mindspore 1.10.0__cp37-none-any.whl → 2.0.0rc1__cp37-none-any.whl

mindspore/ops/_grad/grad_array_ops.py

@@ -17,19 +17,23 @@
 
 import numpy as np
 import mindspore as ms
-from mindspore.ops import composite as C
+from mindspore import Tensor
 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
 from mindspore.ops.composite.multitype_ops.zeros_like_impl import zeros_like
 from mindspore.ops.functional import broadcast_gradient_args
 from mindspore.ops import functional as F
-from mindspore.ops._grad.grad_base import bprop_getters
+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.tensor import RowTensor
-from mindspore.ops._utils.utils import range_op, get_1d_shape, generate_shape_index, is_shape_unknown
-from .._grad.grad_base import dyn_rank, convert_to_tensor, dyn_invert_permutation, dyn_size, dyn_ones, dyn_fill
+from mindspore.common.sparse_tensor import RowTensorInner
+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
 
 reduce_sum = P.ReduceSum()
 unsorted_segment_sum = P.UnsortedSegmentSum()
@@ -40,7 +44,7 @@ reshape = P.Reshape()
 size_op = P.Size()
 invert_permutation = P.InvertPermutation()
 logical_and = P.LogicalAnd()
-is_sub_class = P.IsSubClass()
+is_sub_class = IsSubClass()
 
 
 @bprop_getters.register(P.Fill)
@@ -93,46 +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 RowTensor(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"""
@@ -152,6 +116,7 @@ def get_bprop_dynamicshape(self):
 
     return bprop
 
+
 @bprop_getters.register(P.TensorShape)
 def get_bprop_tensorshape(self):
     """Generate bprop for TensorShape"""
@@ -191,7 +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)
 
@@ -204,7 +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)
 
@@ -217,7 +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),)
 
@@ -230,13 +195,16 @@ def get_bprop_flatten(self):
     flatten_grad = P.Reshape()
 
     def bprop(x, out, dout):
-        dx = flatten_grad(dout, shape_op(x))
+        shape_x = shape_op(x)
+        if F.is_sequence_value_unknown(shape_x):
+            shape_x = dyn_shape_op(x)
+        dx = flatten_grad(dout, shape_x)
         return (dx,)
 
     return bprop
 
 
-@constexpr
+@_primexpr
 def _tile_shape(multiples, shapex):
     """Calculate [1,2], [3, 4] -> [1,3,2,4]."""
     len_muli = len(multiples)
@@ -268,40 +236,45 @@ def _tile_shape(multiples, shapex):
 @bprop_getters.register(P.Tile)
 def get_bprop_tile(self):
     """Generate bprop for Tile"""
-    tuple_to_array = P.TupleToArray()
     cast = P.Cast()
-    stack_op = P.Stack(1)
-    ones = P.Ones()
     concat = P.Concat()
+    stridedslice = P.StridedSlice()
+
+    def get_reduce_axis(r_shape):
+        """
+        reshape grad to r_shape, and reduce along all even dimensions to get the result with input_shape
+        For example:
+        input_shape = [20, 30, 40]
+        multiples = [2, 3, 4]
+        r_shape = [2, 20, 3, 30, 4, 40]
+        axis = [0, 2, 4]
+        """
+        rankr = dyn_shape_op(r_shape)[0]
+        tmp = range_op(0, 20, 2, mstype.int64)
+        return stridedslice(tmp, (0,), F.expand_dims(rankr // 2, 0), (1,))
 
