # Specifies the compression level. A number between 0 and 4.
      compression_level: 2

      # Batch size for the calibration.
      # batch_size: 8
      batch_size: 8

      # Stop after quantization. Used in tests.
      disable_compilation: false

      # List of additional arguments to pass to the model optimizer.
      alls: []

      # Disables calibration.
      disable_calibration: false

      # Hardware architecture to be used.
      hw_arch: hailo8

    # --- RVC2-Specific Arguments ---
    rvc2:
      # Specifies number of shaves.
      number_of_shaves: 8

      # Specifies number of CMX slices.
      number_of_cmx_slices: 8

      # List of additional arguments to pass to the model optimizer.
      # The additional arguments are passed as-is and always take precedence
      # over the default arguments.
      mo_args: []

      # List of additional arguments to pass to the compile_tool.
      compile_tool_args: []

      # Produces .superblob file instead of regular .blob.
      superblob: true

      # If the original model has FP32 weights or biases, they are compressed to FP16. All intermediate data is kept in original precision.
      compress_to_fp16: true

    # --- RVC3-Specific Argument ---
    rvc3:
      # Target device for POT. Can be one of { VPU, ANY }
      pot_target_device: VPU

      # List of additional arguments to pass to the model optimizer.
      # The additional arguments are passed as-is and always take precedence
      # over the default arguments.
      mo_args: []

      # List of additional arguments to pass to the compile_tool.
      compile_tool_args: []

      # If the original model has FP32 weights or biases, they are compressed to FP16. All intermediate data is kept in original precision.
      compress_to_fp16: true

    # --- RVC4-Specific Arguments ---
    rvc4:
      # List of additional arguments to pass to SNPE onnx-to-dlc.
      # The additional arguments are passed as-is and always take precedence
      # over the default arguments.
      snpe_onnx_to_dlc_args: []

      # List of additional arguments to pass to SNPE snpe-dlc-quant.
      snpe_dlc_quant_args: []

      # List of additional arguments to pass to SNPE snpe-dlc-graph-prepare.
      snpe_dlc_graph_prepare_args: []

      # Whether to include the raw images in the intermediate outputs.
      # Warning: the raw images can get very large.
      keep_raw_images: False

      # Selects per-axis-element quantization for the weights
      # and biases of certain layer types.
      # Only Convolution, Deconvolution, and FullyConnected are supported.
      use_per_channel_quantization: True

      # Enables row wise quantization of Matmul and FullyConnected ops.
      use_per_row_quantization: False