Add OP diag [numpy] (#16786)

* Add numpy.diag method

* Fix ci diag problem

* Merge and fix ci Diag problem

* Fix format issue

* Modified Diag method

* Fixed the Diag kernel and name

* Fixed sanity problem

* change K param

* change unittest and fix

* Fix diag code format

* update 3rd party

python/mxnet/_numpy_op_doc.py

@@ -1087,3 +1087,38 @@ def _npx_reshape(a, newshape, reverse=False, order='C'):
     (8, 3, 2, 4, 4, 2)
     """
     pass
+
+
+def _np_diag(array, k = 0):
+    """
+    Extracts a diagonal or constructs a diagonal array.
+    - 1-D arrays: constructs a 2-D array with the input as its diagonal, all other elements are zero.
+    - 2-D arrays: extracts the k-th Diagonal
+
+    Parameters
+    ----------
+    array : ndarray
+        The array to apply diag method.
+    k : offset
+        extracts or constructs kth diagonal given input array
+
+    Examples
+    --------
+    >>> x = np.arange(9).reshape((3,3))
+    >>> x
+    array([[0, 1, 2],
+           [3, 4, 5],
+           [6, 7, 8]])
+    >>> np.diag(x)
+    array([0, 4, 8])
+    >>> np.diag(x, k=1)
+    array([1, 5])
+    >>> np.diag(x, k=-1)
+    array([3, 7])
+
+    >>> np.diag(np.diag(x))
+    array([[0, 0, 0],
+           [0, 4, 0],
+           [0, 0, 8]])
+    """
+    pass

python/mxnet/numpy_dispatch_protocol.py

@@ -93,6 +93,7 @@ def _run_with_array_ufunc_proto(*args, **kwargs):
     'concatenate',
     'copy',
     'cumsum',
+    'diag',
     'dot',
     'expand_dims',
     'fix',

src/operator/numpy/np_matrix_op-inl.h

@@ -28,9 +28,14 @@
 #include <vector>
 #include <algorithm>
 #include <string>
+#include <utility>
 #include "../tensor/matrix_op-inl.h"
 #include "../nn/concat-inl.h"
 #include "../../common/utils.h"
+#include "../mxnet_op.h"
+#include "../operator_common.h"
+#include "../elemwise_op_common.h"
+#include "../tensor/broadcast_reduce_op.h"
 
 namespace mxnet {
 namespace op {
@@ -945,6 +950,206 @@ void NumpyConcatenateBackward(const nnvm::NodeAttrs& attrs,
   });
 }
 
