quark.onnx.tools.convert_shared_initializer_to_unique

Convert ONNX ModelProto with shared initializer to be unique initializer.

param model:

ONNX ModelProto object

param op_types:

The op_type need to copy the shared initializer

return:

converted ONNX ModelProto object

Examples:

Example 1: Convert ONNX ModelProto object:
from quark.onnx.tools import convert_shared_initializer_to_unique
new_onnx_model = convert_shared_initializer_to_unique.convert(onnx_model)

Use the convert_shared_initializer_to_unique.py to duplicate reused initializer in onnx model, so that there do not exist nodes to share initializer

` python convert_shared_initializer_to_unique.py --input $ONNX_MODEL_PATH_WITH_INIT_SHARED --output $ONNX_MODEL_PATH_WITHOUT_INIT_SHARED --op_types ["Cnv", "Gemm"] `

If need to duplicate all op_types in the given onnx model, the op_types could include all op_types or keep None.

` python convert_shared_initializer_to_unique.py --input $ONNX_MODEL_PATH_WITH_INIT_SHARED --output $ONNX_MODEL_PATH_WITHOUT_INIT_SHARED `

The conversion from reused initializer to that one without initializer shared for given node op_types e.g. [“Conv”, “Gemm”]. Empty list [] will include all op_types in the given onnx model, default is []. It is recommended to do conversion to satisfy the compilation need and model quantization FastFinetune need.