[Do not review] [Do not merge] New numpy-compatible sum (apache#14739)

* Add numpy namespace and initial impl of np.sum (not complete)

* Clean up

* Fix import error

* numpy sum

* add test and backward data type support

* add license to test_numpy_op.py

* improve test to reduce flakiness

* fix sanity build

* extra numeric test and imperative test

* add error message for initial argument

python/mxnet/__init__.py

@@ -26,6 +26,7 @@
 from .base import MXNetError
 from .util import is_np_shape, set_np_shape, np_shape, use_np_shape
 from . import base
+from . import numpy
 from . import contrib
 from . import ndarray
 from . import ndarray as nd

python/mxnet/base.py

@@ -561,7 +561,7 @@ def _as_list(obj):
         return [obj]
 
 
-_OP_NAME_PREFIX_LIST = ['_contrib_', '_linalg_', '_sparse_', '_image_', '_random_']
+_OP_NAME_PREFIX_LIST = ['_contrib_', '_linalg_', '_sparse_', '_image_', '_random_', '_numpy_']
 
 
 def _get_op_name_prefix(op_name):
@@ -607,6 +607,15 @@ def _init_op_module(root_namespace, module_name, make_op_func):
     # use mx.nd.contrib or mx.sym.contrib from now on
     contrib_module_name_old = "%s.contrib.%s" % (root_namespace, module_name)
     contrib_module_old = sys.modules[contrib_module_name_old]
+    # special handling of registering numpy ops
+    # only expose mxnet.numpy.op_name to users for imperative mode.
+    # Symbolic mode should be used in Gluon.
+    if module_name == 'ndarray':
+        numpy_module_name = "%s.numpy" % root_namespace
+        numpy_module = sys.modules[numpy_module_name]
+    else:
+        numpy_module_name = None
+        numpy_module = None
     submodule_dict = {}
     for op_name_prefix in _OP_NAME_PREFIX_LIST:
         submodule_dict[op_name_prefix] =\
@@ -645,6 +654,16 @@ def _init_op_module(root_namespace, module_name, make_op_func):
             function.__module__ = contrib_module_name_old
             setattr(contrib_module_old, function.__name__, function)
             contrib_module_old.__all__.append(function.__name__)
+        elif op_name_prefix == '_numpy_' and numpy_module_name is not None:
+            # only register numpy ops under mxnet.numpy in imperative mode
+            hdl = OpHandle()
+            check_call(_LIB.NNGetOpHandle(c_str(name), ctypes.byref(hdl)))
+            # TODO(reminisce): Didn't consider third level module here, e.g. mxnet.numpy.random.
+            func_name = name[len(op_name_prefix):]
+            function = make_op_func(hdl, name, func_name)
+            function.__module__ = numpy_module_name
+            setattr(numpy_module, function.__name__, function)
+            numpy_module.__all__.append(function.__name__)
 
 
 def _generate_op_module_signature(root_namespace, module_name, op_code_gen_func):

python/mxnet/ndarray/__init__.py

@@ -17,7 +17,7 @@
 
 """NDArray API of MXNet."""
 
-from . import _internal, contrib, linalg, op, random, sparse, utils, image, ndarray
+from . import _internal, contrib, linalg, op, random, sparse, utils, image, ndarray, numpy
 # pylint: disable=wildcard-import, redefined-builtin
 try:
     from .gen_op import * # pylint: disable=unused-wildcard-import

python/mxnet/ndarray/numpy.py

@@ -0,0 +1,18 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+__all__ = []

python/mxnet/numpy/__init__.py

@@ -0,0 +1,20 @@
+#!/usr/bin/env python
+
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+__all__ = []

python/mxnet/symbol/__init__.py

@@ -17,7 +17,7 @@
 
 """Symbol API of MXNet."""
 
