mindspore 2.2.14__cp39-cp39-win_amd64.whl → 2.4.0__cp39-cp39-win_amd64.whl

mindspore/ops/operations/_grad_ops.py

@@ -1,4 +1,4 @@
-# Copyright 2020-2022 Huawei Technologies Co., Ltd
+# Copyright 2020-2024 Huawei Technologies Co., Ltd
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,6 +14,7 @@
 # ============================================================================
 
 """Operators for gradients."""
+# pylint: disable=unused-import
 from __future__ import absolute_import
 
 from __future__ import division
@@ -27,6 +28,15 @@ from mindspore import _checkparam as validator
 from mindspore.common import dtype as mstype
 from mindspore.communication.management import GlobalComm
 from mindspore.common._utils import is_shape_unknown, is_dim_unknown
+from ..auto_generate import (AbsGrad, ACosGrad, LogitGrad, AcoshGrad, AsinGrad, AsinhGrad, ReciprocalGrad, RsqrtGrad,
+                             SqrtGrad, BatchNormGrad, BatchNormGradGrad, BiasAddGrad, GeLUGrad, FastGeLUGrad,
+                             AvgPoolGrad, MinimumGrad, LogSoftmaxGrad, PReLUGrad, ReluGrad, ReLU6Grad, EluGrad,
+                             GatherDGradV2, ResizeBilinearGrad, ResizeLinear1DGrad, ResizeNearestNeighborV2Grad,
+                             SigmoidGrad, HSwishGrad, NLLLossGrad, AtanGrad, GridSampler3DGrad, GridSampler2DGrad,
+                             ResizeBicubicGrad, HSigmoidGrad, CholeskyGrad, ResizeNearestNeighborGrad, LayerNormGrad,
+                             HShrinkGrad, LayerNormGradGrad, SiLUGrad, MaximumGrad, MaximumGradGrad, RmsNormGrad,
+                             FlashAttentionScoreGrad, UpsampleTrilinear3DGrad, UpsampleNearest3DGrad, MaskedSelectGrad,
+                             BinaryCrossEntropyGrad, SoftShrinkGrad, SeluGrad)
 
 
 class SparseFillEmptyRowsGrad(Primitive):
@@ -39,92 +49,6 @@ class SparseFillEmptyRowsGrad(Primitive):
                                 outputs=['y_values', 'y_default_value'])
 
 
-class AbsGrad(PrimitiveWithInfer):
-    """Computes gradients for abs operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize AbsGrad"""
-
-
-class ACosGrad(Primitive):
-    """
-    Computes ACosGrad of input element-wise.
-
-    Returns:
-        Tensor, has the same type as input.
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize ACosGrad"""
-        self.init_prim_io_names(inputs=['y', 'dy'], outputs=['z'])
-
-
-class LogitGrad(Primitive):
-    """
-    Computes LogitGrad of input element-wise.
-
-    Returns:
-        Tensor, has the same type as input.
-    """
-    @prim_attr_register
-    def __init__(self, eps=-1.0):
-        """Initialize Exp"""
-        self.init_prim_io_names(inputs=['grad', 'input'], outputs=['dx'])
-        validator.check_value_type("eps", eps, [float], self.name)
-        self.add_prim_attr('eps', eps)
-
-
-class AcoshGrad(Primitive):
-    """Performs grad of Acosh operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize AcoshGrad"""
-        self.init_prim_io_names(inputs=['y', 'dy'], outputs=['z'])
-
-
-class AsinGrad(Primitive):
-    """
-    Computes AsinGrad of input element-wise.
-
-    Returns:
-        Tensor, has the same type as input.
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize AsinGrad"""
-        self.init_prim_io_names(inputs=['y', 'dy'], outputs=['z'])
-
-
-class AsinhGrad(Primitive):
-    """Performs grad of Asinh operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize AsinhGrad"""
-        self.init_prim_io_names(inputs=['y', 'dy'], outputs=['z'])
-
-
-class ReciprocalGrad(Primitive):
-    """Performs grad of Reciprocal operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize ReciprocalGrad"""
-        self.init_prim_io_names(inputs=['y', 'dy'], outputs=['z'])
-
-
-class RsqrtGrad(Primitive):
-    """Performs grad of Rsqrt operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize RsqrtGrad"""
-
-
 class ScaleAndTranslateGrad(Primitive):
     """Performs grad of ScaleAndTranslate operation."""
 
@@ -137,39 +61,15 @@ class ScaleAndTranslateGrad(Primitive):
         validator.check_value_type("antialias", antialias, [bool], self.name)
 
 
-class SoftmaxGrad(ReciprocalGrad):
+class SoftmaxGrad(Primitive):
     """Performs grad of Softmax operation."""
 
-
-class SqrtGrad(Primitive):
-    """Performs grad of Sqrt operation."""
-
     @prim_attr_register
     def __init__(self):
-        """Initialize SqrtGrad"""
+        """Initialize SoftmaxGrad"""
         self.init_prim_io_names(inputs=['y', 'dy'], outputs=['z'])
 
 
-class BatchNormGrad(Primitive):
-    """Performs grad of BatchNorm operation."""
-
-    @prim_attr_register
-    def __init__(self, is_training=False, epsilon=1e-5, data_format='NCHW'):
-        self.is_training = validator.check_value_type('is_training', is_training, (bool,), self.name)
-        self.epsilon = validator.check_float_range(epsilon, 0, 1, validator.INC_RIGHT, 'epsilon', self.name)
-        self.data_format = validator.check_string(data_format, ['NCHW', 'NHWC'], 'format', self.name)
-
-
-class BatchNormGradGrad(Primitive):
-    """Performs grad of BatchNormGrad operation."""
-
-    @prim_attr_register
-    def __init__(self, is_training=False, epsilon=1e-5, data_format='NCHW'):
-        self.is_training = validator.check_value_type('is_training', is_training, (bool,), self.name)
-        self.epsilon = validator.check_float_range(epsilon, 0, 1, validator.INC_RIGHT, 'epsilon', self.name)
-        self.data_format = validator.check_string(data_format, ['NCHW', 'NHWC'], 'format', self.name)
-
-
 class SyncBatchNormGrad(Primitive):
     """Performs grad of SyncBatchNorm operation."""
 
