Support negative inputs for 'Pad' op

[fthielke]

Describe the bug

If I have a Pad op with negative paddings (in this case -16 in all the spatial dimensions of a 3D volume) in my model, onnx_tf.backend.prepare raises the following error:

ValueError: Paddings must be non-negative for 'Pad' (op: 'Pad') with input shapes: [1,32,40,40,40], [5,2] and with computed input tensors: input[1] = <[0 0][0 0][-16 -16][-16 -16][-16 -16]>.

However, this is not in line with the ONNX spec which supports negative paddings (ie. croppings) since version 2: https://github.com/onnx/onnx/blob/master/docs/Changelog.md#Pad-2

Python, ONNX, ONNX-TF, Tensorflow version

• Python version: 3.7.6 (default, Jan 8 2020, 20:23:39) [MSC v.1916 64 bit (AMD64)]
• ONNX version: 1.5.0
• ONNX-TF version: 1.5.0
• Tensorflow version: 2.1.0

[chinhuang007]

Thanks for opening the issue. It is definitely a bug. We will look into a solution. Please note the converted graph will be much bigger since Tensorflow pad operator does not support negative pads.