!1228 Adapt tbe op UnsortedSegmentMin for GE.

Merge pull request !1228 from liuxiao/UnsortedSegmentMin

mindspore/ccsrc/operator/ops.cc

@@ -138,6 +138,7 @@ const PrimitivePtr kPrimSize = std::make_shared<Primitive>("Size");
 const PrimitivePtr kPrimArgMax = std::make_shared<Primitive>("Argmax");
 const PrimitivePtr kPrimPack = std::make_shared<Primitive>("Pack");
 const PrimitivePtr kPrimUnsortedSegmentSum = std::make_shared<Primitive>("UnsortedSegmentSum");
+const PrimitivePtr kPrimUnsortedSegmentMin = std::make_shared<Primitive>("UnsortedSegmentMin");
 const PrimitivePtr kPrimConcatOffset = std::make_shared<Primitive>("ConcatOffset");
 const PrimitivePtr kPrimReshape = std::make_shared<Primitive>("Reshape");
 const PrimitivePtr kPrimTile = std::make_shared<Primitive>("Tile");

mindspore/ccsrc/operator/ops.h

@@ -143,6 +143,7 @@ extern const PrimitivePtr kPrimSize;
 extern const PrimitivePtr kPrimArgMax;
 extern const PrimitivePtr kPrimPack;
 extern const PrimitivePtr kPrimUnpack;
+extern const PrimitivePtr kPrimUnsortedSegmentMin;
 extern const PrimitivePtr kPrimUnsortedSegmentSum;
 extern const PrimitivePtr kPrimConcatOffset;
 extern const PrimitivePtr kPrimReshape;

mindspore/ccsrc/transform/convert.cc

@@ -341,6 +341,7 @@ std::unordered_map<std::string, OpAdapterDescPtr> &DfGraphConvertor::get_adpt_ma
     {prim::kPrimGelu->name(), ADPT_DESC(Gelu)},
     {prim::kPrimGeluGrad->name(), ADPT_DESC(GeluGrad)},
     {string(kNameStridedSlice), ADPT_DESC(StridedSlice)},
+    {prim::kPrimUnsortedSegmentMin->name(), ADPT_DESC(UnsortedSegmentMinD)},
     {prim::kPrimUnsortedSegmentSum->name(), ADPT_DESC(UnsortedSegmentSumD)},
     {string(kNameExpandDims), ADPT_DESC(ExpandDims)},
     {prim::kPrimSqueeze->name(), ADPT_DESC(Squeeze)},

mindspore/ccsrc/transform/op_declare.cc

@@ -1053,6 +1053,12 @@ INPUT_ATTR_MAP(UnsortedSegmentSumD) = {{3, ATTR_DESC(num_segments, AnyTraits<int
 ATTR_MAP(UnsortedSegmentSumD) = EMPTY_ATTR_MAP;
 OUTPUT_MAP(UnsortedSegmentSumD) = {{0, OUTPUT_DESC(y)}};
 
+// UnsortedSegmentMin
+INPUT_MAP(UnsortedSegmentMinD) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(segment_ids)}};
+INPUT_ATTR_MAP(UnsortedSegmentMinD) = {{3, ATTR_DESC(num_segments, AnyTraits<int64_t>())}};
+ATTR_MAP(UnsortedSegmentMinD) = EMPTY_ATTR_MAP;
+OUTPUT_MAP(UnsortedSegmentMinD) = {{0, OUTPUT_DESC(y)}};
+
 // ExpandDims
 INPUT_MAP(ExpandDims) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(axis)}};
 ATTR_MAP(ExpandDims) = EMPTY_ATTR_MAP;

mindspore/ccsrc/transform/op_declare.h

@@ -283,6 +283,9 @@ DECLARE_OP_USE_OUTPUT(StridedSlice)
 DECLARE_OP_ADAPTER(UnsortedSegmentSumD)
 DECLARE_OP_USE_INPUT_ATTR(UnsortedSegmentSumD)
 DECLARE_OP_USE_OUTPUT(UnsortedSegmentSumD)
+DECLARE_OP_ADAPTER(UnsortedSegmentMinD)
+DECLARE_OP_USE_INPUT_ATTR(UnsortedSegmentMinD)
+DECLARE_OP_USE_OUTPUT(UnsortedSegmentMinD)
 DECLARE_OP_ADAPTER(ExpandDims)
 DECLARE_OP_USE_OUTPUT(ExpandDims)
 DECLARE_OP_ADAPTER(Squeeze)