@@ -181,18 +81,6 @@ class SyncBatchNormGrad(Primitive):
         validator.check_int(device_num, 2, validator.GE, "device_num", self.name)
 
 
-class BiasAddGrad(Primitive):
-    """Computes gradients of BiasAdd."""
-
-    @prim_attr_register
-    def __init__(self, data_format="NCHW"):
-        self.init_prim_io_names(inputs=['dout'], outputs=['output'])
-        self.format = validator.check_string(data_format, ['NCHW', 'NHWC', 'NCDHW'], 'format', self.name)
-        if self.format == "NCDHW":
-            self.format = "NCHW"
-        self.add_prim_attr('data_format', self.format)
-
-
 class KLDivLossGrad(Primitive):
     """Computes gradients for `KLDivLoss` operation."""
 
@@ -210,14 +98,6 @@ class KLDivLossGrad(Primitive):
         self.reduction = validator.check_string(reduction, support_mode, 'reduction', self.name)
 
 
-class BinaryCrossEntropyGrad(Primitive):
-    """Computes gradients for `BinaryCrossEntropy` operation."""
-
-    @prim_attr_register
-    def __init__(self, reduction='mean'):
-        self.reduction = validator.check_string(reduction, ['none', 'mean', 'sum'], 'reduction', self.name)
-
-
 class LuUnpackGrad(Primitive):
     """Computes gradients for `LuUnpack` operation."""
 
@@ -713,22 +593,6 @@ class NeighborExchangeV2Grad(PrimitiveWithInfer):
                 'value': None}
 
 
-class GeLUGrad(Primitive):
-    """Gradients of GeLU operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        self.init_prim_io_names(inputs=['dy', 'x', 'y'], outputs=['z'])
-
-
-class FastGeLUGrad(Primitive):
-    """Gradients of FastGeLU operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """init FastGeLUGrad"""
-
-
 class _PoolGrad(PrimitiveWithInfer):
     """Gradients of the max/avg pool operation."""
 
@@ -813,20 +677,6 @@ class AvgPoolGradGe(_PoolGrad):
         return out
 
 
-class AvgPoolGrad(_PoolGrad):
-    """Gradients of the avg pool operation."""
-
-    @prim_attr_register
-    def __init__(self, kernel_size=1, strides=1, pad_mode="VALID", data_format="NCHW"):
-        super(AvgPoolGrad, self).__init__(kernel_size, strides, pad_mode, data_format)
-
-    def infer_shape(self, x1_shape, x2_shape, grad_shape):
-        return x1_shape
-
-    def infer_dtype(self, x1_dtype, x2_dtype, grad_dtype):
-        return x1_dtype
-
-
 class AvgPoolGradV1(Primitive):
     """Gradients of the AvgPoolV1 operation."""
 
@@ -1192,25 +1042,6 @@ class MaxPool3DGradGrad(PrimitiveWithInfer):
         return x_dtype
 
 
-class MaximumGrad(Primitive):
-    """Grad for maximum."""
-
-    @prim_attr_register
-    def __init__(self, grad_x=True, grad_y=True):
-        """Initialize MaximumGrad"""
-        self.init_prim_io_names(inputs=['x1', 'x2', 'grads'], outputs=['y1', 'y2'])
-
-
-class MaximumGradGrad(Primitive):
-    """Grad for maximum grad."""
-
-    @prim_attr_register
-    def __init__(self, grad_x=True, grad_y=True):
-        """Initialize MaximumGradGrad"""
-        super().__init__("MaximumGradGrad")
-        self.init_prim_io_names(inputs=['x1', 'x2', 'dy1', 'dy2'], outputs=['sopd_x1', 'sopd_x2', 'sopd_grad'])
-
-
 class MaxPoolGradWithArgmax(Primitive):
     """Computes the gradients of MaxPoolWithArgmax."""
     @prim_attr_register
@@ -1359,15 +1190,6 @@ class MaxPoolGradGradWithArgmax(_PoolGrad):
         return grad_dtype
 
 
-class MinimumGrad(Primitive):
-    """Grad for minimum."""
-
-    @prim_attr_register
-    def __init__(self, grad_x=True, grad_y=True):
-        """Initialize MinimumGrad"""
-        self.init_prim_io_names(inputs=['x1', 'x2', 'grads'], outputs=['y1', 'y2'])
-
-
 class MinimumGradGrad(Primitive):
     """Grad for minimum_grad."""
     @prim_attr_register
@@ -1406,79 +1228,6 @@ class L2NormalizeGrad(Primitive):
             raise TypeError("The length of axis must be 1, later will support multiple axis!")
 
