Onnxruntime: SliceElimination transformer causes shape inference error

[lucienwang1009]

Describe the bug

I'd like to test some object detection model on the latest runtime when it runs into a shape inference exception with the message:
RuntimeError: [ONNXRuntimeError] : 1 : GENERAL ERROR : [TypeInferenceError] Graph attribute inferencing failed: [ShapeInferenceError] Incompatible dimensions.
If I change back to runtime 0.3.0, everything will be fine.

You can find the model here: model. It's a part of Faster-RCNN model.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 16.04
• ONNX Runtime installed from (source or binary): source or from https://test.pypi.org/project/ort-nightly/
• ONNX Runtime version: 0.3.1
• Python version: 3.6
• Visual Studio version (if applicable):
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version:
• GPU model and memory:

To Reproduce
Describe steps/code to reproduce the behavior:

Expected behavior
A clear and concise description of what you expected to happen.

Screenshots
If applicable, add screenshots to help explain your problem.

Additional context
Add any other context about the problem here.

[snnn]

For any shape inference error, it's either a converter bug, or an ONNX bug. It's unlikely caused by onnxruntime code (unless the TypeInferenceError was thrown from a custom op node)

Do you know which op caused this error?

[RyanUnderhill]

The issue doesn't look like it's an Onnx issue. The inference error was in a 'Div' op (which is confusing since there is no 'Div' in the graph). When I disable the graph transform 'EliminateSlice' it works without error.

Adding Ashwini since she knows about transforms and helped me to narrow it down.

[askhade]

Konstantinos is already looking into this

[askhade]

@lucienwang1009 you can turn off the transformers by setting the level to 0. Which API are you using?
CAPI: OrtSetSessionGraphOptimizationLevel(session_options, 0);
c# api: options.SetSessionGraphOptimizationLevel(0);
python: options.set_graph_optimization_level(0)

[kkaranasos]

From an initial investigation, Slice Elimination is indeed causing the problem. As @askhade says, you can bypass this by setting the optimization level to 0.
Interestingly, the problem is not caused by the "visible" Slice in the top-level graph (this cannot be eliminated actually), but by a Slice that appears in a 74-node sub-graph (if I disable elimination in the subgraphs, it works). I will debug it later today.

[kkaranasos]

PR https://github.com/Microsoft/onnxruntime/pull/918 fixes the issue.

[kkaranasos]

Issue fixed with #918.