[Relay][OP] Gather_nd exposed to relay

docs/langref/relay_op.rst

@@ -92,6 +92,7 @@ This level enables additional math and transform operators.
    tvm.relay.zeros_like
    tvm.relay.ones
    tvm.relay.ones_like
+   tvm.relay.gather_nd
    tvm.relay.full
    tvm.relay.full_like
    tvm.relay.cast
@@ -225,6 +226,7 @@ Level 3 Definitions
 .. autofunction:: tvm.relay.zeros_like
 .. autofunction:: tvm.relay.ones
 .. autofunction:: tvm.relay.ones_like
+.. autofunction:: tvm.relay.gather_nd
 .. autofunction:: tvm.relay.full
 .. autofunction:: tvm.relay.full_like
 .. autofunction:: tvm.relay.cast

python/tvm/relay/frontend/mxnet.py

@@ -1,4 +1,4 @@
-# pylint: disable=invalid-name, import-self, len-as-condition
+# pylint: disable=invalid-name, import-self, len-as-condition, unused-argument
 """MXNet symbol frontend."""
 from __future__ import absolute_import as _abs
 
@@ -631,6 +631,11 @@ def _mx_deformable_convolution(inputs, attrs):
     return res
 
 
+def _mx_gather_nd(inputs, attrs):
+    assert len(inputs) == 2
+    return _op.gather_nd(inputs[0], inputs[1])
+
+
 # Note: due to attribute conversion constraint
 # ops in the identity set must be attribute free
 _identity_list = [
@@ -768,6 +773,7 @@ def _mx_deformable_convolution(inputs, attrs):
     "SoftmaxOutput" : _mx_softmax_output,
     "SoftmaxActivation" : _mx_softmax_activation,
     "smooth_l1"     : _mx_smooth_l1,
+    "gather_nd"     : _mx_gather_nd,
     # vision
     "_contrib_BilinearResize2D" : _mx_upsampling,
     "_contrib_MultiBoxPrior" : _mx_multibox_prior,
@@ -782,7 +788,6 @@ def _mx_deformable_convolution(inputs, attrs):
     # TODO(tvm-tvm): support all operators.
     #
     # "broadcast_to",
-    # "gather_nd",
     # "Crop"          : _crop_like,
 }
 

python/tvm/relay/op/_transform.py

@@ -32,6 +32,7 @@
 _reg.register_schedule("stack", schedule_injective)
 _reg.register_schedule("concatenate", schedule_injective)
 _reg.register_schedule("_contrib_reverse_reshape", schedule_injective)
+_reg.register_schedule("gather_nd", schedule_injective)
 
 # layout_transform
 _reg.register_schedule("layout_transform", schedule_injective)

python/tvm/relay/op/transform.py

@@ -651,3 +651,36 @@ def reverse_reshape(data, newshape):
     if isinstance(newshape, int):
         newshape = [newshape]
     return _make._contrib_reverse_reshape(data, list(newshape))
+
+
+def gather_nd(data, indices):
+    """Gather elements or slices from data and store to a tensor whose shape is
+    defined by indices.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The input data to the operator.
+
+    indices : relay.Expr
+        The shape of output tensor.
+
+    Returns
+    -------
+    ret : relay.Expr
+        The computed result.
+
+    Examples
+    --------
+    .. code-block:: python
+
+        data = [[0, 1], [2, 3]]
+        indices = [[1, 1, 0], [0, 1, 0]]
+        relay.gather_nd(data, indices) = [2, 3, 0]
+
+        data = [[[1, 2], [3, 4]], [[5, 6], [7, 8]]]
+        indices = [[0, 1], [1, 0]]
+        relay.gather_nd(data, indices) = [[3, 4], [5, 6]]
+    """
+
+    return _make.gather_nd(data, indices)

src/relay/op/tensor/transform.cc

@@ -2122,5 +2122,75 @@ example below::
 .set_attr<FTVMCompute>("FTVMCompute", ReshapeCompute)
 .set_attr<TOpPattern>("TOpPattern", kInjective);
 