 
-class LayerNormGrad(Primitive):
-    """
-    Applies the layer Normalization to the input array.
-
-    This operator will calculate the input gradients of layernorm.
-
-    Args:
-        begin_norm_axis (int): The begin axis for the input to apply layernorm. Default: 1.
-        begin_params_axis (int): The begin axis for the parameter input to apply layernorm. Default: 1.
-
-    Returns:
-        tuple[int], tuple of 3 values (the gradients of layernorm input,  gamma, beta).
-    """
-
-    @prim_attr_register
-    def __init__(self, begin_norm_axis=1, begin_params_axis=1):
-        """init"""
-        self.begin_norm_axis = validator.check_value_type('begin_norm_axis', begin_norm_axis, [int], self.name)
-        self.begin_params_axis = validator.check_value_type('begin_params_axis', begin_params_axis, [int], self.name)
-
-
-class LayerNormGradGrad(Primitive):
-    """
-    Gets the gradient of LayerNormGrad operation.
-
-    Args:
-        begin_norm_axis (int): The begin axis for the input to apply layernorm. Default: 1.
-        begin_params_axis (int): The begin axis for the parameter input to apply layernorm. Default: 1.
-
-    Inputs:
-        - **x** (Tensor) - The input tensor to be normalized, float32 or float16.
-        - **dy** (Tensor) - The gradient of LayerNorm's output y, float32 or float16.
-        - **variance** (Tensor) - The variance of x, float32 or float16.
-        - **mean** (Tensor) - The mean of x, float32 or float16.
-        - **gamma** (Tensor) - The original value of weight gamma initialized in LayerNorm, float32 or float16.
-          Default: 'ones'.
-        - **d_dx** (Tensor) - The gradient of dx, where dx is the gradient of LayerNorm's input x, float32 or float16.
-        - **d_dg** (Tensor) - The gradient of dg, where dg is the gradient of LayerNorm's weight gamma,
-          float32 or float16.
-        - **d_db** (Tensor) - The gradient of db, where db is the gradient of LayerNorm's weight beta,
-          float32 or float16.
-
-    Returns:
-        Tuple[Tensor], tuple of 3 Tensors (the gradients of layernormgrad x, dy, gamma).
-
-    Raises:
-        TypeError: If the 8 inputs don't have the same dtype.
-        ValueError: If x, dy, d_dx don't have the same shape.
-        ValueError: If variance, mean don't have the same shape.
-        ValueError: If gamma, d_dg, d_db don't have the same shape.
-
-    Supported Platforms:
-        ``Ascend`` ``GPU`` ``CPU``
-    """
-
-    @prim_attr_register
-    def __init__(self, begin_norm_axis=1, begin_params_axis=1):
-        """init"""
-        self.begin_norm_axis = validator.check_value_type('begin_norm_axis', begin_norm_axis, [int], self.name)
-        self.begin_params_axis = validator.check_value_type('begin_params_axis', begin_params_axis, [int], self.name)
-        self.init_prim_io_names(inputs=['x', 'dy', 'variance', 'mean', 'gamma', 'd_dx', 'd_dg', 'd_db'],
-                                outputs=['sopd_x', 'sopd_dy', 'sopd_gamma'])
-
-
-class LogSoftmaxGrad(Primitive):
-    """Computes gradient for the Log Softmax activation."""
-
-    @prim_attr_register
-    def __init__(self, axis=-1):
-        """Initialize LogSoftmaxGrad"""
-        validator.check_value_type("axis", axis, [int], self.name)
-
-
 class LSTMGradData(Primitive):
     """Computes the data gradients of LSTM."""
 
@@ -1741,27 +1490,6 @@ class DynamicGRUV2Grad(Primitive):
                                 ])
 
 
-class PReLUGrad(Primitive):
-    r"""
-    Gradients of PReLU operation.
-
-    Note:
-        1-dimensional input_x is not supported.
-
-    Inputs:
-        - **y_backprop** (Tensor) - Representing the backprop of the next layer.
-        - **input_x** (Tensor) - Must be the input `input_x` of forward operator PRelu.
-        - **weight** (Tensor) - Float Tensor, w > 0, must be the input `weight` of forward operator PRelu.
-
-    Outputs:
-        Tensor, with the same type as `input_x`.
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        pass
-
-
 class RandomGammaGrad(Primitive):
     r"""
     Computes the derivative of a random sample of Gamma with respect to alpha.:
@@ -1800,180 +1528,6 @@ class RandomGammaGrad(Primitive):
         self.add_prim_attr("side_effect_hidden", True)
 