mindspore/ops/_grad/grad_array_ops.py

@@ -22,6 +22,7 @@ from .. import functional as F
 from .grad_base import bprop_getters
 from ..primitive import constexpr
 from ... import context
+from ...common import dtype as mstype
 
 reduce_sum = P.ReduceSum()
 unsorted_segment_sum = P.UnsortedSegmentSum()
@@ -29,6 +30,7 @@ transpose = P.Transpose()
 shape_op = P.Shape()
 reshape = P.Reshape()
 invert_permutation = P.InvertPermutation()
+logical_and = P.LogicalAnd()
 
 
 @bprop_getters.register(P.Fill)
@@ -456,6 +458,57 @@ def get_bprop_diag_part(self):
     return bprop
 
 
+def _GatherDropNegatives(params,
+                         ids,
+                         zero_clipped_indices=None,
+                         is_positive=None):
+    """Helper function for unsorted segment ops."""
+    maximum = P.Maximum()
+    gather = P.GatherV2()
+    greater_equal = P.GreaterEqual()
+    rank = P.Rank()
+    fill = P.Fill()
+    select = P.Select()
+
+    if zero_clipped_indices is None:
+        zero_clipped_indices = maximum(ids, zeros_like(ids))
+    gathered = gather(params, zero_clipped_indices, 0)
+    if is_positive is None:
+        is_positive = greater_equal(ids, 0)
+        is_positive_shape = shape_op(is_positive)
+        broadcastable_shape = is_positive_shape
+        for _ in range(rank(gathered) - rank(is_positive)):
+            broadcastable_shape += (1,)
+        is_positive = reshape(is_positive, broadcastable_shape)
+        gathered_shape = shape_op(gathered)
+        is_positive = logical_and(is_positive, fill(mstype.bool_, gathered_shape, 1))
+    zero_slice = zeros_like(gathered)
+    return (select(is_positive, gathered, zero_slice), zero_clipped_indices, is_positive)
+
+
+@bprop_getters.register(P.UnsortedSegmentMin)
+def get_bprop_unsorted_segment_min(self):
+    """Generate bprop for UnsortedSegmentMin"""
+    equal = P.Equal()
+    cast = P.Cast()
+    divide = P.RealDiv()
+    get_dtype = P.DType()
+    select = P.Select()
+
+    def bprop(x, segment_ids, num_segments, out, dout):
+        gathered_outputs, zero_clipped_indices, is_positive = _GatherDropNegatives(out, segment_ids)
+        is_selected = equal(x, gathered_outputs)
+        is_selected = logical_and(is_selected, is_positive)
+        num_selected = unsorted_segment_sum(cast(is_selected, get_dtype(dout)),
+                                            segment_ids, num_segments)
+        weighted_grads = divide(dout, num_selected)
+        gathered_grads, _, _ = _GatherDropNegatives(weighted_grads, None,
+                                                    zero_clipped_indices, is_positive)
+        zeros = zeros_like(gathered_grads)
+        return select(is_selected, gathered_grads, zeros), zeros_like(segment_ids), zeros_like(num_segments)
+    return bprop
+
+
 @bprop_getters.register(P.SpaceToBatch)
 def get_bprop_space_to_batch(self):
     """Generate bprop for SpaceToBatch"""

mindspore/ops/operations/__init__.py

@@ -28,7 +28,7 @@ from .array_ops import (Argmax, Argmin, Cast, Concat, Pack, Unpack,
                         ScalarToArray, ScalarToTensor, ScatterNd, ScatterNdUpdate, Select,
                         Shape, Size, Slice, Split,
                         Squeeze, StridedSlice, Tile,
-                        Transpose, TruncatedNormal, TupleToArray,
+                        Transpose, TruncatedNormal, TupleToArray, UnsortedSegmentMin,
                         UnsortedSegmentSum, SpaceToDepth, DepthToSpace, SpaceToBatch, BatchToSpace)
 from .comm_ops import (AllGather, AllReduce, _AlltoAll, ReduceScatter, Broadcast,
                        _MirrorOperator, ReduceOp, _VirtualDataset,
@@ -96,6 +96,7 @@ __all__ = [
     'MaxPool',
     'TopK',
     'Adam',
+    'Softplus',
     'Softmax',
     'LogSoftmax',
     'SoftmaxCrossEntropyWithLogits',
@@ -210,6 +211,7 @@ __all__ = [
     'Size',
     'DepthwiseConv2dNative',
     'UnsortedSegmentSum',
+    'UnsortedSegmentMin',
     "AllGather",
     "AllReduce",
     "ReduceScatter",

