Converting a Sklearn Isolation Forrest to ONNX Model

Today I was faced with a really tricky error when trying to convert my Isolation Forrest model to an ONNX model. I hope this article can save you some time if you are faced with the same situation.

The Scenario

I had a really simple example of fitting an Isolation Forrest and then converting it to an ONNX model as you can see below.

from skl2onnx import to_onnx
from sklearn.ensemble import IsolationForest

model = IsolationForest()
model.fit(X)

model_onnx = to_onnx(
      model,
      initial_types=[('input', FloatTensorType([None, X.shape[1]]))]
)

This led me to the following error:

RuntimeError: The model is using version 4 of domain 'ai.onnx.ml' not supported yet by this library. You need to specify target_opset={'ai.onnx.ml': 3}

The solution approach

First I checked that I had the latest release of skl2onnx, which I did (1.16.0). Then I just followed the recommendation in the log message:

model_onnx = to_onnx(
   model, 
   initial_types=[('input', FloatTensorType([None, temp_scaled_np.shape[1]]))], 
   target_opset={'ai.onnx.ml': 3}
)

Unfortunately, this resulted in the following error:

RuntimeError: op_version must be specified.

I tried to fix this by extending target_opset to {'ai.onnx.ml': 3, 'ai.onnx': 15} (according to ONNX versioning overview). However, this still resulted in the same error.

After some research I found out that setting the opset to '': 15 is the key to success here. So the following statement allowed me to create the ONNX model.

model_onnx = to_onnx(
   model,
   initial_types=[('input', FloatTensorType([None, temp_scaled_np.shape[1]]))], 
   target_opset={'ai.onnx.ml': 3, 'ai.onnx': 15, '':15}
)