     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 shapex or multiples not tuple, it should be dynamic shape.
         if isinstance(multiples, tuple) and isinstance(shapex, tuple):
             r_shape = _tile_shape(multiples, shapex)
+            # 0 represents the start index, and 2 represents the step
+            axis = F.make_range(0, len(r_shape), 2)
         else:
-            if isinstance(multiples, tuple):
-                multiples = tuple_to_array(multiples)
-            multiples = cast(multiples, mstype.int64)
-            len_multi = size_op(multiples)
-            rank = len(shapex)
-            if isinstance(shapex, tuple):
-                shape_tensor = cast(tuple_to_array(shapex), mstype.int64)
-            else:
-                shape_tensor = shapex
-            if len_multi > rank:
-                one_tensor = ones((len_multi - rank,), mstype.int64)
-                shape_tensor = concat((one_tensor, shape_tensor))
-            elif len_multi < rank:
-                one_tensor = ones((rank - len_multi,), mstype.int64)
-                multiples = concat((one_tensor, multiples))
-            tile_shape = stack_op((multiples, shape_tensor))
+            shapex = dyn_shape_op(x)
+            shapey = create_tensor_by_element(multiples)
+            rankx = dyn_rank(x)
+            ranky = dyn_shape_op(shapey)[0]
+            offset = F.expand_dims(ranky - rankx + 1, 0)
+            shape_x = concat((dyn_ones(offset, mstype.int64), shapex))
+            shape_x = shape_x[1:]
+            shapey = concat((P.Ones()((1,), mstype.int64), shapey))
+            shapey = shapey[1:]
+            tile_shape = P.Stack(1)((shapey, shape_x))
             r_shape = P.Reshape()(tile_shape, (-1,))
+            axis = get_reduce_axis(r_shape)
 
-        # 0 represents the start index, and 2 represents the step
-        axis = F.make_range(0, len(r_shape), 2)
         dout_reshaped = P.Reshape()(dout, r_shape)
         dout_origin_dtype = dout_reshaped.dtype
         # Currently, for Ascend and GPU, the reduce_sum's input does not support int16, int32 and int64.
@@ -325,6 +298,8 @@ def get_bprop_embedding_lookup(self):
 
     def bprop_sparse(x, indices, offset, out, dout):
         x_shp = shape_op(x)
+        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)
         if indices_size > 0:
@@ -337,61 +312,43 @@ def get_bprop_embedding_lookup(self):
         actual_dout_shape_changed = new_indices_shape_changed + x_shp_tail
         # Reshape the 'actual_dout' on device
         actual_dout = reshape_op(dout, actual_dout_shape_changed)
-        return RowTensor(new_indices, actual_dout, x_shp), zeros_like(indices), zeros_like(offset)
+        return RowTensorInner(new_indices, actual_dout, x_shp), zeros_like(indices), zeros_like(offset)
 
     return bprop_sparse
 
 
-@constexpr
+@_primexpr
 def make_begin(shp):
     """Creates a tuple with zero according to the shape."""
     begin = tuple([0 for _ in shp])
     return begin
 
 
+def make_dynamic_begin(shp):
+    """Creates a tuple with zero according to the shape."""
+    begin = zeros_like(shp)
+    return begin
+
+
 @bprop_getters.register(P.Padding)
 def get_bprop_padding(self):
     """Grad definition for `Padding` operation."""
 
     def bprop(x, out, dout):
         shp = shape_op(x)
-        begin = make_begin(shp)
+        begin = ()
+        if F.is_sequence_value_unknown(shp):
+            shp = dyn_shape_op(x)
+            begin = make_dynamic_begin(shp)
+        else:
+            begin = make_begin(shp)
         dx = P.Slice()(dout, begin, shp)
         return (dx,)
 
     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
@@ -444,12 +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"""
@@ -461,17 +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:
@@ -480,12 +431,99 @@ def _regenerate_output_shape(x_shp, ind_shp, axis):
     return out_shape
 
 
+def _dyn_regenerate_output_shape(x_shp, ind_shp, axis):
+    """Get reshape new_shape"""
+    rank = dyn_shape_op(x_shp)[0]
+    if axis < 0:
+        axis += rank
+    out_shape = P.Concat(0)((x_shp[:axis], ind_shp, x_shp[axis + 1:]))
+    return out_shape
+
+
+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))
+    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 = F.hstack((index[:batch_dims], perm_part1, index[batch_dims:axis], index[ind_end:]))
+    return perm
+
+
+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 = 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):
     """Generate bprop for GatherV2"""
 
