mindspore 2.0.0a0__cp37-none-any.whl → 2.0.0rc1__cp37-none-any.whl

mindspore/ops/_grad/grad_array_ops.py

@@ -18,7 +18,6 @@
 import numpy as np
 import mindspore as ms
 from mindspore import Tensor
-from mindspore.ops import composite as C
 from mindspore.ops import operations as P
 from mindspore.ops.operations import _grad_ops as G
 from mindspore.ops.operations.array_ops import Fills, NonZero
@@ -27,11 +26,11 @@ from mindspore.ops.functional import broadcast_gradient_args
 from mindspore.ops import functional as F
 from mindspore.ops._grad.grad_base import bprop_getters, create_tensor_by_element
 from mindspore.ops.primitive import constexpr
+from mindspore.ops.primitive import _primexpr
 from mindspore.common import dtype as mstype
 from mindspore.common.sparse_tensor import RowTensorInner
-from mindspore.ops._utils.utils import range_op, get_1d_shape, generate_shape_index, is_shape_unknown
-from mindspore.ops._grad.grad_base import dyn_rank, convert_to_tensor, dyn_invert_permutation, dyn_size, dyn_ones, \
-    dyn_fill
+from mindspore.ops._utils.utils import range_op, get_1d_shape, generate_shape_index
+from mindspore.ops._grad.grad_base import dyn_rank, convert_to_tensor, dyn_ones, dyn_fill
 from mindspore.ops._grad.grad_base import sum_grad_reduce_axis
 from mindspore.ops.operations._inner_ops import DynamicBroadcastGradientArgs
 from ..operations._inner_ops import DynamicBroadcastGradientArgs, IsSubClass
@@ -98,47 +97,6 @@ def get_bprop_dtype(self):
     return bprop
 
 
-dout_cast = C.MultitypeFuncGraph("dout_cast")
-
-
-@dout_cast.register("Tensor", "Tensor")
-def dout_cast_tensor(dout, x):
-    """Casts dout to the dtype of x for Tensor."""
-    cast = P.Cast()
-    get_dtype = P.DType()
-    dx = cast(dout, get_dtype(x))
-    return dx
-
-
-@dout_cast.register("Number", "Number")
-def dout_cast_number(dout, x):
-    """Casts dout to the dtype of x for Number."""
-    cast = P.Cast()
-    get_dtype = P.DType()
-    dx = cast(dout, get_dtype(x))
-    return dx
-
-
-@dout_cast.register("RowTensor", "Tensor")
-def dout_cast_row_tensor(dout, x):
-    """Casts dout values to the dtype of x for RowTensor."""
-    cast = P.Cast()
-    get_dtype = P.DType()
-    values = cast(dout.values, get_dtype(x))
-    return RowTensorInner(dout.indices, values, dout.dense_shape)
-
-
-@bprop_getters.register(P.Cast)
-def get_bprop_cast(self):
-    """Generate bprop for Cast"""
-
-    def bprop(x, t, out, dout):
-        dx = dout_cast(dout, x)
-        return dx, zeros_like(t)
-
-    return bprop
-
-
 @bprop_getters.register(P.Shape)
 def get_bprop_shape(self):
     """Generate bprop for Shape"""
@@ -198,7 +156,7 @@ def get_bprop_reshape(self):
 
     def bprop(x, shp, out, dout):
         shapex = shape_op(x)
-        if is_shape_unknown(shapex):
+        if F.is_sequence_value_unknown(shapex):
             shapex = dyn_shape_op(x)
         return reshape(dout, shapex), zeros_like(shp)
 
@@ -211,7 +169,7 @@ def get_bprop_expand_dims(self):
 
     def bprop(x, axis, out, dout):
         shapex = shape_op(x)
-        if is_shape_unknown(shapex):
+        if F.is_sequence_value_unknown(shapex):
             shapex = dyn_shape_op(x)
         return reshape(dout, shapex), zeros_like(axis)
 
@@ -224,7 +182,7 @@ def get_bprop_squeeze(self):
 
     def bprop(x, out, dout):
         shapex = shape_op(x)
-        if is_shape_unknown(shapex):
+        if F.is_sequence_value_unknown(shapex):
             shapex = dyn_shape_op(x)
         return (reshape(dout, shapex),)
 