 
-class ReluGrad(Primitive):
-    """Performs grad of Relu operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize ReluGrad"""
-        self.init_prim_io_names(inputs=['y_backprop', 'x'], outputs=['output'])
-
-
-class SiLUGrad(Primitive):
-    """Performs grad of SiLU operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize SiLUGrad"""
-        self.init_prim_io_names(inputs=['dout', 'out'], outputs=['output'])
-
-
-class ReLU6Grad(Primitive):
-    """Performs grad of ReLU6 operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        self.init_prim_io_names(inputs=['y_grad', 'x'], outputs=['output'])
-
-
-class ReluGradV2(Primitive):
-    """Performs grad of ReLUV2 operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        self.init_prim_io_names(inputs=['gradients', 'mask'], outputs=['output'])
-
-
-class EluGrad(Primitive):
-    """Performs grad of Elu operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize EluGrad"""
-        self.init_prim_io_names(inputs=['y_backprop', 'x'], outputs=['output'])
-
-
-class GatherDGrad(Primitive):
-    """Performs grad of GatherD operation."""
-
-    @prim_attr_register
-    def __init__(self, dim=0, shape=None):
-        """Initialize GatherDGrad"""
-        validator.check_is_int(dim, int)
-        self.add_prim_attr("dim", dim)
-        self.dim = dim
-        self.out_shape = shape
-        self.init_prim_io_names(inputs=['index', 'grad'], outputs=['output'])
-
-
-class GatherDGradV2(Primitive):
-    """Performs grad of GatherD operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize GatherDGradV2"""
-        self.init_prim_io_names(inputs=['x', 'dim', 'index', 'grad'], outputs=['output'])
-
-
-class ResizeBilinearGrad(Primitive):
-    """Performs grad of ResizeBilinear operation."""
-
-    @prim_attr_register
-    def __init__(self, align_corners=False, half_pixel_centers=False):
-        """init"""
-        validator.check_value_type("align_corners", align_corners, [bool], self.name)
-        validator.check_value_type("half_pixel_centers", half_pixel_centers, [bool], self.name)
-        self.align_corners = validator.check_value_type("align_corners", align_corners, [bool], self.name)
-        self.half_pixel_centers = validator.check_value_type("half_pixel_centers",
-                                                             half_pixel_centers, [bool], self.name)
-        self.init_prim_io_names(inputs=['grads', 'original_image'], outputs=['y'])
-        if half_pixel_centers and align_corners:
-            raise ValueError(f"If half_pixel_centers is True, align_corners must be False, but got {align_corners}")
-
-
-class ResizeNearestNeighborGrad(Primitive):
-    """
-    Compute gradient of `ResizeNearestNeighbor` operator.
-
-    Note:
-        The shape of input parameter `size` must be (height, width).
-
-    Args:
-        align_corners (bool): Whether the centers of the 4 corner pixels of the input
-            and output tensors are aligned. Default: ``False``.
-    """
-
-    @prim_attr_register
-    def __init__(self, align_corners=False):
-        """Initialize ResizeNearestNeighborGrad"""
-        self.init_prim_io_names(inputs=['grads', 'size'], outputs=['y'])
-
-
-class ResizeLinear1DGrad(Primitive):
-    """
-    Compute gradient of `ResizeLinear1D` operator.
-
-    .. warning::
-        This is an experimental API that is subject to change.
-
-    Args:
-        coordinate_transformation_mode (string): Default is 'align_corners'. Describes how to transform the coordinate
-            in the resized tensor to the coordinate in the original tensor. Other optional: 'half_pixel'.
-    """
-
-    @prim_attr_register
-    def __init__(self, coordinate_transformation_mode="align_corners"):
-        """Initialize ResizeLinear1DGrad"""
-        self.init_prim_io_names(
-            inputs=['grads', 'input_x'], outputs=['y'])
-        validator.check_value_type(
-            "coordinate_transformation_mode", coordinate_transformation_mode, [str], self.name)
-        validator.check_string(coordinate_transformation_mode, ["align_corners", "half_pixel"],
-                               "coordinate_transformation_mode", self.name)
-
-
-class ResizeNearestNeighborV2Grad(Primitive):
-    """
-    Compute gradient of `ResizeNearestNeighborV2` operator.
-
-    Args:
-        align_corners (bool): Whether the centers of the 4 corner pixels of the input
-            and output tensors are aligned. Default: ``False``.
-        half_pixel_centers (bool): Default: ``False``.
-    """
-
-    @prim_attr_register
-    def __init__(self, align_corners=False, half_pixel_centers=False):
-        """Initialize ResizeNearestNeighborV2Grad"""
-        self.init_prim_io_names(inputs=['grads', 'size'], outputs=['y'])
-        validator.check_value_type('align_corners', align_corners, [bool], self.name)
-        validator.check_value_type('half_pixel_centers', half_pixel_centers, [bool], self.name)
-
-
-class UpsampleNearest3DGrad(Primitive):
-    """
-    Upsample the 3-D gradient data  with the nearest neighbor interpolation algorithm.
-
-    Note:
-        Only one of 'scales' and 'output_size' can be specified, and it is an error if both are specified.
-
-    Inputs:
-        - **dy** (Tensor) - Tensor of shape [N, C, D, H, W], Must be one of the following types:
-            float16, float32, float64.
-        - **input_size** (listInt): An required listInt, which contain 5 elements:
-            [min_batch, channels, depth, height, width].
-            Must: input_size[0] == dy_tensor_size[0], input_size[1] == dy_tensor_size[1].
-        - **output_size** (listInt): An optional listInt. Default: ``None``.
-            It contains 3 elements: depth, height, width, whose elements should be the same as `dy`.
-            Must:
-            dy_tensor_size[2] == floor(input_size[2] * scales[0]) == output_size[0],
-            dy_tensor_size[3] == floor(input_size[3] * scales[1]) == output_size[1],
-            dy_tensor_size[4] == floor(input_size[4] * scales[2]) == output_size[2].
-        - **scales** (listFloat): An optional listFloat. Default: ``None``.
-            The scale array along each dimension, contain 3 elements: scale_depth, scale_height, scale_width.
-            The number of elements of 'scales' should be the same as the rank of `dy`.
-
-    Outputs:
-        - **dx**- (Tensor) - A 5-D tensor. Has the same type as `dy`, shape depends on `input_size`.
-    """
-    @prim_attr_register
-    def __init__(self):
-        """Initialize UpsampleNearest3DGrad."""
-        self.init_prim_io_names(
-            inputs=['dy', 'input_size', 'output_size', 'scales'],
-            outputs=['dx'])
-
-
 class ROIAlignGrad(Primitive):
     """
     ROIAlignGrad operator.
@@ -2034,15 +1588,6 @@ class PsROIPoolingGrad(PrimitiveWithInfer):
         return ydiff_type
 
 
-class SigmoidGrad(Primitive):
-    """Gets the gradient of Sigmoid operation."""
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize SigmoidGrad"""
-        self.init_prim_io_names(inputs=['y', 'dy'], outputs=['output'])
-
-
 class _ActivationGrad(PrimitiveWithInfer):
     """_ActivationGrad base class."""
 
@@ -2060,14 +1605,6 @@ class _ActivationGrad(PrimitiveWithInfer):
         return x_dtype
 
 
-class HSwishGrad(Primitive):
-    """Gets the gradient of HSwish operation."""
-    @prim_attr_register
-    def __init__(self):
-        """Initialize HSwishGrad"""
-        self.init_prim_io_names(inputs=['y_grad', 'x'], outputs=['output'])
-
-
 class SigmoidCrossEntropyWithLogitsGrad(Primitive):
     """Computes the gradients of `SigmoidCrossEntropyWithLogits`."""
 
@@ -2102,19 +1639,6 @@ class SliceGrad(PrimitiveWithInfer):
                 'value': None}
 
 
-class NLLLossGrad(PrimitiveWithInfer):
-    """Computes the gradients of `NLLLoss`."""
-
-    @prim_attr_register
-    def __init__(self, reduction="mean", ignore_index=-100):
-        """Initialize NLLLoss"""
-        self.init_prim_io_names(inputs=['x', 'loss_grad', 'target', 'weight', 'total_weight'], outputs=['x_grad'])
-        self.reduction = validator.check_string(reduction, ['none', 'sum', 'mean'], 'reduction', self.name)
-        self.ignore_index = ignore_index
-        self.add_prim_attr('reduction', self.reduction)
-        self.add_prim_attr('ignore_index', self.ignore_index)
-
-
 class SmoothL1LossGrad(Primitive):
     """Computes gradient for prediction on SmoothL1Loss."""
 
