Can't concat ragged tensors with more than one dimension.

[magicaltoast]

Describe the bug Can't concat ragged tensors with more than one dimension.

I want to decode results from the object detector. I am using tf.ragged.boolean_mask(bboxes, pos_inds) to gather results above threshold but my detector has multiple scales like yolo family so I need to concatenate it to one tensor to perform nonmax suppression. There's is no problem with scores because is a 1d tensor but it fails to convert bounding boxes.

That feature would allow avoiding builder plate code for decoding object detection results because the decoding step would be inside onnx model.

System information

• OS Platform and Distribution: Ubuntu 21.04
• Tensorflow Version: 2.7.0
• Python version: 3.8.8

To Reproduce

i1 = Input((None, 4), ragged=True, batch_size=1)
i2 = Input((None, 4), ragged=True, batch_size=1)
# None in shape above means that i have unknow number of detection per scale, 4 is for bounding box
o = tf.concat([i1, i2], axis=1)
m = Model([i1, i2], o)
m.save("model")

python -m tf2onnx.convert --saved-model model --output model.onnx --opset 15

Error

ValueError: make_sure failure: RaggedGather conversion only supports tensors with no dense dimensions

[hwangdeyu]

Now the ONNX doesn't support ragged tensors. Converter can convert rank-1 ragged tensors but currently can't do larger than rank-1. We could potentially support higher ranks with a loop, but that would probably not be very efficient now.