-from . import _internal, contrib, linalg, op, random, sparse, image, symbol
+from . import _internal, contrib, linalg, op, random, sparse, image, symbol, numpy
 # pylint: disable=wildcard-import, redefined-builtin
 try:
     from .gen_op import * # pylint: disable=unused-wildcard-import

python/mxnet/symbol/numpy.py

@@ -0,0 +1,18 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+__all__ = []

src/operator/numpy/np_broadcast_reduce_op.h

@@ -0,0 +1,218 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ *  Copyright (c) 2015 by Contributors
+ * \file broadcast_reduce_op.h
+ * \brief Function definition of broadcast and reduce operators
+ */
+#ifndef MXNET_OPERATOR_NUMPY_NP_BROADCAST_REDUCE_OP_H_
+#define MXNET_OPERATOR_NUMPY_NP_BROADCAST_REDUCE_OP_H_
+
+#include <algorithm>
+#include <vector>
+#include "../tensor/broadcast_reduce_op.h"
+
+namespace mxnet {
+namespace op {
+
+struct NumpyReduceAxesParam : public dmlc::Parameter<NumpyReduceAxesParam> {
+  dmlc::optional<mxnet::Tuple<int>> axis;
+  dmlc::optional<int> dtype;
+  bool keepdims;
+  dmlc::optional<double> initial;
+  DMLC_DECLARE_PARAMETER(NumpyReduceAxesParam) {
+    DMLC_DECLARE_FIELD(axis)
+      .set_default(dmlc::optional<mxnet::Tuple<int>>())
+      .describe("Axis or axes along which a sum is performed. The default, axis=None, will sum "
+                "all of the elements of the input array. If axis is negative it counts from the "
+                "last to the first axis.");
+    DMLC_DECLARE_FIELD(dtype)
+      .add_enum("float16", mshadow::kFloat16)
+      .add_enum("float32", mshadow::kFloat32)
+      .add_enum("float64", mshadow::kFloat64)
+      .add_enum("int8", mshadow::kInt8)
+      .add_enum("int32", mshadow::kInt32)
+      .add_enum("int64", mshadow::kInt64)
+      .set_default(dmlc::optional<int>())
+      .describe("The type of the returned array and of the accumulator in which the elements are "
+                "summed. The dtype of a is used by default unless a has an integer dtype of less "
+                "precision than the default platform integer. In that case, if a is signed then "
+                "the platform integer is used while if a is unsigned then an unsigned integer of "
+                "the same precision as the platform integer is used.");
+    DMLC_DECLARE_FIELD(keepdims).set_default(false)
+      .describe("If this is set to `True`, the reduced axes are left "
+                "in the result as dimension with size one.");
+    DMLC_DECLARE_FIELD(initial).set_default(dmlc::optional<double>())
+      .describe("Starting value for the sum.");
+  }
+};
+
+inline TShape NumpyReduceAxesShapeImpl(const TShape& ishape,
+                                       const dmlc::optional<mxnet::Tuple<int>>& axis,
+                                       bool keepdims) {
+  // TODO(junwu): improve the logic
+  // If input is a scalar, output should be a scalar too
+  if (ishape.ndim() == 0) {
+    if (axis.has_value()) {
+      const mxnet::Tuple<int>& axes = axis.value();
+      if (axes.ndim() > 0) {
+        CHECK_EQ(axes.ndim(), 1);
+        CHECK(axes[0] == 0 || axes[0] == -1);
+      }
+    }
+    return TShape(0, -1);
+  }
+
+  // axis=None, do global reduction
+  if (!axis.has_value()) {
+    if (keepdims) {
+      return TShape(ishape.ndim(), 1);
+    } else {
+      return TShape(0, -1);
+    }
+  }
+
+  // axis = (), will return identity(input)
+  if (axis.value().ndim() == 0) {
+    return ishape;
+  }
+
+  // axis has value
+  mxnet::Tuple<int> axes(axis.value());
+  for (index_t i = 0; i < axes.ndim(); i++) {
+    if (axes[i] < 0) {
+      axes[i] += ishape.ndim();
+    }
+  }
+  std::sort(axes.begin(), axes.end());
+
+  for (index_t i = 1; i < axes.ndim(); i++) {
+    CHECK_LT(axes[i-1], axes[i])
+        << "Reduction axes have duplicates "
+        << axes;
+  }
+  CHECK_LT(axes[axes.ndim()-1], ishape.ndim())
+      << "Reduction axis " << axes[axes.ndim()-1]