@@ -238,7 +196,7 @@ def get_bprop_flatten(self):
 
     def bprop(x, out, dout):
         shape_x = shape_op(x)
-        if is_shape_unknown(shape_x):
+        if F.is_sequence_value_unknown(shape_x):
             shape_x = dyn_shape_op(x)
         dx = flatten_grad(dout, shape_x)
         return (dx,)
@@ -246,7 +204,7 @@ def get_bprop_flatten(self):
     return bprop
 
 
-@constexpr
+@_primexpr
 def _tile_shape(multiples, shapex):
     """Calculate [1,2], [3, 4] -> [1,3,2,4]."""
     len_muli = len(multiples)
@@ -297,7 +255,7 @@ def get_bprop_tile(self):
 
     def bprop(x, multiples, out, dout):
         shapex = shape_op(x)
-        if is_shape_unknown(shapex):
+        if F.is_sequence_value_unknown(shapex):
             shapex = dyn_shape_op(x)
         if isinstance(multiples, tuple) and isinstance(shapex, tuple):
             r_shape = _tile_shape(multiples, shapex)
@@ -340,7 +298,7 @@ def get_bprop_embedding_lookup(self):
 
     def bprop_sparse(x, indices, offset, out, dout):
         x_shp = shape_op(x)
-        if is_shape_unknown(x_shp):
+        if F.is_sequence_value_unknown(x_shp):
             raise RuntimeError("Now, EmbeddingLookup op's grad don't support Dynamic Sense!")
         new_indices = sub_op(indices, offset)
         indices_size = size_op(new_indices)
@@ -359,7 +317,7 @@ def get_bprop_embedding_lookup(self):
     return bprop_sparse
 
 
-@constexpr
+@_primexpr
 def make_begin(shp):
     """Creates a tuple with zero according to the shape."""
     begin = tuple([0 for _ in shp])
@@ -379,7 +337,7 @@ def get_bprop_padding(self):
     def bprop(x, out, dout):
         shp = shape_op(x)
         begin = ()
-        if is_shape_unknown(shp):
+        if F.is_sequence_value_unknown(shp):
             shp = dyn_shape_op(x)
             begin = make_dynamic_begin(shp)
         else:
@@ -390,36 +348,7 @@ def get_bprop_padding(self):
     return bprop
 
 
-@constexpr
-def _transpose_perm_positive(perm):
-    res = []
-    for value in perm:
-        value = value if (value >= 0) else (value + len(perm))
-        res.append(value)
-    return tuple(res)
-
-
-def _dyn_transpose_perm_positive(perm):
-    return (perm + dyn_size(perm)) % (dyn_size(perm))
-
-
-@bprop_getters.register(P.Transpose)
-def get_bprop_transpose(self):
-    """Generate bprop for Transpose"""
-
-    def bprop(x, perm, out, dout):
-        is_mutable, perm = convert_to_tensor(perm)
-        if is_mutable:
-            perm = _dyn_transpose_perm_positive(perm)
-            return transpose(dout, dyn_invert_permutation(perm)), zeros_like(perm)
-
-        perm = _transpose_perm_positive(perm)
-        return transpose(dout, invert_permutation(perm)), zeros_like(perm)
-
-    return bprop
-
-
-@constexpr
+@_primexpr
 def _concat_grad_uniform(input_shapes, input_nums):
     """Helper function for bprop of Concat"""
     is_uniform = True
@@ -472,12 +401,6 @@ def get_bprop_concat(self):
     return bprop
 
 
-@constexpr
-def _slice_grad_pad(begins, sizes, shapes):
-    pads = tuple((begin, shape - begin - size) for begin, size, shape in zip(begins, sizes, shapes))
-    return pads
-
-
 @bprop_getters.register(P.Slice)
 def get_bprop_slice(self):
     """Generate bprop for Slice"""
@@ -489,17 +412,17 @@ def get_bprop_slice(self):
     return bprop
 
 
-@constexpr
-def _generate_inverse_index(x_shape, axis):
+@_primexpr
+def _generate_inverse_index(x_shape, axis, batch_dims=0):
     x_rank = len(x_shape)
     index = tuple(range(x_rank))
     if axis < 0:
         axis += x_rank
-    perm = index[1:1 + axis] + (0,) + index[1 + axis:]
+    perm = index[:batch_dims] + index[batch_dims + 1:1 + axis] + (index[batch_dims],) + index[1 + axis:]
     return perm
 
