[RELAY][FRONTEND][OP] support repeat in ONNX

[XiaolongMeng]

since it does not support tile op in onnx ,I write a convert, but it seems something wrong.

what I want to convert is
query = simple_cat.repeat(1, 100, 1)
which in pytorch.

what I add to onnx.py is

class Repeat(OnnxOpConverter):
    """ Operator converter for Repeat.
    """
    @classmethod
    def _impl_v1(cls, inputs, attr, params):
        repeats = tuple(params[inputs[1].name_hint].asnumpy())
        for i, repeat in enumerate(repeats):
            if repeat != 1:
                inputs[0] = _op.repeat(inputs[0], repeats=repeat, axis=i)
        return inputs[0]

and add this to _get_convert_map

but when I run
tvm_output = intrp.evaluate(sym)(inp, **params).asnumpy()

It occurred errors.

%4 = repeat(%3, meta[relay.attrs.RepeatAttrs][0]) # an internal invariant was violdated while typechecking your program [08:49:22] /notebooks/notebook/serving/tvm/include/tvm/expr_operator.h:535: cannot make const for type 80

Stack trace returned 10 entries:
[bt] (0) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0x73fc4d) [0x7f6a81f8bc4d]
[bt] (1) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0x74089d) [0x7f6a81f8c89d]
[bt] (2) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0x77d62d) [0x7f6a81fc962d]
[bt] (3) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0x77d716) [0x7f6a81fc9716]
[bt] (4) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0xa6d3e9) [0x7f6a822b93e9]
[bt] (5) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0xa59f74) [0x7f6a822a5f74]
[bt] (6) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0xc2b087) [0x7f6a82477087]
[bt] (7) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0xc45373) [0x7f6a82491373]
[bt] (8) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(tvm::relay::InferType(tvm::relay::Function const&, tvm::relay::Module const&, tvm::relay::GlobalVar const&)+0x32b) [0x7f6a824920db]
[bt] (9) /root/.local/lib/python3.6/site-packages/tvm-0.6.dev0-py3.6-linux-x86_64.egg/tvm/libtvm.so(+0xbb6720) [0x7f6a82402720]

so how can I fix this,thank you!

[XiaolongMeng]

emmm, I find what's wrong...It's the input shape, it must match the input name, I use 1,2,3, but in onnx,it's input.1, input.2..... so,maybe I can post a PR? @tqchen

[cchung100m]

Hi @XiaolongMeng

I would appreciate if you can help to review the proposed PR.

many thank,

[XiaolongMeng]

maybe you can call tqchen ...

[tqchen]

#3941 thanks to @cchung100m