DepthwiseConv2d translation adds Reshape and Transpose operators to the kernel that are invalid

[NikolasMarkou]

Describe the bug
When translating the DepthwiseConv2d or DepthwiseConv2dNative operations 2 extra ops are added to the kernel to get into the correct form. These are the Reshape and Transpose operators. These are problematic for the following reasons.

1. The Transpose order (3, 2, 0, 1) is invalid under certain versions of tensorrt
2. The Reshape operator is invalid under certain versions of tensorrt (5.1 does not support the int64)
3. Additional complexity with 2 extra layers per DepthwiseConv2d operator

Urgency
This is urgent and blocker for anybody using the onnx system as a intermediate step towards other runtimes

System information

• Linux Mint 18.04
• Tensorflow Version: 1.15
• Python version: 3.6.9

To Reproduce
Create any DepthwiseConv2dNative or DepthwiseConv2d operator in tensorflow. I use Opset 7 but this is true for any version

Expected behavior
I expected the kernel to be changed/reshaped internally and not use additional operators

Screenshots
Tensorflow

Onnx

Additional context
My target system is tensorrt 5.1 and tensorrt 7.0

[guschmue]

This is via the tf2onnx command line or via python api ?
I'm sure our optimizer would rewrite a sequence transpose->clip->transpose to clip.
Maybe this was via python api and the optimizer wasn't called (some sample how to call it is here).

Since onnx only supports NCHW we need insert those transposes. The way we do this is that we wrap the impacted ops with a transpose and the optimizer later tries to push them down to the output of the graph until they hit another transpose so we can remove the 2 transposes. That works for most models. Sometimes there is a model where some transposes will be left either because it is just not possible to remove them or the optimizer doesn't handle a specific situation. Later we should be able to fix.

[NikolasMarkou]

Hi @guschmue this was performed using the tf2onnx command line tool, the reason the transpose -> clip -> transpose sequence is not optimized is because of the other Issue / PR i opened (the optimizer failed due to other reasons)