 
-@constexpr
+@_primexpr
 def _regenerate_output_shape(x_shp, ind_shp, axis):
     rank = len(x_shp)
     if axis < 0:
@@ -517,29 +440,52 @@ def _dyn_regenerate_output_shape(x_shp, ind_shp, axis):
     return out_shape
 
 
-def _dyn_generate_shape_index(out_shape, indices_shape, axis):
+def _dyn_generate_shape_index(out_shape, indices_shape, axis, batch_dims=0):
     """Get tranpose order"""
     out_rank = F.reshape(dyn_shape_op(out_shape), ())
     ind_rank = F.reshape(dyn_shape_op(indices_shape), ())
     if axis < 0:
         axis += out_rank - ind_rank + 1
     perm_part1 = P.Range()(F.cast(0, mstype.int32), F.cast(20, mstype.int32), F.cast(1, mstype.int32))
-    perm_part1 = perm_part1[axis: axis + ind_rank]
+    ind_end = axis + ind_rank - batch_dims
+    perm_part1 = perm_part1[axis: ind_end]
     index = P.Range()(F.cast(0, mstype.int32), F.cast(out_rank, mstype.int32), F.cast(1, mstype.int32))
-    perm = P.Concat(0)((perm_part1, index[:axis], index[axis + ind_rank:]))
+    perm = F.hstack((index[:batch_dims], perm_part1, index[batch_dims:axis], index[ind_end:]))
     return perm
 
 
-def _dyn_generate_inverse_index(x_shp, axis):
+def _dyn_generate_inverse_index(x_shp, axis, batch_dims=0):
     """Get tranpose order"""
     x_rank = F.reshape(dyn_shape_op(x_shp), ())
     index = P.Range()(F.cast(0, mstype.int32), F.cast(x_rank, mstype.int32), F.cast(1, mstype.int32))
     if axis < 0:
         axis += x_rank
-    perm = P.Concat(0)((index[1: 1 + axis], Tensor([0], dtype=mstype.int32), index[1 + axis:]))
+    perm = F.hstack((index[:batch_dims], index[batch_dims + 1:1 + axis], index[batch_dims], index[1 + axis:]))
     return perm
 
 
+def calculate_batch_gather(values, indices, x_shape, axis, batch_dims):
+    """Calculate gather grad with batch_dims"""
+    values_shape = dyn_shape_op(values)
+    batch_size = F.prod(x_shape[:batch_dims])
+    batch_size = F.cast(batch_size, mstype.int32)
+    axis_dim = F.cast(x_shape[axis], mstype.int32)
+
+    # Move batch dimension to first non-batch dimension
+    values = values.reshape((-1,) + values.shape[batch_dims:])
+    indices = indices.reshape((-1,) + indices.shape[batch_dims:])
+    offset = P.Range()(F.cast(0, mstype.int32), batch_size * axis_dim, axis_dim)
+    offset_shape = F.hstack([batch_size] + [Tensor(1, dtype=mstype.int32) for _ in range(len(indices.shape) - 1)])
+    offset = reshape(offset, offset_shape)
+    indices = indices + offset
+    num_segments = batch_size * axis_dim
+    params_grad = unsorted_segment_sum(values, indices, num_segments)
+    grad_shape = dyn_shape_op(params_grad)
+    ret_shape = F.hstack([values_shape[:batch_dims], F.cast(axis_dim, mstype.int64), grad_shape[1:]])
+    params_grad = reshape(params_grad, ret_shape)
+    return params_grad
+
+
 @bprop_getters.register(P.Gather)
 @bprop_getters.register(P.GatherV2)
 def get_bprop_gather_v2(self):
@@ -551,6 +497,9 @@ def get_bprop_gather_v2(self):
         x_shp = dyn_shape_op(x)
         ind_shp = dyn_shape_op(indices)
         out_shp = dyn_shape_op(dout)
+        batch_dims = self.batch_dims
+        if batch_dims < 0:
+            batch_dims += F.reshape(dyn_shape_op(ind_shp), ())
 
         if F.rank(dout) == 0:
             dout = P.ExpandDims()(dout, -1)
@@ -560,17 +509,20 @@ def get_bprop_gather_v2(self):
             dout = reshape(dout, out_shp)
 