@@ -2134,35 +1658,6 @@ class SoftMarginLossGrad(Primitive):
         self.reduction = validator.check_string(reduction, ['none', 'sum', 'mean'], 'reduction', self.name)
 
 
-class StridedSliceV2Grad(Primitive):
-    """
-    Performs grad of StridedSliceV2 operation.
-
-    Inputs:
-        - **shapex** (Tensor) - StridedSliceV2 shape of input
-        - **begin** (tuple[int]) - A tuple which represents the location where to start. Only
-          constant value is allowed.
-        - **end** (tuple[int]) - A tuple or which represents the maximum location where to end.
-          Only constant value is allowed.
-        - **strides** (tuple[int]) - A tuple which represents the stride is continuously added
-          before reaching the maximum location. Only constant value is allowed.
-        - **dy** (Tensor) - The output of StridedSliceV2
-
-    Outputs:
-        Tensor, the shape same as the input of StridedSliceV2
-    """
-
-    @prim_attr_register
-    def __init__(self,
-                 begin_mask=0,
-                 end_mask=0,
-                 ellipsis_mask=0,
-                 new_axis_mask=0,
-                 shrink_axis_mask=0):
-        """Initialize StridedSliceV2Grad"""
-        self.init_prim_io_names(inputs=['shapex', 'begin', 'end', 'strides', 'dy'], outputs=['output'])
-
-
 class StridedSliceGrad(Primitive):
     """
     Performs grad of StridedSlice operation.
@@ -2301,19 +1796,6 @@ class RefToEmbed(Primitive):
         pass
 
 
-class AtanGrad(Primitive):
-    """
-    Computes AtanGrad of input element-wise.
-
-    Returns:
-        Tensor, has the same type as input.
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize AtanGrad"""
-
-
 class BasicLSTMCellCStateGrad(PrimitiveWithInfer):
     """Computes the state gradients of BasicLSTMCell."""
 
@@ -2480,51 +1962,6 @@ class MvlgammaGrad(Primitive):
         self.p = validator.check_value_type('p', p, [int], self.name)
 
 
-class MaskedSelectGrad(PrimitiveWithInfer):
-    """Computes gradient for MaskedSelect."""
-
-    @prim_attr_register
-    def __init__(self):
-        pass
-
-    def infer_shape(self, x, mask, grad):
-        return x
-
-    def infer_dtype(self, x, mask, grad):
-        return x
-
-
-class SoftShrinkGrad(Primitive):
-    r"""
-          Gradients for SoftShrink operation.
-
-          Args:
-              lambd – The \lambdaλ (must be no less than zero) value for the Softshrink formulation. Default: 0.5.
-
-          Inputs:
-              - **input_grad** (Tensor) - The input gradient.
-              - **input_x** (Tensor) - The input of SoftShrink with data type of float16 or float32.
-                Any number of additional dimensions.
-
-          Outputs:
-              output - Tensor, has the same shape and data type as input_x.
-
-          Raises:
-              TypeError: If lambd is not a float.
-              TypeError: If dtype of input_x is neither float16 nor float32.
-              ValueError: If lambd is less than to 0.
-
-          Supported Platforms:
-              ``Ascend``
-      """
-
-    @prim_attr_register
-    def __init__(self, lambd=0.5):
-        self.init_prim_io_names(inputs=['input_grad', 'input_x'], outputs=['output'])
-        validator.check_value_type("lambd", lambd, [float], self.name)
-        validator.check_number("lambd", lambd, 0, validator.GE, self.name)
-
-
 class CdistGrad(Primitive):
     """Computes gradient for Cdist."""
 
@@ -2616,40 +2053,6 @@ class MultilabelMarginLossGrad(Primitive):
         self.init_prim_io_names(inputs=['y_grad', 'x', 'target', 'is_target'], outputs=['x_grad'])
 
 
-class HShrinkGrad(Primitive):
-    """
-    Computes gradients for HShrinkGrad operation.
-
-    Args:
-        lambd (float): the λ value for the Hardshrink formulation. Default: 0.5
-
-    Inputs:
-        - **Gradients** (Tensor) - the gradients of loss to output of HShrink function.
-          Currently gradients data type only support float16 and float32.
-        - **Features** (Tensor) - Must be the input `input_x` of the forward operator HSHrink.
-          Currently features data type only support float16 and float32.
-
-    Outputs:
-        backprops - Tensor, with the same shape and data type as `features`.
-
-    Rasise:
-        ValueError: If `lambd` is not a float.
-        ValueError: If shape of `gradients` is not the same as `features`.
-        TypeError: If dtype of `gradients` is not the same as `features`.
-        TypeError: If dtype of `gradients` or `features` is neither float16 nor float32.
-
-    Supported Platforms:
-        ``Ascend`` ``GPU`` ``CPU``
-    """
-
-    @prim_attr_register
-    def __init__(self, lambd=0.5):
-        validator.check_value_type("lambd", lambd, [float], self.name)
-        if lambd < 0.0:
-            lambd = 0.0
-            self.add_prim_attr('lambd', lambd)
-
-
 class Dilation2DBackpropInput(Primitive):
     """
     Computes the gradient of morphological 2-D dilation with respect to the input.
@@ -2962,6 +2365,12 @@ class MultiMarginLossGrad(Primitive):
     Supported Platforms:
         ``Ascend``  ``CPU``
     """
+    __mindspore_signature__ = (
+        sig.make_sig('y_grad'),
+        sig.make_sig('x'),
+        sig.make_sig('target'),
+        sig.make_sig('weight', default=None)
+    )
 