+      << " Exceeds input dimensions " << ishape;
+  CHECK_GE(axes[0], 0)
+      << "Reduction axis " << axis.value()
+      << " Exceeds input dimensions " << ishape;
+
+  TShape oshape;
+  if (keepdims) {
+    oshape = TShape(ishape);
+  } else {
+    oshape = TShape(ishape.ndim() - axes.ndim(), -1);
+  }
+
+  if (keepdims) {
+    for (index_t i = 0; i < axes.ndim(); ++i) {
+      oshape[axes[i]] = 1;
+    }
+  } else {
+    for (index_t i = 0, j = 0, k = 0; i < ishape.ndim(); ++i) {
+      if (j < axes.ndim() && i == axes[j]) {
+        ++j;
+        continue;
+      }
+      oshape[k++] = ishape[i];
+    }
+  }
+  return oshape;
+}
+
+inline bool NumpyReduceAxesShape(const nnvm::NodeAttrs& attrs,
+                                 std::vector<TShape> *in_attrs,
+                                 std::vector<TShape> *out_attrs) {
+  CHECK_EQ(in_attrs->size(), 1U);
+  CHECK_EQ(out_attrs->size(), 1U);
+  if (!shape_is_known(in_attrs->at(0))) {
+    return false;
+  }
+  const NumpyReduceAxesParam& param = nnvm::get<NumpyReduceAxesParam>(attrs.parsed);
+  SHAPE_ASSIGN_CHECK(*out_attrs, 0,
+                     NumpyReduceAxesShapeImpl((*in_attrs)[0], param.axis, param.keepdims));
+  return shape_is_known(out_attrs->at(0));
+}
+
+template<bool safe_acc_hint = false>
+inline bool NeedSafeAcc(int itype, int otype) {
+  bool rule = (itype != otype) || (itype != mshadow::kFloat32 && itype != mshadow::kFloat64);
+  return safe_acc_hint && rule;
+}
+
+template<typename xpu, typename reducer, bool safe_acc_hint = false, bool normalize = false,
+         typename OP = op::mshadow_op::identity>
+void NumpyReduceAxesCompute(const nnvm::NodeAttrs& attrs,
+                            const OpContext& ctx,
+                            const std::vector<TBlob>& inputs,
+                            const std::vector<OpReqType>& req,
+                            const std::vector<TBlob>& outputs) {
+  const NumpyReduceAxesParam& param = nnvm::get<NumpyReduceAxesParam>(attrs.parsed);
+  if (param.initial.has_value()) {
+    LOG(FATAL) << "initial is not supported yet";
+  }
+  if (param.axis.has_value() && param.axis.value().ndim() == 0) {
+    UnaryOp::IdentityCompute<xpu>(attrs, ctx, inputs, req, outputs);
+  }
+  TShape small;
+  if (param.keepdims) {
+    small = outputs[0].shape_;
+  } else {
+    small = NumpyReduceAxesShapeImpl(inputs[0].shape_, param.axis, true);
+  }
+
+  if (NeedSafeAcc<safe_acc_hint>(inputs[0].type_flag_, outputs[0].type_flag_)) {
+    ReduceAxesComputeImpl<xpu, reducer, true, normalize, OP>(ctx, inputs, req, outputs, small);
+  } else {
+    ReduceAxesComputeImpl<xpu, reducer, false, normalize, OP>(ctx, inputs, req, outputs, small);
+  }
+}
+
+template<typename xpu, bool normalize = false>
+inline void NumpyReduceAxesBackwardUseNone(const nnvm::NodeAttrs& attrs,
+                                           const OpContext& ctx,
+                                           const std::vector<TBlob>& inputs,
+                                           const std::vector<OpReqType>& req,
+                                           const std::vector<TBlob>& outputs) {
+  using namespace mshadow;
+  using namespace mshadow::expr;
+  const NumpyReduceAxesParam& param = nnvm::get<NumpyReduceAxesParam>(attrs.parsed);
+  TShape small;
+  if (param.keepdims) {
+    small = inputs[0].shape_;
+  } else {
+    small = NumpyReduceAxesShapeImpl(outputs[0].shape_, param.axis, true);
+  }
+
+  BroadcastComputeImpl<xpu>(attrs, ctx, inputs, req, outputs, small);
+  if (normalize) {
+    Stream<xpu> *s = ctx.get_stream<xpu>();
+    MSHADOW_TYPE_SWITCH(outputs[0].type_flag_, IType, {
+      Tensor<xpu, 1, IType> igrad = outputs[0].FlatTo1D<xpu, IType>(s);
+      printf("output size: %lu input_size: %lu\n", outputs[0].Size(), inputs[0].Size());
+      igrad /= scalar<IType>(outputs[0].Size()/inputs[0].Size());
+    });
+  }
+}
+
+}  // namespace op
+}  // namespace mxnet
+#endif  // MXNET_OPERATOR_NUMPY_NP_BROADCAST_REDUCE_OP_H_