the cond operator doesn't support a function that doesn't require inputs.

[zheng-da]

The python functions of the cond operator has to access some variables outside the function.

def test_no_inputs():
    class TestBlock(gluon.HybridBlock):
        def __init__(self, prefix=None, params=None):
            super(TestBlock, self).__init__(prefix=prefix, params=params)

        def hybrid_forward(self, F, data):
            def then_func():
                return F.ones((1)).expand_dims(0)
            def else_func():
                return data
            return F.contrib.cond(data == 0, then_func, else_func)

    block = TestBlock()
    block.initialize(ctx=default_context())
    block.hybridize()
    data = mx.nd.ones((1)).expand_dims(0)
    print(data)
    block(data)

The following code will run into an error as follows.

======================================================================
ERROR: __main__.test_no_inputs
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/usr/local/lib/python2.7/site-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "tests/python/unittest/test_contrib_control_flow.py", line 47, in test_no_inputs
    block(data)
  File "/Users/dzzhen/Workspace/incubator-mxnet/python/mxnet/gluon/block.py", line 541, in __call__
    out = self.forward(*args)
  File "/Users/dzzhen/Workspace/incubator-mxnet/python/mxnet/gluon/block.py", line 908, in forward
    return self._call_cached_op(x, *args)
  File "/Users/dzzhen/Workspace/incubator-mxnet/python/mxnet/gluon/block.py", line 814, in _call_cached_op
    out = self._cached_op(*cargs)
  File "/Users/dzzhen/Workspace/incubator-mxnet/python/mxnet/_ctypes/ndarray.py", line 150, in __call__
    ctypes.byref(out_stypes)))
  File "/Users/dzzhen/Workspace/incubator-mxnet/python/mxnet/base.py", line 255, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
MXNetError: [16:00:03] src/imperative/cached_op.cc:127: Check failed: idx.input_nodes().size() >= 1 (0 vs. 1) CachedOp requires at least 1 input

Stack trace returned 10 entries:
[bt] (0) 0   libmxnet.so                         0x00000001053cfe50 dmlc::StackTrace() + 272
[bt] (1) 1   libmxnet.so                         0x00000001053cfbff dmlc::LogMessageFatal::~LogMessageFatal() + 47
[bt] (2) 2   libmxnet.so                         0x000000010673209c mxnet::CachedOp::CachedOp(nnvm::Symbol const&, std::__1::vector<std::__1::pair<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> >, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > >, std::__1::allocator<std::__1::pair<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> >, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > > > > const&) + 3164
[bt] (3) 3   libmxnet.so                         0x00000001068d6865 mxnet::op::LoopState::MakeSharedOp(nnvm::Symbol const&) + 453
[bt] (4) 4   libmxnet.so                         0x0000000106cbdb74 mxnet::op::LoopState::LoopState(nnvm::Symbol const&) + 164
[bt] (5) 5   libmxnet.so                         0x00000001068da510 mxnet::op::CondState::CondState(mxnet::op::CondParam const&, nnvm::Symbol const&, nnvm::Symbol const&, nnvm::Symbol const&) + 64
[bt] (6) 6   libmxnet.so                         0x00000001068da41b mxnet::OpStatePtr mxnet::OpStatePtr::Create<mxnet::op::CondState, mxnet::op::CondParam const&, nnvm::Symbol&, nnvm::Symbol&, nnvm::Symbol&>(mxnet::op::CondParam const&&&, nnvm::Symbol&&&, nnvm::Symbol&&&, nnvm::Symbol&&&) + 75
[bt] (7) 7   libmxnet.so                         0x00000001068d1731 mxnet::op::CreateCondState(nnvm::NodeAttrs const&, mxnet::Context, std::__1::vector<nnvm::TShape, std::__1::allocator<nnvm::TShape> > const&, std::__1::vector<int, std::__1::allocator<int> > const&) + 49
[bt] (8) 8   libmxnet.so                         0x0000000105550b4e std::__1::__function::__func<mxnet::OpStatePtr (*)(nnvm::NodeAttrs const&, mxnet::Context, std::__1::vector<nnvm::TShape, std::__1::allocator<nnvm::TShape> > const&, std::__1::vector<int, std::__1::allocator<int> > const&), std::__1::allocator<mxnet::OpStatePtr (*)(nnvm::NodeAttrs const&, mxnet::Context, std::__1::vector<nnvm::TShape, std::__1::allocator<nnvm::TShape> > const&, std::__1::vector<int, std::__1::allocator<int> > const&)>, mxnet::OpStatePtr (nnvm::NodeAttrs const&, mxnet::Context, std::__1::vector<nnvm::TShape, std::__1::allocator<nnvm::TShape> > const&, std::__1::vector<int, std::__1::allocator<int> > const&)>::operator()(nnvm::NodeAttrs const&, mxnet::Context&&, std::__1::vector<nnvm::TShape, std::__1::allocator<nnvm::TShape> > const&, std::__1::vector<int, std::__1::allocator<int> > const&) + 30
[bt] (9) 9   libmxnet.so                         0x000000010676ca6e mxnet::imperative::RunGraph(bool, nnvm::IndexedGraph const&, std::__1::vector<mxnet::NDArray*, std::__1::allocator<mxnet::NDArray*> >, unsigned long, unsigned long, std::__1::vector<mxnet::OpReqType, std::__1::allocator<mxnet::OpReqType> >&&, std::__1::vector<unsigned int, std::__1::allocator<unsigned int> >&&, std::__1::vector<mxnet::OpStatePtr, std::__1::allocator<mxnet::OpStatePtr> >*, std::__1::vector<mxnet::DispatchMode, std::__1::allocator<mxnet::DispatchMode> > const&, bool) + 2462

[zheng-da]

The reason that the cond operator fails is that CachedOp, which the cond operator uses to execute a subgraph, requires to have inputs for every subgraph.

[zheng-da]

@junrushao1994

[vdantu]

@mxnet-label-bot [Python, Question]

[junrushao]

@zheng-da Thank you so much for providing this example and very detailed explanation!

To be clear, please allow me to summarize the problem as follows: currently, the implementation of three control flow operators, foreach, cond and while_loop, requires every subgraph to have at least one input.

To my understanding, this is actually a restriction of CachedOp: our implementation simply creates a CachedOp for each subgraph, but CachedOp refuses to create because it requires at least one input. This is reflected in the stack trace information you provide:

[bt] (2) 2   libmxnet.so                         0x000000010673209c mxnet::CachedOp::CachedOp(nnvm::Symbol const&, std::__1::vector<std::__1::pair<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> >, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > >, std::__1::allocator<std::__1::pair<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> >, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > > > > const&) + 3164
[bt] (3) 3   libmxnet.so                         0x00000001068d6865 mxnet::op::LoopState::MakeSharedOp(nnvm::Symbol const&) + 453
[bt] (4) 4   libmxnet.so                         0x0000000106cbdb74 mxnet::op::LoopState::LoopState(nnvm::Symbol const&) + 164

I guess it will be great if it could be fixed in ongoing work #11641.

[zheng-da]

Yes, the problem is that CachedOp can't handle a subgraph that doesn't require inputs. There are two possible fixes: one is to find nodes such as zeros() and ones() and keep them outside the subgraph; the other is to modify CachedOp to handle the case. The problem of handling subgraph without inputs in CachedOp is that we need to find out what context we should use.