     @prim_attr_register
     def __init__(self, p=1, margin=1.0, reduction="mean"):
@@ -2972,96 +2381,8 @@ class MultiMarginLossGrad(Primitive):
         self.reduction = validator.check_string(reduction, ['none', 'sum', 'mean'], 'reduction', self.name)
         self.init_prim_io_names(inputs=['y_grad', 'x', 'target', 'weight'], outputs=['x_grad'])
 
-
-class UpsampleTrilinear3DGrad(Primitive):
-    r"""
-    Upsample the 3-D gradient data with trilinear interpolation algorithm.
-
-    Note:
-        One of 'scales' and 'output_size' must be specified. And it is an error if both are specified.
-
-    Args:
-        align_corners (bool): An optional bool. Default: ``False``.
-
-    Inputs:
-        - **dy** (Tensor) - Tensor of shape [N, C, D, H, W]. Must be one of the following types:
-          float16, float32, float64.
-        - **input_size** (Union[tuple[int], list[int]]): An required listInt which contains 5 elements:
-          [batch, channels, depth, height, width]. Must:
-          input_size[0] == dy_tensor_size[0]
-          input_size[1] == dy_tensor_size[1].
-        - **output_size** (Union[tuple[int], list[int]]): An optional listInt. Default: ``None``.
-          It contains 3 elements: depth, height, width, whose elements should be the same as `dy`. Must:
-          dy_tensor_size[2] == floor(input_size[2] * scales[0]) == output_size[0]
-          dy_tensor_size[3] == floor(input_size[3] * scales[1]) == output_size[1]
-          dy_tensor_size[4] == floor(input_size[4] * scales[2]) == output_size[2].
-        - **scales** (Union[tuple[float], list[float]]): An optional listFloat. Default: ``None``.
-          The scale array along each dimension, contain 3 elements: scale_depth, scale_height, scale_width.
-          The number of elements of 'scales' should be the same as the rank of input `dy`.
-
-    Outputs:
-        - **dx** (Tensor) - A Tensor with shape depending on intput_size, and its' dtype is the same as `dy`.
-    """
-    @prim_attr_register
-    def __init__(self, align_corners=False):
-        """Initialize UpsampleTrilinear3DGrad."""
-        self.init_prim_io_names(
-            inputs=['dy', 'input_size', 'output_size', 'scales'],
-            outputs=['dx'])
-        self.align_corners = align_corners
-        self.add_prim_attr('align_corners', self.align_corners)
-
-
-class GridSampler3DGrad(Primitive):
-    """
-    Computes gradients for GridSampler3D operation.
-
-    Args:
-        interpolation_mode (str): An optional string specifying the interpolation method. The optional values are
-            "bilinear" or "nearest". Default: "bilinear".
-        padding_mode (str): An optional string specifying the pad method. The optional values are "zeros", "border" or
-            "reflection". Default: "zeros".
-        align_corners (bool): An optional bool. If "true", the centers of the corner pixels of the input and output
-            tensors are aligned. Defaults to "false".
-
-    Inputs:
-        - **grad** (Tensor) - A 5-D tensor whose dtype is float32 or float64 and whose shape is :math:`(N, C, D_{out},
-          H_{out}, W_{out})`. The shape is inconsistent with the shape of the output result of forward calculation.
-        - **input_x** (Tensor) - A 5-D tensor whose dtype is the same as `grad` and whose shape is :math:`(N, C,
-          D_{in}, H_{in}, W_{in})`.
-        - **grid** (Tensor) - A 5-D tensor whose dtype is the same as `grad` and whose shape is :math:`(N, D_{out},
-          H_{out}, W_{out}, 3)`.
-
-    Outputs:
-        - **dx** (Tensor) - A 5-D tensor whose dtype and shape are the same as `input_x`.
-        - **dgrid** (Tensor) - A 5-D tensor whose dtype and shape are the same as `grid`.
-
-    Raises:
-        TypeError: If `grad`, `input_x` or `grid` is not a Tensor.
-        TypeError: If the dtypes of `grad`, `input_x` and `grid` are inconsistent.
-        TypeError: If the dtype of `grad`, `input_x` or `grid` is not a valid type.
-        TypeError: If `align_corners` is not a boolean value.
-        ValueError: If the rank of `grad`, `input_x` or `grid` is not equal to 5.
-        ValueError: If the first dimension of `grad`, `input_x` and `grid` are inconsistent.
-        ValueError: If the last dimension of `grid` is not equal to 3.
-        ValueError: If `interpolation_mode` is not "bilinear", "nearest" or a string value.
-        ValueError: If `padding_mode` is not "zeros", "border", "reflection" or a string value.
-        ValueError: If the shape of `grad` is inconsistent with the shape of the output result of forward calculation.
-
-    Supported Platforms:
-        ``GPU`` ``CPU``
-    """
-
-    @prim_attr_register
-    def __init__(self, interpolation_mode='bilinear', padding_mode='zeros', align_corners=False):
-        """Initialize GridSampler3DGrad."""
-        validator.check_string(interpolation_mode, ['bilinear', 'nearest'], 'interpolation_mode', self.name)
-        validator.check_string(padding_mode, ['zeros', 'border', 'reflection'], 'padding_mode', self.name)
-        validator.check_bool(align_corners, 'align_corners', self.name)
-        self.init_prim_io_names(inputs=['grad', 'input_x', 'grid'], outputs=['dx', 'dgrid'])
-        self.add_prim_attr('interpolation_mode', interpolation_mode)
-        self.add_prim_attr('padding_mode', padding_mode)
-        self.add_prim_attr('align_corners', align_corners)
+    def __call__(self, y_grad, x, target, weight=None):
+        return super().__call__(y_grad, x, target, weight)
 
 
 class SparseSegmentMeanGrad(Primitive):
@@ -3466,136 +2787,6 @@ class SparseSegmentSqrtNGrad(Primitive):
         self.init_prim_io_names(inputs=['x', 'indices', 'segment_ids', 'output_dim0'], outputs=['y'])
 