mindspore/ops/operations/array_ops.py

@@ -1253,6 +1253,54 @@ class UnsortedSegmentSum(PrimitiveWithInfer):
         return out
 
 
+class UnsortedSegmentMin(PrimitiveWithInfer):
+    """
+    Computes the minimum along segments of a tensor.
+
+    If the given segment_ids is negative, the value will be ignored.
+
+    Inputs:
+        - **input_x** (Tensor) - The shape is :math:`(x_1, x_2, ..., x_R)`.
+        - **segment_ids** (Tensor) - A `1-D` tensor whose shape is a prefix of `x_shape`.
+        - **num_segments** (int) - The value spcifies the number of distinct `segment_ids`.
+
+    Outputs:
+        Tensor, Set the number of `num_segments` as `N`, the shape is :math:`(N, x_2, ..., x_R)`.
+
+    Examples:
+        >>> input_x = Tensor(np.array([[1, 2, 3], [4, 5, 6], [4, 2, 1]]).astype(np.float32))
+        >>> segment_ids = Tensor(np.array([0, 1, 1]).np.int32)
+        >>> num_segments = 2
+        >>> unsorted_segment_min = P.UnsortedSegmentMin()
+        >>> unsorted_segment_min(input_x, segment_ids, num_segments)
+        [[1., 2., 3.], [4., 2., 1.]]
+    """
+
+    @prim_attr_register
+    def __init__(self):
+        """init UnsortedSegmentMin"""
+        self.init_prim_io_names(inputs=['x', 'segment_ids', 'num_segments'], outputs=['y'])
+
+    def __infer__(self, x, segment_ids, num_segments):
+        x_type = x['dtype']
+        x_shape = x['shape']
+        segment_ids_shape = segment_ids['shape']
+        valid_type = [mstype.float16, mstype.float32, mstype.int32]
+        validator.check_tensor_type_same({"x": x['dtype']}, valid_type, self.name)
+        validator.check_tensor_type_same({"segment_ids": segment_ids['dtype']}, [mstype.int32], self.name)
+        validator.check_integer("rank of segment_ids_shape", len(segment_ids_shape), 1, Rel.EQ, self.name)
+        num_segments_v = num_segments['value']
+        validator.check_value_type('num_segments', num_segments_v, [int], self.name)
+        validator.check_integer("num_segments", num_segments_v, 0, Rel.GT, self.name)
+        segment_ids_shape_len = len(segment_ids_shape)
+        out_shape = [num_segments_v]
+        out_shape += x_shape[segment_ids_shape_len:]
+        out = {'shape': out_shape,
+               'dtype': x_type,
+               'value': None}
+        return out
+
+
 class Concat(PrimitiveWithInfer):
     r"""
     Concat tensor in specified axis.

tests/ut/python/ops/test_ops.py

@@ -778,6 +778,11 @@ test_case_nn_ops = [
         'desc_inputs': [[3, 2, 1, 3], Tensor(np.array([[0, 1], [0, 1], [0, 1]]).astype(np.int32))],
         'desc_bprop': [[4, 1, 3]],
         'skip': ['backward']}),
+    ('UnsortedSegmentMin', {
+        'block': P.UnsortedSegmentMin(),
+        'desc_const': [4],
+        'desc_inputs': [[3, 2, 1, 3], Tensor(np.array([1, 2, 3]).astype(np.int32))],
+        'desc_bprop': [[4, 2, 1, 3]]}),
     ('DropoutGenMask', {
         'block': P.DropoutGenMask(),
         'desc_const': [(2, 2), Tensor(0.5, mstype.float32)],