+// gather_nd operator
+bool GatherNDRel(const Array<Type>& types,
+                 int num_inputs,
+                 const Attrs& attrs,
+                 const TypeReporter& reporter) {
+  // `types` contains: [data, indices, result]
+  CHECK_EQ(types.size(), 3);
+  const auto* data = types[0].as<TensorTypeNode>();
+  const auto* indices = types[1].as<TensorTypeNode>();
+  if (data == nullptr) {
+    CHECK(types[0].as<IncompleteTypeNode>())
+        << "GatherND: expect input data type to be TensorType but get "
+        << types[0];
+    return false;
+  }
+  if (indices == nullptr) {
+    CHECK(types[1].as<IncompleteTypeNode>())
+        << "GatherND: expect indices type to be TensorType but get "
+        << types[1];
+    return false;
+  }
+  const size_t ndim = data->shape.size();
+  const IntImm* mdim = data->shape[0].as<IntImm>();
+  const size_t kdim = indices->shape.size() - 1;
+  CHECK(size_t(mdim->value) <= ndim)
+        << "GatherND: indices shape does satisfy.";
+
+  Array<IndexExpr> oshape;
+  for (size_t i = 1; i < kdim + 1; ++i)
+      oshape.push_back(indices->shape[i]);
+  for (size_t i = mdim->value; i < ndim; ++i)
+      oshape.push_back(data->shape[i]);
+  reporter->Assign(types[2], TensorTypeNode::make(oshape, data->dtype));
+  return true;
+}
+
+Array<Tensor> GatherNDCompute(const Attrs& attrs,
+                              const Array<Tensor>& inputs,
+                              const Type& out_type,
+                              const Target& target) {
+  return { topi::gather_nd(inputs[0], inputs[1]) };
+}
+
+Expr MakeGatherND(Expr data,
+                  Expr indices) {
+  static const Op& op = Op::Get("gather_nd");
+  return CallNode::make(op, {data, indices}, {});
+}
+
+TVM_REGISTER_API("relay.op._make.gather_nd")
+.set_body([](const TVMArgs& args, TVMRetValue* rv) {
+    runtime::detail::unpack_call<Expr, 2>(MakeGatherND, args, rv);
+});
+
+RELAY_REGISTER_OP("gather_nd")
+.describe(R"code(Gather elements or slices from data and store to
+                 a tensor whose shape is defined by indices.
+
+Given data with shape (X_0, X_1, ..., X_{N-1}) and indices with
+shape (M, Y_0, ..., Y_{K-1}), the output will have shape
+(Y_0, ..., Y_{K-1}, X_M, ..., X_{N-1}), where M <= N. If M == N,
+output shape will simply be (Y_0, ..., Y_{K-1}).
+)code" TVM_ADD_FILELINE)
+.set_num_inputs(2)
+.add_argument("data", "Tensor", "The input tensor.")
+.set_support_level(3)
+.add_type_rel("GatherND", GatherNDRel)
+.set_attr<FTVMCompute>("FTVMCompute", GatherNDCompute)
+.set_attr<TOpPattern>("TOpPattern", kInjective);
+
 }  // namespace relay
 }  // namespace tvm

tests/python/frontend/mxnet/test_forward.py

@@ -491,6 +491,21 @@ def verify(shape, indices_src, axis, mode="clip"):
     verify((3,4), [-1, 5], 1)
     verify((3,4), [-1, 5], 1, mode="wrap")
 
+def test_forward_gather_nd():
+    def verify(xshape, yshape, y_data):
+        x_data = np.random.uniform(size=xshape).astype("float32")
+        ref_res = mx.nd.gather_nd(mx.nd.array(x_data), mx.nd.array(y_data))
+        mx_sym = mx.sym.gather_nd(mx.sym.var("x_data"), mx.sym.var("y_data"))
+        new_sym, _ = relay.frontend.from_mxnet(mx_sym, {"x_data": xshape, "y_data": yshape}, {"x_data": "float32", "y_data": "int32"})
+        for target, ctx in ctx_list():
+            for kind in ["graph", "debug"]:
+                intrp = relay.create_executor(kind, ctx=ctx, target=target)
+                op_res = intrp.evaluate(new_sym)(x_data, y_data)
+                tvm.testing.assert_allclose(op_res.asnumpy(), ref_res.asnumpy())
+    verify((2, 2), (2, 3), [[1, 1, 0], [0, 1, 0]])
+    verify((2, 2, 2), (2, 2), [[0, 1], [1, 0]])
+
+
 if __name__ == '__main__':
     test_forward_mlp()
     test_forward_vgg()
@@ -527,3 +542,4 @@ def verify(shape, indices_src, axis, mode="clip"):
     test_forward_embedding()
     test_forward_smooth_l1()
     test_forward_take()
+    test_forward_gather_nd()

tests/python/relay/test_op_level3.py

@@ -553,7 +553,6 @@ def verify_stack(dshapes, axis):
     verify_stack([(2, 2, 3, 4), (2, 2, 3, 4), (2, 2, 3, 4), (2, 2, 3, 4)], -1)
 
 
-
 def test_reverse():
     def verify_reverse(dshape, axis):
         x = relay.var("x", relay.TensorType(dshape, "float32"))
@@ -573,6 +572,25 @@ def verify_reverse(dshape, axis):
     verify_reverse((2, 3, 4), -1)
 
 
+def test_gather_nd():
+    def verify_gather_nd(xshape, yshape, y_data):
+        x = relay.var("x", relay.TensorType(xshape, "float32"))
+        y = relay.var("y", relay.TensorType(yshape, "int32"))
+        z = relay.gather_nd(x, y)
+
+        func = relay.Function([x, y], z)
+        x_data = np.random.uniform(size=xshape).astype("float32")
+        ref_res = x_data[y_data]
+
+        for target, ctx in ctx_list():
+            for kind in ["graph", "debug"]:
+                intrp = relay.create_executor(kind, ctx=ctx, target=target)
+                op_res = intrp.evaluate(func)(x_data, y_data)
+                tvm.testing.assert_allclose(op_res.asnumpy(), ref_res, rtol=1e-5)
+    verify_gather_nd((2, 2), (2, 3), [[1, 1, 0], [0, 1, 0]])
+    verify_gather_nd((2, 2, 2), (2, 2), [[0, 1], [1, 0]])
+
+
 if __name__ == "__main__":
     test_cast()
     test_zeros_ones()
@@ -601,3 +619,4 @@ def verify_reverse(dshape, axis):
     test_stack()
     test_tile()
     test_repeat()
+    test_gather_nd()