+struct NumpyDiagParam : public dmlc::Parameter<NumpyDiagParam> {
+  int k;
+  DMLC_DECLARE_PARAMETER(NumpyDiagParam) {
+    DMLC_DECLARE_FIELD(k).set_default(0)
+      .describe("Diagonal in question. The default is 0. "
+                "Use k>0 for diagonals above the main diagonal, "
+                "and k<0 for diagonals below the main diagonal. ");
+  }
+};
+
+inline mxnet::TShape NumpyDiagShapeImpl(const mxnet::TShape &ishape,
+                                        const int k) {
+  CHECK_LE(ishape.ndim(), 2) << "Input must be 1- or 2-d";
+
+  if (ishape.ndim() == 1) {
+    auto s = ishape[0] + std::abs(k);
+    return mxnet::TShape({s, s});
+  }
+
+  auto h = ishape[0];
+  auto w = ishape[1];
+
+  if (k > 0) {
+    w -= k;
+  } else if (k < 0) {
+    h += k;
+  }
+  dim_t a = 0;
+  auto s = std::max(std::min(h, w), a);
+  // s is the length of diagonal with k as the offset
+
+  int32_t n_dim = ishape.ndim() - 1;
+  mxnet::TShape oshape(n_dim, -1);
+  oshape[n_dim - 1] = s;
+  return oshape;
+}
+
+inline bool NumpyDiagOpShape(const nnvm::NodeAttrs &attrs,
+                             mxnet::ShapeVector *in_attrs,
+                             mxnet::ShapeVector *out_attrs) {
+  CHECK_EQ(in_attrs->size(), 1U);
+  CHECK_EQ(out_attrs->size(), 1U);
+
+  const mxnet::TShape &ishape = (*in_attrs)[0];
+  if (!mxnet::ndim_is_known(ishape)) {
+    return false;
+  }
+
+  const NumpyDiagParam &param = nnvm::get<NumpyDiagParam>(attrs.parsed);
+  mxnet::TShape oshape = NumpyDiagShapeImpl(ishape, param.k);
+
+  if (shape_is_none(oshape)) {
+    LOG(FATAL) << "Diagonal does not exist.";
+  }
+  SHAPE_ASSIGN_CHECK(*out_attrs, 0, oshape);
+
+  return shape_is_known(out_attrs->at(0));
+}
+
+inline bool NumpyDiagOpType(const nnvm::NodeAttrs &attrs,
+                            std::vector<int> *in_attrs,
+                            std::vector<int> *out_attrs) {
+  CHECK_EQ(in_attrs->size(), 1U);
+  CHECK_EQ(out_attrs->size(), 1U);
+
+  TYPE_ASSIGN_CHECK(*out_attrs, 0, (*in_attrs)[0]);
+  TYPE_ASSIGN_CHECK(*in_attrs, 0, (*out_attrs)[0]);
+  return (*out_attrs)[0] != -1;
+}
+
+template <int ndim, int req, bool back>
+struct diag {
+  template <typename DType>
+  MSHADOW_XINLINE static void Map(index_t i, DType *out, const DType *a,
+                                  index_t stride, index_t offset) {
+    using namespace mxnet_op;
+    index_t j = offset + stride * i;
+
+    if (back) {
+      KERNEL_ASSIGN(out[j], req, a[i]);
+    } else {
+      KERNEL_ASSIGN(out[i], req, a[j]);
+    }
+  }
+};
+
+template <int req, bool back>
+struct diag_gen {
+  template <typename DType>
+  MSHADOW_XINLINE static void Map(index_t i, DType *out, const DType *a,
+                                  mshadow::Shape<2> oshape, int k) {
+    using namespace mxnet_op;
+
+    auto j = unravel(i, oshape);
+    if (j[1] == (j[0] + k)) {
+      auto l = j[0] < j[1] ? j[0] : j[1];
+      if (back) {
+        KERNEL_ASSIGN(out[l], req, a[i]);
+      } else {
+        KERNEL_ASSIGN(out[i], req, a[l]);
+      }
+    } else if (!back) {
+      KERNEL_ASSIGN(out[i], req, static_cast<DType>(0));
+    }
+  }
+};
+
+template <typename xpu, bool back>
+void NumpyDiagOpImpl(const TBlob &in_data,
+                     const TBlob &out_data,
