implement round ste

Author: Jopyth

python/mxnet/ndarray/ndarray.py

@@ -1406,6 +1406,14 @@ def round(self, *args, **kwargs):
         """
         return op.round(self, *args, **kwargs)
 
+    def round_ste(self, *args, **kwargs):
+        """Convenience fluent method for :py:func:`round_ste`.
+
+        The arguments are the same as for :py:func:`round_ste`, with
+        this array as data.
+        """
+        return op.round_ste(self, *args, **kwargs)
+
     def rint(self, *args, **kwargs):
         """Convenience fluent method for :py:func:`rint`.
 

python/mxnet/symbol/symbol.py

@@ -2150,6 +2150,14 @@ def round(self, *args, **kwargs):
         """
         return op.round(self, *args, **kwargs)
 
+    def round_ste(self, *args, **kwargs):
+        """Convenience fluent method for :py:func:`round_ste`.
+
+        The arguments are the same as for :py:func:`round_ste`, with
+        this array as data.
+        """
+        return op.round_ste(self, *args, **kwargs)
+
     def rint(self, *args, **kwargs):
         """Convenience fluent method for :py:func:`rint`.
 

smd_hpi/tests/test_functions.py

@@ -15,3 +15,18 @@ def test_det_sign():
         assert_almost_equal(exp_y, y.asnumpy())
     y.backward()
     assert_almost_equal(exp_grad, x.grad.asnumpy())
+
+
+def test_round_ste():
+    npy = np.random.uniform(-10, 10, (2, 3, 4))
+
+    exp_y = np.round(npy)
+    exp_grad = np.ones_like(npy)
+
+    x = mx.nd.array(npy)
+    x.attach_grad()
+    with autograd.record():
+        y = x.round_ste()
+        assert_almost_equal(exp_y, y.asnumpy())
+    y.backward()
+    assert_almost_equal(exp_grad, x.grad.asnumpy())

src/operator/operator_tune.cc

@@ -273,6 +273,7 @@ IMPLEMENT_UNARY_WORKLOAD_BWD(mxnet::op::mshadow_op::sign_grad);  // NOLINT()
 IMPLEMENT_UNARY_WORKLOAD_FWD(mxnet::op::mshadow_op::det_sign);  // NOLINT()
 IMPLEMENT_UNARY_WORKLOAD_BWD(mxnet::op::mshadow_op::det_sign);  // NOLINT()
 IMPLEMENT_UNARY_WORKLOAD_FWD(mxnet::op::mshadow_op::round);  // NOLINT()
+IMPLEMENT_UNARY_WORKLOAD_BWD(mxnet::op::mshadow_op::round);  // NOLINT()
 IMPLEMENT_UNARY_WORKLOAD_FWD(mxnet::op::mshadow_op::floor);  // NOLINT()
 IMPLEMENT_UNARY_WORKLOAD_FWD(mxnet::op::mshadow_op::trunc);  // NOLINT()
 IMPLEMENT_UNARY_WORKLOAD_FWD(mxnet::op::mshadow_op::rint);  // NOLINT()

src/operator/tensor/elemwise_unary_op_basic.cc

@@ -723,6 +723,25 @@ The storage type of ``round`` output depends upon the input storage type:
 )code" ADD_FILELINE)
 .set_attr<nnvm::FGradient>("FGradient", MakeZeroGradNodes);
 
+// round_ste
+MXNET_OPERATOR_REGISTER_UNARY_WITH_RSP(round_ste, cpu, mshadow_op::round)
+MXNET_ADD_SPARSE_OP_ALIAS(round_ste)
+.describe(R"code(Returns element-wise rounded value to the nearest integer of the input but with STE.
+
+Example::
+
+   round_ste([-1.5, 1.5, -1.9, 1.9, 2.1]) = [-2.,  2., -2.,  2.,  2.]
+
+The storage type of ``round_ste`` output depends upon the input storage type:
+
+  - round_ste(default) = default
+  - round_ste(row_sparse) = row_sparse
+
+)code" ADD_FILELINE)
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_round_ste"});
+
+MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU(_backward_round_ste, unary_bwd<mshadow_op::identity_grad>);
+
 // rint
 MXNET_OPERATOR_REGISTER_UNARY_WITH_RSP_CSR(rint, cpu, mshadow_op::rint)
 .describe(R"code(Returns element-wise rounded value to the nearest integer of the input.

src/operator/tensor/elemwise_unary_op_basic.cu

@@ -173,6 +173,15 @@ NNVM_REGISTER_OP(round)
 .set_attr<FCompute>("FCompute<gpu>", UnaryOp::Compute<gpu, mshadow_op::round>)
 .set_attr<FComputeEx>("FComputeEx<gpu>", UnaryOp::ComputeEx<gpu, mshadow_op::round>);
 
+// round_ste
+NNVM_REGISTER_OP(round_ste)
+.set_attr<FCompute>("FCompute<gpu>", UnaryOp::Compute<gpu, mshadow_op::round>)
+.set_attr<FComputeEx>("FComputeEx<gpu>", UnaryOp::ComputeEx<gpu, mshadow_op::round>);
+
+NNVM_REGISTER_OP(_backward_round_ste)
+.set_attr<FCompute>("FCompute<gpu>", ElemwiseBinaryOp::Compute<
+  gpu, unary_bwd<mshadow_op::identity_grad> >);
+
 // ceil
 NNVM_REGISTER_OP(ceil)
 .set_attr<FCompute>("FCompute<gpu>", UnaryOp::Compute<gpu, mshadow_op::ceil>)