         # Example: out_shape:(3,2,3) axis 1 -> (1,0,2)
-        perm_1 = _dyn_generate_shape_index(out_shp, ind_shp, axis)
+        perm_1 = _dyn_generate_shape_index(out_shp, ind_shp, axis, batch_dims)
         values_transpose = transpose(dout, perm_1)
-        params_grad = unsorted_segment_sum(values_transpose, indices, x_shp[axis])
-        perm_2 = _dyn_generate_inverse_index(x_shp, axis)
+        if batch_dims > 0:
+            params_grad = calculate_batch_gather(values_transpose, indices, x_shp, axis, batch_dims)
+        else:
+            params_grad = unsorted_segment_sum(values_transpose, indices, x_shp[axis])
+        perm_2 = _dyn_generate_inverse_index(x_shp, axis, batch_dims)
         params_grad = transpose(params_grad, perm_2)
         return params_grad, zeros_like(orig_indices), zeros_like(axis)
 
     def bprop(x, indices, axis, out, dout):
         is_mutable, axis = convert_to_tensor(axis)
-        if (is_shape_unknown(shape_op(x)) or is_shape_unknown(shape_op(indices)) or \
-            is_shape_unknown(shape_op(dout))) and is_mutable:
+        if (F.is_sequence_value_unknown(shape_op(x)) or F.is_sequence_value_unknown(shape_op(indices)) or \
+                F.is_sequence_value_unknown(shape_op(dout))) and is_mutable:
             return _dyn_bprop_gather_v2(x, indices, axis, dout)
         orig_indices = indices
         if F.rank(dout) == 0:
@@ -585,15 +537,19 @@ def get_bprop_gather_v2(self):
         x_shp = shape_op(x)
         out_shp = shape_op(dout)
         ind_shp = shape_op(indices)
+        batch_dims = self.batch_dims
+        if batch_dims < 0:
+            batch_dims += len(ind_shp)
         # Example: out_shape:(3,2,3) axis 1 -> (1,0,2)
-        perm_1 = generate_shape_index(out_shp, ind_shp, axis)
+        perm_1 = generate_shape_index(out_shp, ind_shp, axis, batch_dims)
         values_transpose = transpose(dout, perm_1)
-        if is_shape_unknown(shape_op(x)):
-            params_grad = unsorted_segment_sum(values_transpose, indices, dyn_shape_op(x)[axis])
+        dyn_x_sape = dyn_shape_op(x)
+        if batch_dims > 0:
+            params_grad = calculate_batch_gather(values_transpose, indices, dyn_x_sape, axis, batch_dims)
         else:
-            params_grad = unsorted_segment_sum(values_transpose, indices, shape_op(x)[axis])
+            params_grad = unsorted_segment_sum(values_transpose, indices, dyn_x_sape[axis])
         # Example: out_shape:(3,2,3) axis 2 -> (1,2,0)
-        perm_2 = _generate_inverse_index(x_shp, axis)
+        perm_2 = _generate_inverse_index(x_shp, axis, batch_dims)
         params_grad = transpose(params_grad, perm_2)
         return params_grad, zeros_like(orig_indices), zeros_like(axis)
 
@@ -854,7 +810,7 @@ def get_bprop_strided_slice(self):
 
     def bprop(x, begin, end, strides, out, dout):
         x_shape = shape_op(x)
-        if is_shape_unknown(x_shape):
+        if F.is_sequence_value_unknown(x_shape):
             x_shape = dyn_shape_op(x)
         dx = input_grad(dout, x_shape, begin, end, strides)
         return dx, zeros_like(begin), zeros_like(end), zeros_like(strides)
@@ -926,7 +882,7 @@ def get_bprop_resize_nearest_neighbor(self):
     tensor_shape = P.TensorShape()
 
     def bprop(inputs, out, dout):
-        if (-1 in shape_op(inputs)) or (-2 in shape_op(inputs)):
+        if F.is_sequence_value_unknown(shape_op(inputs)) or F.is_sequence_shape_unknown(shape_op(inputs)):
             shp = tensor_shape(inputs)
         else:
             shp = shape_op(inputs)
@@ -944,7 +900,7 @@ def get_bprop_gather_nd(self):
 
     def bprop(x, indices, out, dout):
         shp = shape_op(x)
-        if is_shape_unknown(shp):
+        if F.is_sequence_value_unknown(shp):
             shp = dyn_shape_op(x)
         return op(indices, dout, shp), zeros_like(indices)
 
