Describe the bug
Pad operator: Negative pads value not supported (shape is correct values look like garbage memory).

Urgency
Pad node with a negative pad amount are usefull to perform crop. As negative pad is relative to the size of the input while, a slice operation is hard-coded, meaning a network cannot change input resolution later. ONNX specs support negative pad amount, having ONNX runtime support it as well is a must have for us.

System information

• OS: Windows 10
• ONNX Runtime installed from (source or binary): From binary (pip)
• ONNX Runtime version: tryed both 0.5 and 1.0 (ie same problem with both)
• ONNX version : 1.2.2
• Python version: Python 3.6.8
• Visual Studio version (if applicable): NA
• GCC/Compiler version (if compiling from source): NA
• CUDA/cuDNN version: NA
• GPU model and memory: NA

To Reproduce
Run the attached python script negative_pad_onnxruntime_repro_bug.py.txt
Look at output tensor values.

Expected behavior
According to ONNX Negative Padding should be supported:
https://github.com/onnx/onnx/blob/master/docs/Operators.md#Pad

Running the script attached shows that it is not the case, ie the output tensor shape is correct while values look like uninitialized memory.

Screenshots

Additional context