Why `RecursionLimit = 10` in `OnnxTransformer`?

[s-tory]

What is the purpose that Google.Protobuf.CodedInputStream.RecursionLimit is set 10 at the following code?

machinelearning/src/Microsoft.ML.OnnxTransformer/OnnxUtils.cs

Line 208 in 2a6cf9d

using (var codedStream = Google.Protobuf.CodedInputStream.CreateWithLimits(modelStream, Int32.MaxValue, 10))

---

It is set 100 by default in protocol-buffers C#-wrapper.

https://github.com/protocolbuffers/protobuf/blob/10599e6c8dde8a9875258e03054a696d53cadebd/csharp/src/Google.Protobuf/CodedInputStream.cs#L83

        internal const int DefaultRecursionLimit = 100;

---

I could not load some network / .onnx file by the following exception be caused by RecursionLimit = 10 without the monkey-patched Microsoft.ML.OnnxTransformer.dll.

Google.Protobuf.InvalidProtocolBufferException: Protocol message had too many levels of nesting.  May be malicious.  Use CodedInputStream.SetRecursionLimit() to increase the depth limit.

I think that the networks are including Inception construction especially.
(ex. tensorflow/models`s Faster-RCNN-Inception-V2 converted)

Thank you for coding great tools!

[michaelgsharp]

Thanks for bringing this up. Let me discuss this with my team and see if we can figure out why we were setting this to 10.

[s-tory]

@michaelgsharp
Did you figure out anything?

[darth-vader-lg]

I'm having exactly the same problem with one my efficientdet-d1 custom object detection model by TensorFlow 2.4.1 (the same model built as SSD Mobile Net V2 doesn't have any problem).
Just for now I exported it as a frozen_graph.pb and all is ok.
Frozen because with saved_model format there is another issue: I can use it for inference without any problem but it cannot be saved after as an ML.NET standard zip repository. At the saving time all the pipe is correctly written in the zip but the saved_model file is missing...