+                     const mxnet::TShape &ishape,
+                     const mxnet::TShape &oshape,
+                     index_t dsize,
+                     const int &k,
+                     mxnet_op::Stream<xpu> *s,
+                     const OpReqType &req) {
+  using namespace mxnet_op;
+  using namespace mshadow;
+  if (ishape.ndim() > 1) {
+    index_t stride1 = ishape[1], stride2 = 1;
+    // stride1 + stride2 is the stride for
+    // iterating over the diagonal in question
+
+    // the extra index offset introduced by k
+    index_t offset;
+    if (k > 0) {
+      offset = stride2 * k;
+    } else if (k < 0) {
+      offset = stride1 * -k;
+    } else {
+      offset = 0;
+    }
+
+    MSHADOW_TYPE_SWITCH(out_data.type_flag_, DType, {
+      MXNET_ASSIGN_REQ_SWITCH(req, req_type, {
+        if (back && req != kAddTo && req != kNullOp) {
+          out_data.FlatTo1D<xpu, DType>(s) = 0;
+        }
+
+        Kernel<diag<2, req_type, back>, xpu>::Launch(
+            s, dsize, out_data.dptr<DType>(), in_data.dptr<DType>(),
+            stride1 + stride2, offset);
+      });
+    });
+  } else {
+    MSHADOW_TYPE_SWITCH(out_data.type_flag_, DType, {
+      MXNET_ASSIGN_REQ_SWITCH(req, req_type, {
+        Kernel<diag_gen<req_type, back>, xpu>::Launch(
+            s, dsize, out_data.dptr<DType>(), in_data.dptr<DType>(),
+            Shape2(oshape[0], oshape[1]), k);
+      });
+    });
+  }
+}
+
+template <typename xpu>
+void NumpyDiagOpForward(const nnvm::NodeAttrs &attrs,
+                        const OpContext &ctx,
+                        const std::vector<TBlob> &inputs,
+                        const std::vector<OpReqType> &req,
+                        const std::vector<TBlob> &outputs) {
+  using namespace mxnet_op;
+  using namespace mshadow;
+  CHECK_EQ(inputs.size(), 1U);
+  CHECK_EQ(outputs.size(), 1U);
+  CHECK_EQ(req.size(), 1U);
+  CHECK_EQ(req[0], kWriteTo);
+  Stream<xpu> *s = ctx.get_stream<xpu>();
+  const TBlob &in_data = inputs[0];
+  const TBlob &out_data = outputs[0];
+  const mxnet::TShape &ishape = inputs[0].shape_;
+  const mxnet::TShape &oshape = outputs[0].shape_;
+  const NumpyDiagParam &param = nnvm::get<NumpyDiagParam>(attrs.parsed);
+
+  NumpyDiagOpImpl<xpu, false>(in_data, out_data, ishape, oshape,
+                              out_data.Size(), param.k, s, req[0]);
+}
+
+template <typename xpu>
+void NumpyDiagOpBackward(const nnvm::NodeAttrs &attrs,
+                         const OpContext &ctx,
+                         const std::vector<TBlob> &inputs,
+                         const std::vector<OpReqType> &req,
+                         const std::vector<TBlob> &outputs) {
+  using namespace mxnet_op;
+  using namespace mshadow;
+  CHECK_EQ(inputs.size(), 1U);
+  CHECK_EQ(outputs.size(), 1U);
+  Stream<xpu> *s = ctx.get_stream<xpu>();
+
+  const TBlob &in_data = inputs[0];
+  const TBlob &out_data = outputs[0];
+  const mxnet::TShape &ishape = inputs[0].shape_;
+  const mxnet::TShape &oshape = outputs[0].shape_;
+  const NumpyDiagParam &param = nnvm::get<NumpyDiagParam>(attrs.parsed);
+
+  NumpyDiagOpImpl<xpu, true>(in_data, out_data, oshape, ishape,
+                             in_data.Size(), param.k, s, req[0]);
+}
+
 }  // namespace op
 }  // namespace mxnet
 

