[relay]where op error:can_dispatch(n):NodeFunctor calls un-registered function on type Any

Questions
#1

I use relay.frontend.from_tensorflow,the following error occurred:

File "tvm/python/tvm/relay/frontend/tensorflow.py", in from_tensorflow
         mod, params = g.from_tensorflow(graph, layout, shape, outputs)
File "tvm/python/tvm/relay/frontend/tensorflow.py", in from_tensorflow
         for node_item in self._nodes[node.name]]
File "tvm/python/tvm/relay/frontend/common.py", in infer_shape
         return get_const_tuple(out_type.checked_type.shape)
File "tvm/topi/python/topi/util.py", in infer_shape,in get_const_tuple
        elem = tvm.tir.ir_pass.Simplify(elem)
xxxxx(etc.)
TVMError: Check failed:error:can_dispatch(n):NodeFunctor calls un-registered function on type Any

TVM: latest version
Tensorflow:1.14
Model: bert

I print the other information: OP : Where ( len(inputs) == 1, so it’s argwhere )

[
    CallNode(
        Op(argwhere),
        [CallNode(
                 Op(not_equal),
                  [CallNode(
                         Op(full),
                         [Constant(1),],
                         relay.attrs.InitOpAttrs(0x98720b8),
                         []),
                    CallNode(
                         Op(zeros_like),
                         [CallNode(Op(full),[Constant(1),],relay.attrs.InitOpAttrs(0x98720b8),[])]
                         relay.attrs.InitOpAttrs(0x98720b8),
                         []),
                  ],
                  (nullptr),[]
        )],
        (nullptr),
        []
    )
]

input mod info:

%0 = full(1,shape=[128],dtype="int64");
free_var % zeros_like: Tensor[(128),int 64]
not_equal(%0, % zeros_like)

input shape:

[128]

I add print in infer_shape function from “tvm/python/tvm/relay/frontend/common.py”:

error out_type.checked_type = Tensor[(?, 1), int32]
out_type.checked_type.shape: [?, 1]

The output shape of argwhere op should be dynamic,it depends on the actual calculation.

So,Is there a problem with the ‘where’ implementation of relay ?

1 Like
#2

I think that’s because of the function TransformShape in the file data_layout.cc. It’s used to split shape like from NCHW to NCHWc. But when you want to convert your output shape from NCHW to NHWC for example, it will cause error. So you need to add your own shape transform function.

1 Like
#3

Can you explain more about this? I have same issue

#4

May I ask how to fix this problem? I have same issue

#5

It seems there is a bug in original TransformShape function. What I did is to add a transposeshape function, just tranpose its shape by using name hint of original axis. That means which split the original function with two parts. one is to split, and another one is transpose if I remenber correctly