broadcast_like requires inputs to have same dtype

[leezu]

Description

broadcast_like "Broadcasts lhs to have the same shape as rhs." Thus only the shape information of lhs and rhs array should matter. However, it also asserts that the dtypes are equal.

Error Message

[08:58:11] /home/ANT.AMAZON.COM/lausen/src/mxnet-master/src/io/../operator/elemwise_op_common.h:135: Check failed: assign(&dattr, vec.at(i)): Incompatible attr in node  at 1-th input: expected float32, got int64
Stack trace:
  [bt] (0) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x43) [0x7f8b58c46d63]
  [bt] (1) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(mxnet::op::ElemwiseAttr<int, &mxnet::op::type_is_none, &mxnet::op::type_assign, true, &mxnet::op::type_string[abi:cxx11], -1, -1>(nnvm::NodeAttrs const&, std::vector<int, std::allocator<int> >*, std::vector<int, std::allocator<int> >*, int const&)::{lambda(std::vector<int, std::allocator<int> > const&, unsigned long, char const*)#1}::operator()(std::vector<int, std::allocator<int> > const&, unsigned long, char const*) const+0x275) [0x7f8b58e26e75]
  [bt] (2) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(bool mxnet::op::ElemwiseAttr<int, &mxnet::op::type_is_none, &mxnet::op::type_assign, true, &mxnet::op::type_string[abi:cxx11], -1, -1>(nnvm::NodeAttrs const&, std::vector<int, std::allocator<int> >*, std::vector<int, std::allocator<int> >*, int const&)+0x1b3) [0x7f8b58fd7783]
  [bt] (3) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(bool mxnet::op::ElemwiseType<2, 1>(nnvm::NodeAttrs const&, std::vector<int, std::allocator<int> >*, std::vector<int, std::allocator<int> >*)+0x23f) [0x7f8b59021c6f]
  [bt] (4) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(mxnet::imperative::SetShapeType(mxnet::Context const&, nnvm::NodeAttrs const&, std::vector<mxnet::NDArray*, std::allocator<mxnet::NDArray*> > const&, std::vector<mxnet::NDArray*, std::allocator<mxnet::NDArray*> > const&, mxnet::DispatchMode*)+0x57e) [0x7f8b58df513e]
  [bt] (5) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(mxnet::Imperative::Invoke(mxnet::Context const&, nnvm::NodeAttrs const&, std::vector<mxnet::NDArray*, std::allocator<mxnet::NDArray*> > const&, std::vector<mxnet::NDArray*, std::allocator<mxnet::NDArray*> > const&)+0xeb) [0x7f8b58de6a4b]
  [bt] (6) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(MXImperativeInvokeImpl(void*, int, void**, int*, void***, int, char const**, char const**)+0x538) [0x7f8b58c82388]
  [bt] (7) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/../../build/libmxnet.so(MXImperativeInvokeEx+0x8b) [0x7f8b58c833cb]
  [bt] (8) /home/ANT.AMAZON.COM/lausen/src/mxnet-master/python/mxnet/_cy3/ndarray.cpython-37m-x86_64-linux-gnu.so(+0x13be8) [0x7f8b2f1f0be8]

To Reproduce

mx.nd.broadcast_like(mx.nd.ones((1,)), mx.nd.arange(10, dtype=int))

[leezu]

mx.np.broadcast_arrays(mx.np.ones((1,)), mx.np.arange(10, dtype=int)) works correctly.

[leezu]

The reason that this bug is important: When writing a HybridBlock, the dtype of some input array depends on the user. But some other internally used array may have a fixed dtype that could be unequal of the input dtype. Workaround is to add an extra cast prior to broadcast_like.

[samskalicky]

Since this issue has 2 parts:

• Should not assert, should throw an error/exception instead
• broadcast_like should be enhanced [Feature request] to support different types Assigning to Mu for the first part, fixing error messages @zachgk assign [@mli ]

[szha]

this issue is related to an op that is deprecated in 2.0, so I'm marking this as 1.x. this issue is probably not a good first issue without pointers on where to look to address it.