 
-class GridSampler2DGrad(Primitive):
-    """
-    Computes gradients for GridSampler2D operation.
-
-      Args:
-        interpolation_mode (str): An optional string specifying the interpolation method. The optional values are
-            "bilinear" or "nearest". Default: "bilinear".
-        padding_mode (str): An optional string specifying the pad method. The optional values are "zeros", "border" or
-            "reflection". Default: "zeros".
-        align_corners (bool): An optional bool. If "true", the centers of the corner pixels of the input and output
-            tensors are aligned. Defaults to "false".
-
-      Inputs:
-        - **grad** (Tensor) - A 4-D tensor whose dtype is float16 or float32 and whose shape is :math:`(N, C,
-          H_{out}, W_{out})`. The shape is inconsistent with the shape of the output result of forward calculation.
-        - **input_x** (Tensor) - A 4-D tensor whose dtype is the same as `grad` and whose shape is :math:`(N, C,
-          H_{in}, W_{in})`.
-        - **grid** (Tensor) - A 4-D tensor whose dtype is the same as `grad` and whose
-          shape is :math:`(N, H_{out}, W_{out}, 2)`.
-
-    Outputs:
-        - **dx** (Tensor) - A 4-D tensor whose dtype and shape are the same as `input_x`.
-        - **dgrid** (Tensor) - A 4-D tensor whose dtype and shape are the same as `grid`.
-
-    Raises:
-        TypeError: If `grad`, `input_x` or `grid` is not a Tensor.
-        TypeError: If the dtypes of `grad`, `input_x` and `grid` are inconsistent.
-        TypeError: If the dtype of `grad`, `input_x` or `grid` is not a valid type.
-        TypeError: If `align_corners` is not a boolean value.
-        ValueError: If the rank of `grad`, `input_x` or `grid` is not equal to 4.
-        ValueError: If the first dimension of `grad`, `input_x` and `grid` are inconsistent.
-        ValueError: If the last dimension of `grid` is not equal to 2.
-        ValueError: If `interpolation_mode` is not "bilinear", "nearest" or a string value.
-        ValueError: If `padding_mode` is not "zeros", "border", "reflection" or a string value.
-        ValueError: If the shape of `grad` is inconsistent with the shape of the output result of forward calculation.
-
-    Supported Platforms:
-        ``GPU`` ``CPU``
-    """
-
-    @prim_attr_register
-    def __init__(self, interpolation_mode='bilinear', padding_mode='zeros', align_corners=False):
-        """Initialize GridSampler2DGrad."""
-        validator.check_string(interpolation_mode, ['bilinear', 'nearest'], 'interpolation_mode', self.name)
-        validator.check_string(padding_mode, ['zeros', 'border', 'reflection'], 'padding_mode', self.name)
-        validator.check_bool(align_corners, 'align_corners', self.name)
-        self.init_prim_io_names(inputs=['grad', 'input_x', 'grid'], outputs=['dx', 'dgrid'])
-        self.add_prim_attr('interpolation_mode', interpolation_mode)
-        self.add_prim_attr('padding_mode', padding_mode)
-        self.add_prim_attr('align_corners', align_corners)
-
-
-class ResizeBicubicGrad(Primitive):
-    """
-    Computes gradients for ResizeBicubicGrad operation.
-
-    Args:
-        align_corners (bool):If true, the centers of the 4 corner pixels of the input
-    and output tensors are aligned, preserving the values at the corner pixels.Default: ``False``.
-        half_pixel_centers (bool): An optional bool. Default: ``False``.
-
-    Inputs:
-        - **grads** (Tensor) - A Tensor of type float. 4-D with shape
-          [batch, height, width,channels]. The format must be NHWC.
-        - **original_image** (Tensor) - A Tensor. Must be one of the following types: float,double.
-          4-D with shape [batch, orig_height, orig_width, channels], The image tensor that was resized.
-          The format must be NHWC.
-
-    Outputs:
-        A 4-D Tensor , with the same shape and data type as `original_image`.
-
-    Rasise:
-        TypeError: If `grads` is not allowed.
-        TypeError: If `original_image` is not allowed.
-        ValueError: If `images` dim is not 4.
-        ValueError: If `size` dim is not 4.
-
-    Supported Platforms:
-        ``Ascend`` ``GPU`` ``CPU``
-    """
-    @prim_attr_register
-    def __init__(self, align_corners=False, half_pixel_centers=False):
-        """Initialize CropAndResize"""
-        validator.check_value_type('align_corners', align_corners, bool, self.name)
-        validator.check_value_type('half_pixel_centers', half_pixel_centers, bool, self.name)
-        self.init_prim_io_names(inputs=['grads', 'original_image'], outputs=['y'])
-
-    def __infer__(self, grads, original_image):
-        # get shape
-        grads_shape = list(grads['shape'])
-        original_image_shape = list(original_image['shape'])
-        # get value
-        if grads['value'] is None:
-            raise ValueError(
-                f"For '{self.name}', the 'grads' cannot be None, but got {grads['value']}."
-            )
-        if original_image['value'] is None:
-            raise ValueError(
-                f"For '{self.name}', the 'original_image' cannot be None, but got {original_image['value']}."
-            )
-        # get dtype
-        grads_dtype = grads['dtype']
-        original_image_dtype = original_image['dtype']
-        # check dytpe
-        validator.check_tensor_dtype_valid("grads", grads_dtype,
-                                           [mstype.float32], self.name)
-        validator.check_tensor_dtype_valid("original_image", original_image_dtype,
-                                           [mstype.float32, mstype.float64], self.name)
-        # check input shape rank
-        validator.check("grads rank", len(grads_shape), "expected", 4, validator.EQ, self.name)
-        validator.check("original_image rank", len(original_image_shape), "expected", 4, validator.EQ, self.name)
-        validator.check("batch_size equal", grads_shape[0], "expected",
-                        original_image_shape[0], validator.EQ, self.name)
-        validator.check("channel equal", grads_shape[3], "expected", original_image_shape[3], validator.EQ, self.name)
-        # check original_image_shape and grads_shape
-        validator.check("original_image[0] and grads[0]", original_image_shape[0],
-                        "expected", grads_shape[0], validator.EQ, self.name)
-        validator.check("original_image[3] and grads[3]", original_image_shape[3],
-                        "expected", grads_shape[3], validator.EQ, self.name)
-
-        batch_size = grads_shape[0]
-        height = original_image_shape[1]
-        width = original_image_shape[2]
-        channel = grads_shape[3]
-        out_shape = (batch_size, height, width, channel)
-        return {'shape': out_shape,
-                'dtype': original_image_dtype,
-                'value': None}
-
-
 class SparseSliceGrad(Primitive):
     r"""
     Computes gradients for SparseSlice operation.
@@ -3717,13 +2908,6 @@ class AffineGridGrad(Primitive):
         self.init_prim_io_names(inputs=['y_grad', 'x_size'], outputs=['x_grad'])
 
