将导出的ONNX模型输入从float转变成double类型

model_onnx = convert_sklearn(
    pipe,
    "pipeline_catboost",
    [("input", DoubleTensorType([None, 33]))], 
    target_opset={"": 12, "ai.onnx.ml": 2}, final_types = [("output_label", Int64TensorType([None,])),("output_probability1", DoubleTensorType([None,2]))]
)

可能会遇到的报错：

Load model from pipeline_lightgbm.onnx failed:Type Error: Type (seq(map(int64,tensor(double)))) of output arg (output_probability) of node (ZipMap) does not match expected type (seq(map(int64,tensor(float)))).

然后尝试将
final_types = [(“output_label”, Int64TensorType([None,])),(“output_probability1”, DoubleTensorType([None,2]))]
改成 final_types = [(“output_label”, Int64TensorType([None,])),(“output_probability1”, FloatTensorType([None,2]))]
仍报错

InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from sklrfc.onnx failed:This is an invalid model. Type Error: Type ‘seq(map(int64,tensor(float)))’ of input parameter (output_probability) of operator (Cast) in node (Cast2) is invalid.

加上 在convert_sklearn函数中加上 options={‘zipmap’: False}即解决

大概的代码：

model_onnx = convert_sklearn(
    pipe,
    "pipeline_catboost",
    [("input", DoubleTensorType([None, 33]))], 
    target_opset={"": 12, "ai.onnx.ml": 2}, final_types = [("output_label", Int64TensorType([None,])),("output_probability1", FloatTensorType([None,2]))]
    , options={'zipmap': False}
)

sess = rt.InferenceSession("pipeline_catboost.onnx", providers=["CPUExecutionProvider"])

pred_onx = sess.run(None, {"input": X[:5].values.astype(numpy.double)})

参考：https://stackoverflow.com/questions/69351596/final-types-for-randomforestclassifier-skl2onnx/69394645#69394645