@@ -1127,7 +1083,7 @@ def _gather_drop_negatives(params,
         is_positive = greater_equal(ids, 0)
         is_positive_shape = shape_op(is_positive)
         gathered_shape = shape_op(gathered)
-        if is_shape_unknown(gathered_shape) or is_shape_unknown(is_positive_shape):
+        if F.is_sequence_value_unknown(gathered_shape) or F.is_sequence_value_unknown(is_positive_shape):
             gathered_shape = dyn_shape_op(gathered)
             rank_gathered = dyn_rank(gathered)
             fill_gathered = dyn_fill(mstype.int64, gathered_shape, 1)
@@ -1212,6 +1168,8 @@ def get_bprop_unsorted_segment_prod(self):
     unsorted_segment_prod = P.UnsortedSegmentProd()
 
     def bprop(x, segment_ids, num_segments, out, dout):
+        if x.dtype == mstype.complex64 or x.dtype == mstype.complex128:
+            raise TypeError("For 'UnsortedSegmentProd', complex number is not supported for gradient currently.")
         if x.dtype == mstype.complex64 or x.dtype == mstype.complex128:
             is_zero = equal(x, F.scalar_to_tensor(0).astype(x.dtype))
         else:
@@ -1301,19 +1259,30 @@ def get_bprop_broadcast_to(self):
         x_shape = shape_op(x)
         dout_shape = shape_op(dout)
         broadcast_shape = shape_op(out)
-        dynamic = is_shape_unknown(x_shape) or is_shape_unknown(dout_shape)
+        dynamic = F.is_sequence_value_unknown(x_shape) or F.is_sequence_value_unknown(dout_shape)
         if not dynamic and x_shape == dout_shape:
             return (dout,)
-        dynamic = dynamic or is_shape_unknown(broadcast_shape)
+        dynamic = dynamic or F.is_sequence_value_unknown(broadcast_shape)
+        out_type = dout.dtype
         if not dynamic:
             _, reduction_axes = broadcast_gradient_args(broadcast_shape, x_shape)
-            reduced_grad = reduce_keep_dim(dout, reduction_axes)
+            if out_type in (ms.int16, ms.int32, ms.int64):
+                dout = P.Cast()(dout, ms.float32)
+                reduced_grad = reduce_keep_dim(dout, reduction_axes)
+                reduced_grad = P.Cast()(reduced_grad, out_type)
+            else:
+                reduced_grad = reduce_keep_dim(dout, reduction_axes)
             dx = reshape(reduced_grad, x_shape)
         else:
             x_shape = dyn_shape_op(x)
             broadcast_shape = dyn_shape_op(out)
             _, reduction_axes = DynamicBroadcastGradientArgs()(broadcast_shape, x_shape)
-            reduced_grad = sum_grad_reduce_axis(dout, reduction_axes, keep_dims=True)
+            if out_type in (ms.int16, ms.int32, ms.int64):
+                dout = P.Cast()(dout, ms.float32)
+                reduced_grad = sum_grad_reduce_axis(dout, reduction_axes, keep_dims=True)
+                reduced_grad = P.Cast()(reduced_grad, out_type)
+            else:
+                reduced_grad = sum_grad_reduce_axis(dout, reduction_axes, keep_dims=True)
             dx = reshape(reduced_grad, x_shape)
         return (dx,)
 

mindspore/ops/_grad/grad_base.py

@@ -21,7 +21,9 @@ from mindspore.common import Tensor
 from mindspore.ops import operations as P
 from mindspore.common import dtype as mstype
 from mindspore.ops.operations._inner_ops import DynamicBroadcastTo
-
+from mindspore.ops import functional as F
+from mindspore.ops.operations import _sequence_ops as seq_op
+import mindspore as ms
 dyn_shape = P.TensorShape()
 cast = P.Cast()
 