+    def _dyn_bprop_gather_v2(x, indices, axis, dout):
+        """dyn shape bprop for GatherV2"""
+        orig_indices = indices
+        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)
+        if F.rank(indices) == 0:
+            indices = P.ExpandDims()(indices, -1)
+            out_shp = _dyn_regenerate_output_shape(x_shp, ind_shp, axis)
+            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, batch_dims)
+        values_transpose = transpose(dout, perm_1)
+        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 (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:
             dout = P.ExpandDims()(dout, -1)
@@ -499,15 +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)
 
@@ -519,17 +561,19 @@ def get_bprop_gather_d(self):
     """Generate bprop for GatherD"""
 
     def bprop(x, dim, index, out, dout):
-        dx = G.GatherDGradV2(dim)(x, index, dout)
+        dx = G.GatherDGradV2()(x, dim, index, dout)
         return dx, zeros_like(dim), zeros_like(index)
 
     return bprop
 
+
 @bprop_getters.register(G.GatherDGrad)
 def get_bprop_gather_d_grad(self):
     """Generate bprop for GatherDGrad"""
     op = P.Gather()
     dim = self.dim
     x_shp = self.out_shape
+
     def bprop(index, x, out, dout):
         index_shp = shape_op(index)
         dim_before_axis = 1
@@ -538,7 +582,7 @@ def get_bprop_gather_d_grad(self):
         dim_at_axis_index = index_shp[dim]
         dim_at_axis_output = x_shp[dim]
         dim_after_axis = 1
-        for i in range(dim+1, len(x_shp)):
+        for i in range(dim + 1, len(x_shp)):
             dim_after_axis *= x_shp[i]
         element = dim_before_axis * dim_at_axis_index * dim_after_axis
         id_ = range_op(0, element, 1, index.dtype)
@@ -547,7 +591,7 @@ def get_bprop_gather_d_grad(self):
         j = P.Cast()(index < 0, index.dtype)
         j_read = dim_at_axis_index * j + index
         j_read = P.Reshape()(j_read, (-1,))
-        read_id = i*dim_at_axis_output*dim_after_axis + j_read * dim_after_axis + k
+        read_id = i * dim_at_axis_output * dim_after_axis + j_read * dim_after_axis + k
         dout = P.Reshape()(dout, (-1,))
         dx = op(dout, read_id, 0)
         dx = P.Reshape()(dx, shape_op(x))
@@ -561,6 +605,7 @@ def get_bprop_gather_d_grad_v2(self):
     """Generate bprop for GatherDGradV2"""
     op = P.Gather()
     dim = self.dim
+
     def bprop(index, x, out, dout):
         index_shp = shape_op(index)
         dim_before_axis = 1
@@ -570,7 +615,7 @@ def get_bprop_gather_d_grad_v2(self):
         dim_at_axis_index = index_shp[dim]
         dim_at_axis_output = x_shp[dim]
         dim_after_axis = 1
-        for i in range(dim+1, len(x_shp)):
+        for i in range(dim + 1, len(x_shp)):
             dim_after_axis *= x_shp[i]
         element = dim_before_axis * dim_at_axis_index * dim_after_axis
         id_ = range_op(0, element, 1, index.dtype)
@@ -579,7 +624,7 @@ def get_bprop_gather_d_grad_v2(self):
         j = P.Cast()(index < 0, index.dtype)
         j_read = dim_at_axis_index * j + index
         j_read = P.Reshape()(j_read, (-1,))
-        read_id = i*dim_at_axis_output*dim_after_axis + j_read * dim_after_axis + k
+        read_id = i * dim_at_axis_output * dim_after_axis + j_read * dim_after_axis + k
         dout = P.Reshape()(dout, (-1,))
         dx = op(dout, read_id, 0)
         dx = P.Reshape()(dx, shape_op(x))
@@ -587,6 +632,7 @@ def get_bprop_gather_d_grad_v2(self):
 
