Append dim even both dim value and param are not set

[xuzijian629]

Description

• In data propagation, fix appendDimToTensorShapeProto so that a dimension is appended even when both dim value and param are not known.

Motivation and Context

• I got a ShapeInferenceError caused by here, where the input_data is not nullptr but its dims have neither dim_values nor dim_params.

[askhade]

Curious - Are you using ONNX's graph level shape inference? meaning are you using onnx shape inference apis? When using onnx graph level shape inference I don't know how you could land in this situation this because - either the input data is an initializer in which case all dimensions are dim values or the input data is dynamic in which case graph level shape inference will create a symbol for every unknown dim... (https://github.com/onnx/onnx/blob/master/onnx/shape_inference/implementation.cc#L141)

However, not all runtimes implement symbol generation and therefore it may be best to include your fix.

[askhade]

can you simply add a dim? tsp.mutable_dim()->Add() instead of Copy

[askhade]

Added a comment

[askhade]

@xuzijian629 : The PR looks good to me. I added a minor comment. As soon as you address it we can merge this. Thanks for your contribution.

[askhade]

@xuzijian629: Any updates?

[askhade]

Any updates?

[xuzijian629]

Sorry, I missed the notification. I will soon make an update 🙇

[xuzijian629]

#3828 (comment)

Yes, I encountered this problem when I was using ONNX's graph level shape inference.

I got the bug in the following part of a model.

Since I cropped the graph, the inputs of Gather and Sub are hidden. Note that the inputs to Unsqueezes are all 0-dimensional (i.e., scalar).

[xuzijian629]

I summarized the above explanation into code.

Note that before running this, you may need to merge #3784 because the following code uses data propagation for ConstantOfShape.

import onnx
from onnx import parser, helper, shape_inference

input = '''
   agraph (float[3, n] x) => ()
   {
        shape = Shape(x)
        one = Constant<value = int64 {1}>()
        scalar = Gather(shape, one)
        subbed = Sub(scalar, one)
        unsqueeze_dim = Constant<value = int64[1] {0}>()
        unsqueezed = Unsqueeze(subbed, unsqueeze_dim)
        concat_out = Concat<axis = 0>(shape, unsqueezed)
        out = ConstantOfShape(concat_out)
   }
'''
graph = parser.parse_graph(input)
original_model = helper.make_model(graph, producer_name='onnx-examples')
onnx.checker.check_model(original_model)

# Apply shape inference on the model
inferred_model = shape_inference.infer_shapes(original_model, data_prop=True)
onnx.checker.check_model(inferred_model)
print(inferred_model.graph.value_info)

With this PR, the output shape of out is [3, n, unk__0], where as the existing ONNX infers as [3, n], which has incorrect rank.

[xuzijian629]

Please take a look! @askhade 🙇

[gramalingam]

Thanks for fixing this! This looks correct. But I think the whole function can be simplified to

   * tsp.add_dim() = dim;

[xuzijian629]

Thank you!