@@ -91,6 +93,26 @@ def convert_to_tensor(data):
     if isinstance(data, Tensor):
         return True, data
 
+    if isinstance(data, list):
+        if F.is_sequence_value_unknown(data):
+            data_tensor = seq_op.ListToTensor()(data, ms.int64)
+            return True, data_tensor
+        return False, data
+    if isinstance(data, tuple):
+        if F.is_sequence_value_unknown(data):
+            data_tensor = seq_op.TupleToTensor()(data, ms.int64)
+            return True, data_tensor
+        return False, data
+    if isinstance(data, int):
+        if not F.isconstant(data):
+            data_tensor = F.scalar_to_tensor(data, ms.int64)
+            return True, data_tensor
+        return False, data
+    if isinstance(data, float):
+        if not F.isconstant(data):
+            data_tensor = F.scalar_to_tensor(data, ms.float32)
+            return True, data_tensor
+        return False, data
     return False, data
 
 

mindspore/ops/_grad/grad_clip_ops.py

@@ -16,15 +16,14 @@
 
 """Define the grad rules of clip operations."""
 from __future__ import absolute_import
-
 from mindspore.ops._grad.grad_base import bprop_getters, dyn_fill
 from mindspore.ops import operations as P
+from mindspore.ops import functional as F
 from mindspore.ops.operations import _inner_ops as inner
 from mindspore.ops.operations import _grad_ops as G
 from mindspore.common import dtype as mstype
 from mindspore.ops._grad.grad_math_ops import _sum_grad
 from mindspore.ops._grad.grad_math_ops import binop_grad_common
-from mindspore.ops._utils.utils import is_shape_unknown
 
 
 @bprop_getters.register(inner.ClipByNorm)
@@ -72,7 +71,7 @@ def get_bprop_clip_by_norm(self):
         temp_num = 2.0
         temp_out = mul_op(reduce_sum_dout_x, cast_x)
         shape_cast_x = shape_op(cast_x)
-        if is_shape_unknown(shape_cast_x):
+        if F.is_sequence_value_unknown(shape_cast_x):
             fill_x = dyn_fill(type_op(temp_out), dyn_shape_op(cast_x), temp_num)
         else:
             fill_x = fill_op(type_op(temp_out), shape_cast_x, temp_num)

mindspore/ops/_grad/grad_comm_ops.py

@@ -207,16 +207,16 @@ def get_bprop_mirror_micro_step_operator(self):
                 z = F.depend(z, dout)
                 real_grad = all_reduce(z)
                 real_grad = F.tensor_mul(real_grad, scale)
-                assign(z, real_grad)
-                assign_out = z
+                if opt_shard:
+                    return (real_grad, cast(out_tensor, dtype(z)))
+                return F.depend((cast(out_tensor, dtype(x)), cast(out_tensor, dtype(z))), assign(z, real_grad))
         else:
             if issubclass_(F.typeof(dout), mstype.tensor):
                 z = F.depend(z, dout)
                 real_grad = all_reduce(z)
-                assign(z, real_grad)
-                assign_out = z
-        if opt_shard:
-            return (real_grad, cast(out_tensor, dtype(z)))
+                if opt_shard:
+                    return (real_grad, cast(out_tensor, dtype(z)))
+                return F.depend((cast(out_tensor, dtype(x)), cast(out_tensor, dtype(z))), assign(z, real_grad))
         return F.depend((cast(out_tensor, dtype(x)), cast(out_tensor, dtype(z))), assign_out)
     return bprop
 
@@ -289,6 +289,7 @@ def get_bprop_mini_step_all_gather(self):
                 dx = F.depend(dx, z)
         else:
             dx = dout
+
         return (dx, zeros_like(z))
 