src/operator/numpy/np_matrix_op.cc

@@ -36,6 +36,7 @@ DMLC_REGISTER_PARAMETER(NumpyMoveaxisParam);
 DMLC_REGISTER_PARAMETER(NumpyRot90Param);
 DMLC_REGISTER_PARAMETER(NumpyReshapeParam);
 DMLC_REGISTER_PARAMETER(NumpyXReshapeParam);
+DMLC_REGISTER_PARAMETER(NumpyDiagParam);
 
 
 bool NumpyTransposeShape(const nnvm::NodeAttrs& attrs,
@@ -1302,5 +1303,27 @@ NNVM_REGISTER_OP(_npi_hsplit_backward)
 })
 .set_attr<FCompute>("FCompute<cpu>", HSplitOpBackward<cpu>);
 
+NNVM_REGISTER_OP(_np_diag)
+.set_attr_parser(ParamParser<NumpyDiagParam>)
+.set_num_inputs(1)
+.set_num_outputs(1)
+.set_attr<nnvm::FListInputNames>("FListInputNames",
+    [](const NodeAttrs &attrs) {
+    return std::vector<std::string>{"data"};
+})
+.set_attr<mxnet::FInferShape>("FInferShape", NumpyDiagOpShape)
+.set_attr<nnvm::FInferType>("FInferType", NumpyDiagOpType)
+.set_attr<FCompute>("FCompute<cpu>", NumpyDiagOpForward<cpu>)
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseNone{"_backward_diag"})
+.add_argument("data", "NDArray-or-Symbol", "Input ndarray")
+.add_arguments(NumpyDiagParam::__FIELDS__());
+
+NNVM_REGISTER_OP(_backward_np_diag)
+.set_attr_parser(ParamParser<NumpyDiagParam>)
+.set_num_inputs(1)
+.set_num_outputs(1)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<FCompute>("FCompute<cpu>", NumpyDiagOpBackward<cpu>);
+
 }  // namespace op
 }  // namespace mxnet

src/operator/numpy/np_matrix_op.cu

@@ -118,5 +118,11 @@ NNVM_REGISTER_OP(_npi_hsplit_backward)
 NNVM_REGISTER_OP(_npx_reshape)
 .set_attr<FCompute>("FCompute<gpu>", UnaryOp::IdentityCompute<gpu>);
 
+NNVM_REGISTER_OP(_np_diag)
+.set_attr<FCompute>("FCompute<gpu>", NumpyDiagOpForward<gpu>);
+
+NNVM_REGISTER_OP(_backward_np_diag)
+.set_attr<FCompute>("FCompute<gpu>", NumpyDiagOpBackward<gpu>);
+
 }  // namespace op
 }  // namespace mxnet

tests/python/unittest/test_numpy_interoperability.py

@@ -57,6 +57,35 @@ def get_workloads(name):
         return OpArgMngr._args.get(name, None)
 
 
+def _add_workload_diag():
+    def get_mat(n):
+        data = _np.arange(n)
+        data = _np.add.outer(data, data)
+        return data
+
+    A = np.array([[1, 2], [3, 4], [5, 6]])
+    vals = (100 * np.arange(5)).astype('l')
+    vals_c = (100 * np.array(get_mat(5)) + 1).astype('l')
+    vals_f = _np.array((100 * get_mat(5) + 1), order ='F', dtype ='l')
+    vals_f = np.array(vals_f)
+
+    OpArgMngr.add_workload('diag', A, k= 2)
+    OpArgMngr.add_workload('diag', A, k= 1)
+    OpArgMngr.add_workload('diag', A, k= 0)
+    OpArgMngr.add_workload('diag', A, k= -1)
+    OpArgMngr.add_workload('diag', A, k= -2)
+    OpArgMngr.add_workload('diag', A, k= -3)
+    OpArgMngr.add_workload('diag', vals, k= 0)
+    OpArgMngr.add_workload('diag', vals, k= 2)
+    OpArgMngr.add_workload('diag', vals, k= -2)
+    OpArgMngr.add_workload('diag', vals_c, k= 0)
+    OpArgMngr.add_workload('diag', vals_c, k= 2)
+    OpArgMngr.add_workload('diag', vals_c, k= -2)
+    OpArgMngr.add_workload('diag', vals_f, k= 0)
+    OpArgMngr.add_workload('diag', vals_f, k= 2)
+    OpArgMngr.add_workload('diag', vals_f, k= -2)
+
+
 def _add_workload_concatenate(array_pool):
     OpArgMngr.add_workload('concatenate', [array_pool['4x1'], array_pool['4x1']])
     OpArgMngr.add_workload('concatenate', [array_pool['4x1'], array_pool['4x1']], axis=1)
@@ -89,6 +118,7 @@ def _add_workload_concatenate(array_pool):
 
 
 def _add_workload_append():
+
     def get_new_shape(shape, axis):
         shape_lst = list(shape)
         if axis is not None:
@@ -1226,6 +1256,7 @@ def _prepare_workloads():
     _add_workload_copy()
     _add_workload_cumsum()
     _add_workload_ravel()
+    _add_workload_diag()
     _add_workload_dot()
     _add_workload_expand_dims()
     _add_workload_fix()