Missing output shape when using infer_shapes_path on a model with external data?

[daveliddell]

Hi! I'm having trouble converting code that previously used infer_shapes to use infer_shapes_path instead because the model I'm working with is too large for infer_shapes. However, on a small test model, it seems that now the output shape isn't getting set. I'm using onnx 1.15 with python 3.10 on ubuntu 22 or 20 (doesn't seem to matter).

Here's the model (with data externalized):

def create_model(model_file: Path, data_path: Path):
    const = make_node('Constant', [], ['c_shape'], 'const',
        value=numpy_helper.from_array(numpy.array([4], dtype=numpy.int64)))
    cofshape = make_node('ConstantOfShape', ['c_shape'], ['c_out'], 'cofshape',
        value=numpy_helper.from_array(numpy.array([1], dtype=numpy.int64)))

    outval = make_tensor_value_info('c_out', TensorProto.INT64, [None])
    graph = make_graph([const, cofshape], 'constgraph', [], [outval])

    onnx_model = make_model(graph)
    convert_model_to_external_data(onnx_model, all_tensors_to_one_file=True,
        location=data_path, size_threshold=0, convert_attribute=True)
    onnx.save(onnx_model, model_file)

Here's the code to load it and infer shapes:

    onnx.shape_inference.infer_shapes_path(args.input_file, temp_file)
    inferred_model = onnx.load(temp_file, load_external_data=False)
    onnx.load_external_data_for_model(inferred_model, data_dir)

Here's a dump of the output of the resulting model:

  output {
    name: "c_out"
    type {
      tensor_type {
        elem_type: 7
        shape {
          dim {   <-- was expecting "dim_value: 4" here
          }
        }
      }
    }
  }

Some experiments I've tried:

• infer_shapes_path + internal data --> PASS
• infer_shapes_path + external data --> FAIL
• infer_shapes + external data --> PASS

Here's how we were using infer_shapes before:

    raw_model = onnx.load(file_path)
    inferred_model = onnx.shape_inference.infer_shapes(raw_model)

Before I start debugging, I was hoping someone might at least point me in the right direction. Thanks!