     return bprop
@@ -299,21 +300,33 @@ def get_bprop_micro_step_all_gather(self):
     """Generate bprop for _MicroStepAllGather"""
     fusion = self.get_attr_dict()["fusion"]
     mean_flag = self.get_attr_dict()["mean_flag"]
-    do_mirror = self.get_attr_dict()["do_mirror"]
-    scale = 1.0 / self.rank_size
-    all_reduce = AllReduce(ReduceOp.SUM, self.group).add_prim_attr("fusion", fusion)
-    rank = get_rank(self.group)
-    dev_num = get_group_size(self.group)
-    split = P.Split(output_num=dev_num)
-    if self.instance_name:
-        instance_name = "grad_" + self.instance_name
-        all_reduce.set_prim_instance_name(instance_name)
+    do_mirror = False
+    if self.group != "":
+        do_mirror = self.get_attr_dict()["do_mirror"]
+    if do_mirror:
+        scale = 1.0 / self.rank_size
+        all_reduce = AllReduce(ReduceOp.SUM, self.group).add_prim_attr("fusion", fusion)
+        rank = get_rank(self.group)
+        dev_num = get_group_size(self.group)
+        split = P.Split(output_num=dev_num)
+        if self.instance_name:
+            instance_name = "grad_" + self.instance_name
+            all_reduce.set_prim_instance_name(instance_name)
     cast = P.Cast()
     dtype = P.DType()
     out_tensor = Tensor(1.0, mstype.float16)
+    with_mirror_operator = self.get_attr_dict()["with_mirror_operator"]
 
     # z: accu_grad
     def bprop(x, z, out, dout):
+        if with_mirror_operator:
+            if not do_mirror:
+                return (dout, cast(out_tensor, dtype(z)))
+            real_grad = all_reduce(dout)
+            real_grad = split(real_grad)[rank]
+            if mean_flag:
+                real_grad = F.tensor_mul(real_grad, scale)
+            return (real_grad, cast(out_tensor, dtype(z)))
         z = F.depend(z, dout)
         if not do_mirror:
             return (z, cast(out_tensor, dtype(z)))
@@ -451,7 +464,9 @@ def get_bprop_mirror_operator(self):
     group = self.get_attr_dict()['group']
     dev_num = self.get_attr_dict()['dev_num']
     mean_flag = self.get_attr_dict()['mean_flag']
+    dev_num_r = 1.0
     if dev_num > 1:
+        dev_num_r = 1.0 / dev_num
         all_reduce = AllReduce(group=group)
         all_gather = AllGather(group=group)
         mul = P.Mul()
@@ -473,15 +488,11 @@ def get_bprop_mirror_operator(self):
         if mean_flag:
             if issubclass_(F.typeof(dout), mstype.tensor):
                 dx = all_reduce(dout)
-                float_one = F.scalar_cast(1.0, F.dtype(dx))
-                num = F.scalar_cast(dev_num, F.dtype(dx))
-                dx = mul(dx, cast(F.scalar_to_tensor(float_one/num), F.dtype(dx)))
+                dx = mul(dx, cast(F.scalar_to_tensor(dev_num_r), F.dtype(dx)))
             else:
                 indices = all_gather(dout.indices)
                 grad = all_gather(dout.values)
-                float_one = F.scalar_cast(1.0, F.dtype(grad))
-                num = F.scalar_cast(dev_num, F.dtype(grad))
-                grad = mul(grad, cast(F.scalar_to_tensor(float_one/num), F.dtype(grad)))
+                grad = mul(grad, cast(F.scalar_to_tensor(dev_num_r), F.dtype(grad)))
                 dx = RowTensorInner(indices, grad, dout.dense_shape)
         else:
             if issubclass_(F.typeof(dout), mstype.tensor):
@@ -504,6 +515,7 @@ def get_bprop_mirror_mini_step_operator(self):
     group = self.group
     dev_num = self.dev_num
     mean_flag = self.mean_flag
+    dev_num_r = 1.0 / dev_num
 
     all_reduce = AllReduce(group=group)
     mul = P.Mul()
@@ -529,9 +541,7 @@ def get_bprop_mirror_mini_step_operator(self):
                     dx = real_grad
                 else:
                     dx = dout
-                float_one = F.scalar_cast(1.0, F.dtype(dx))
-                num = F.scalar_cast(dev_num, F.dtype(dx))
-                dx = mul(dx, cast(F.scalar_to_tensor(float_one/num), F.dtype(dx)))
+                dx = mul(dx, cast(F.scalar_to_tensor(dev_num_r), F.dtype(dx)))
             else:
                 dx = zeros_like(x)  # The grad accumulation do not support row tensor now
         else: