[translate ONNX to Relax]AttributeError: <class 'tvm.relax.expr.Call'> has no attribute data

#1

When I perform a transformation on an ONNX node, I encounter the following error. I checked the input of this Mul operation, and it looks like this.

image
image1225×381 26.6 KB
According to the code, the input here needs to be of type relax.PrimValue, but the actual input is indeed tvm.relax.expr.Call. I would like to ask how to resolve this error. Thank you very much.

the error is:\

Traceback (most recent call last):
  File "/lpai/code/llava_tvm/debug.py", line 39, in <module>
    relax_mod = from_onnx(onnx_model, shape_dict=encoder_shape_dyna, dtype_dict=dtype_dict, keep_params_in_input=True)
  File "/lpai/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 3703, in from_onnx
    return g.from_onnx(graph, opset)
  File "/lpai/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 3320, in from_onnx
    self._construct_nodes(graph)
  File "/lpai/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 3504, in _construct_nodes
    op = self._convert_operator(op_name, inputs, attr, self.opset)
  File "/lpai/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 3609, in _convert_operator
    sym = op_function(self.bb, inputs, attrs, [self._nodes, self._params])
  File "/lpai/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 381, in _impl_v7
    return cls.base_impl(bb, inputs, attr, params)
  File "/lpai/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 344, in base_impl
    x = _to_numpy(inputs[0])
  File "/lpai/tvm/python/tvm/relax/frontend/onnx/onnx_frontend.py", line 324, in _to_numpy
    return x.data.numpy()
  File "/lpai/tvm/python/tvm/runtime/object.py", line 75, in __getattr__
    raise AttributeError(f"{type(self)} has no attribute {name}") from None
AttributeError: <class 'tvm.relax.expr.Call'> has no attribute data
#2

I’ve encountered this problem as well, and I believe the main issue is that the TVM frontend doesn’t support dynamic shapes for the ReduceSum operator in ONNX.

#3

Thank you for your response. I hadn’t noticed this before, but I think you have a point. I attempted to forcibly modify the logic with the following code to make it work here, but many issues arose afterward. I think I should take a different approach to handle this ONNX model. :joy: class BinaryBase(OnnxOpConverter): “”“Converts an onnx BinaryBase node into an equivalent Relax expression.”""

    numpy_op: Callable = None
    relax_op: Callable = None

    @classmethod
    def base_impl(cls, bb, inputs, attr, params):
        """Base implementation for binary operations."""
        if cls.numpy_op is None or cls.relax_op is None:
            raise ValueError("Numpy and Relax operators must be defined for BinaryBase.")
        if (len(inputs) == 2 and isinstance(inputs[0], relax.expr.Call) and isinstance(inputs[1], relax.PrimValue)):
            # import pdb; pdb.set_trace()
            x = inputs[0]
            y = inputs[1]
            if isinstance(y, relax.expr.Var):
                return cls.relax_op(x, y)
            if isinstance(y.value, (tir.IntImm, tir.FloatImm)):
                y = relax.const(y.value.value, dtype=x.struct_info.dtype)
                return cls.relax_op(x, y)

            return cls.relax_op(x, y)

        if all([isinstance(inp, relax.Constant) for inp in inputs]):
            output = cls.numpy_op(  # pylint: disable=not-callable
                inputs[0].data.numpy(), inputs[1].data.numpy()
            )
            return relax.const(output, inputs[0].struct_info.dtype)
        if any([isinstance(inp, relax.PrimValue) for inp in inputs]):
            x = _to_numpy(inputs[0])
            y = _to_numpy(inputs[1])
            return relax.PrimValue(cls.numpy_op(x, y))  # pylint: disable=not-callable

        return cls.relax_op(inputs[0], inputs[1])  # pylint: disable=not-callable