[Torch] Pool ops, convert strides and pool_size to int

[alexwong]

Was working on a SVDD model and ran into some issues with pool operators.

  File "tests/python/frontend/pytorch/test_svdd.py", line 39, in <module>
    mod, params = relay.frontend.from_pytorch(model, shape_dict)
  File "/home/ubuntu/tvm/tvm/python/tvm/relay/frontend/pytorch.py", line 3160, in from_pytorch
    ret = converter.convert_operators(_get_operator_nodes(graph.nodes()), outputs, ret_name)[0]
  File "/home/ubuntu/tvm/tvm/python/tvm/relay/frontend/pytorch.py", line 2582, in convert_operators
    inputs, _get_input_types(op_node, outputs, default_dtype=self.default_dtype)
  File "/home/ubuntu/tvm/tvm/python/tvm/relay/frontend/pytorch.py", line 1338, in avg_pool2d
    return func(data)
  File "/home/ubuntu/tvm/tvm/python/tvm/relay/frontend/pytorch.py", line 1332, in func
    count_include_pad=count_include_pad,
  File "/home/ubuntu/tvm/tvm/python/tvm/relay/op/nn/nn.py", line 1026, in avg_pool2d
    return _make.avg_pool2d(data, pool_size, strides, padding, layout, ceil_mode, count_include_pad)
  File "/home/ubuntu/tvm/tvm/python/tvm/_ffi/_ctypes/packed_func.py", line 237, in __call__
    raise get_last_ffi_error()
tvm._ffi.base.TVMError: Traceback (most recent call last):
  [bt] (4) /home/ubuntu/tvm/tvm/build/libtvm.so(TVMFuncCall+0x5b) [0x7f6ec5b4f8db]
  [bt] (3) /home/ubuntu/tvm/tvm/build/libtvm.so(+0xe41916) [0x7f6ec564d916]
  [bt] (2) /home/ubuntu/tvm/tvm/build/libtvm.so(tvm::runtime::TVMMovableArgValueWithContext_::operator tvm::runtime::Array<tvm::PrimExpr, void><tvm::runtime::Array<tvm::PrimExpr, void> >() const+0x67) [0x7f6ec4df0217]
  [bt] (1) /home/ubuntu/tvm/tvm/build/libtvm.so(tvm::runtime::Array<tvm::PrimExpr, void> tvm::runtime::TVMPODValue_::AsObjectRef<tvm::runtime::Array<tvm::PrimExpr, void> >() const+0x438) [0x7f6ec4df00a8]
  [bt] (0) /home/ubuntu/tvm/tvm/build/libtvm.so(+0x5d5612) [0x7f6ec4de1612]
  File "/home/ubuntu/tvm/tvm/include/tvm/runtime/packed_func.h", line 687
TVMError: In function relay.op.nn._make.avg_pool2d: error while converting argument 2: [10:04:02] /home/ubuntu/tvm/tvm/include/tvm/runtime/packed_func.h:1564: 
---------------------------------------------------------------
An internal invariant was violated during the execution of TVM.
Please read TVM's error reporting guidelines.
More details can be found here: https://discuss.tvm.ai/t/error-reporting/7793.
---------------------------------------------------------------
  Check failed: !checked_type.defined() == false: Expected Array[PrimExpr], but got Array[index 0: relay.Constant]

I found that if the PT graph has aten::Int as the input value for either strides or pool_size, then we get above.

 %125 : int = aten::Int(%124)
 %126 : int[] = prim::ListConstruct(%116, %119)
 %127 : int[] = prim::ListConstruct(%122, %125)
 %128 : int[] = prim::ListConstruct(%111, %111)
 %x.6 : Tensor = aten::avg_pool2d(%x.5, %126, %127, %128, %108, %107, %106)

This fixes that case by converting relay constants to ints. @masahi

[masahi]

@alexwong can you add a test?

[alexwong]

@alexwong can you add a test?

Yep, added a test in maxpool2d. It creates a call expr which hits the new function.

[masahi]

Thanks @alexwong