 
-class HSigmoidGrad(Primitive):
-    """Gets the gradient of HSigmoid operation."""
-    @prim_attr_register
-    def __init__(self):
-        """Initialize HSigmoidGrad"""
-        self.init_prim_io_names(inputs=['grads', 'input_x'], outputs=['output'])
-
 
 class GluGrad(Primitive):
     """
@@ -3737,46 +2921,6 @@ class GluGrad(Primitive):
         validator.check_value_type("axis", axis, [int], self.name)
 
 
-class CholeskyGrad(Primitive):
-    r"""
-    Computes the reverse mode backpropgated gradient of the Cholesky algorithm.
-
-    Inputs:
-        - **x** (Tensor) - A tensor with float32 or float64 data type.
-        - **grad** (Tensor) - A tensor with float32 or float64 data type. `x` should have
-          the same dtype with `a`.
-
-    Outputs:
-        Tensor, has the same dtype as `a` and `x`.
-
-    Raises:
-        TypeError: If x is not Tensor.
-        TypeError: If grad is not Tensor.
-        TypeError: If dtype of input x and grad is not float64 nor float32,
-        TypeError: If x has different dtype with grad.
-        ValueError: If input tensor's last two dims are not equal,
-        ValueError: If the shape of x and grad mismatch.
-
-    Supported Platforms:
-        ``Ascend``
-
-    Examples:
-        >>> x = Tensor(np.array([[4, 2],[2, 3]]), mstype.float64)
-        >>> grad = Tensor(np.array([[4, 2],[2, 3]]), mstype.float64)
-        >>> choleskygrad = G.CholeskyGrad()
-        >>> output = choleskygrad(x, grad)
-        >>> print (output)
-        [[0.5 0. ]
-         [0.  0.5]]
-
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize CholeskyGrad"""
-        self.init_prim_io_names(inputs=['x', 'grad'], outputs=['y'])
-
-
 class MapTensorGetGrad(Primitive):
     """
     Computes gradients for MapTensorGet operation.
@@ -3832,53 +2976,3 @@ class WKVGrad(Primitive):
         """Initialize WKVGrad."""
         self.init_prim_io_names(inputs=["time_first", "time_decay", "key", "value", "gy"],
                                 outputs=["gw", "gu", "gk", "gv"])
-
-
-class FlashAttentionScoreGrad(Primitive):
-    r"""
-    Calculates the gradient of FlashAttentionScore operation.
-    .. warning::
-        This is an experimental API that is subject to change or deletion.
-
-    Supported Platforms:
-        ``Ascend``
-    """
-    @prim_attr_register
-    def __init__(self, head_num, keep_prob=1.0, scale_value=1.0, pre_tokens=65536, next_tokens=65536, inner_precise=1,
-                 input_layout='BSH', sparse_mode=0):
-        """Initialize FlashAttentionScoreGrad."""
-        validator.check_value_type('head_num', head_num, [int], self.name)
-        validator.check_value_type('keep_prob', keep_prob, [int, float], self.name)
-        validator.check_float(keep_prob, 0.0, validator.GE, "keep_prob", self.name)
-        validator.check_float(keep_prob, 1.0, validator.LE, "keep_prob", self.name)
-        validator.check_value_type('scale_value', scale_value, [float], self.name)
-        validator.check_value_type('pre_tokens', pre_tokens, [int], self.name)
-        validator.check_value_type('next_tokens', next_tokens, [int], self.name)
-        validator.check_value_type('inner_precise', inner_precise, [int], self.name)
-        validator.check_value_type('sparse_mode', sparse_mode, [int], self.name)
-        if inner_precise not in [0, 1]:
-            raise ValueError(f"Attribute 'inner_precise' must be either 0 or 1, but got {inner_precise}")
-        validator.check_value_type('input_layout', input_layout, [str], self.name)
-        if input_layout not in ["BSH", "BNSD"]:
-            raise ValueError(f"Attribute 'input_layout' must be either 'BSH' or 'BNSD', but got {input_layout}")
-        self.init_prim_io_names(inputs=['query', 'key', 'value', 'dy', 'pse_shift', 'drop_mask', "padding_mask",
-                                        'attn_mask', 'softmax_max', 'softmax_sum', 'softmax_out', 'attention_in',
-                                        'prefix'],
-                                outputs=['dq', 'dk', 'dv', 'dpse'])
-
-
-class RmsNormGrad(Primitive):
-    r"""
-    Calculates the gradient of RmsNorm operation.
-    .. warning::
-        This is an experimental API that is subject to change or deletion.
-
-    Supported Platforms:
-        ``Ascend``
-    """
-
-    @prim_attr_register
-    def __init__(self):
-        """Initialize RmsNormGrad."""
-        self.init_prim_io_names(inputs=["dy", "x", "rstd", "gamma"],
-                                outputs=["dx", "dgamma"])