     return bprop
 
+
 @bprop_getters.register(P.SparseGatherV2)
 def get_bprop_sparse_gather_v2(self):
     """Generate bprop for SparseGatherV2"""
@@ -602,7 +648,7 @@ def get_bprop_sparse_gather_v2(self):
             values_shape = indices_size + x_tail_shp
             values = reshape(dout, values_shape)
             indices_new = reshape(indices, indices_size)
-            return RowTensor(indices_new, values, x_shp), zeros_like(indices), zeros_like(axis)
+            return RowTensorInner(indices_new, values, x_shp), zeros_like(indices), zeros_like(axis)
         if F.rank(dout) == 0:
             dout = P.ExpandDims()(dout, -1)
         if F.rank(indices) == 0:
@@ -715,7 +761,7 @@ def get_bprop_stack(self):
     axis = self.axis
 
     def bprop(x, out, dout):
-        stack_grad = P.Unstack(axis)
+        stack_grad = P.Unstack(num=len(x), axis=axis)
         out = stack_grad(dout)
         if is_sub_class(F.typeof(x), ms.list_):
             ret = []
@@ -764,7 +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)
@@ -780,9 +826,10 @@ def get_bprop_strided_slice_grad(self):
                                    ellipsis_mask=self.ellipsis_mask,
                                    new_axis_mask=self.new_axis_mask,
                                    shrink_axis_mask=self.shrink_axis_mask)
+
     def bprop(dy, shapex, begin, end, strides, out, dout):
         return strided_slice(dout, begin, end, strides), zeros_like(shapex), zeros_like(begin), zeros_like(end), \
-            zeros_like(strides)
+               zeros_like(strides)
 
     return bprop
 
@@ -835,7 +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)
@@ -853,7 +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)
 
@@ -1029,18 +1076,14 @@ def _gather_drop_negatives(params,
     select = P.Select()
 
     if zero_clipped_indices is None:
-        if is_shape_unknown(shape_op(ids)):
-            zero_ids = dyn_fill(ids.dtype, dyn_shape_op(ids), 0)
-        else:
-            zero_ids = zeros_like(ids)
-        zero_clipped_indices = maximum(ids, zero_ids)
+        zero_clipped_indices = maximum(ids, zeros_like(ids))
     gathered = gather(params, zero_clipped_indices, 0)
     zero_slice = zeros_like(gathered)
     if is_positive is None:
         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)
@@ -1052,7 +1095,6 @@ def _gather_drop_negatives(params,
                 is_positive_shape = P.Concat(-1)((is_positive_shape, padded_shape))
             is_positive = reshape(is_positive, is_positive_shape)
             is_positive = logical_and(is_positive, F.cast(fill_gathered, mstype.bool_))
-            zero_slice = dyn_fill(gathered.dtype, gathered_shape, 0)
         else:
             broadcastable_shape = is_positive_shape
             for _ in range(rank(gathered) - rank(is_positive)):
@@ -1087,13 +1129,7 @@ def get_bprop_unsorted_segment_sum(self):
     """Generate bprop for UnsortedSegmentSum"""
 
     def bprop(x, segment_ids, num_segments, out, dout):
-        segment_shape = shape_op(segment_ids)
-        if is_shape_unknown(segment_shape):
-            segment_shape = dyn_shape_op(segment_ids)
-            zeros_segment = dyn_fill(segment_ids.dtype, segment_shape, 0)
-        else:
-            zeros_segment = zeros_like(segment_ids)
-        return _gather_drop_negatives(dout, segment_ids, None, None)[0], zeros_segment, \
+        return _gather_drop_negatives(dout, segment_ids, None, None)[0], zeros_like(segment_ids), \
                zeros_like(num_segments)
 
     return bprop
@@ -1132,6 +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:
@@ -1221,12 +1259,31 @@ def get_bprop_broadcast_to(self):
         x_shape = shape_op(x)
         dout_shape = shape_op(dout)
         broadcast_shape = shape_op(out)
-
-        if x_shape == 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,)
-        _, reduction_axes = broadcast_gradient_args(broadcast_shape, x_shape)
-        reduced_grad = reduce_keep_dim(dout, reduction_axes)
-        dx = reshape(reduced_grad, x_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)
+            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)
+            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,)
 
     return bprop