* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file group_norm_swish_grad_proto.h
* \brief
*/
#ifndef OPS_NORM_GROUP_NORM_SWISH_GRAD_H_
#define OPS_NORM_GROUP_NORM_SWISH_GRAD_H_
#include "graph/operator_reg.h"
namespace ge {
* @brief Performs the backward operation of group normalization and swish.
* @par Inputs:
* Six input, including:
* @li dy: A Tensor. Group grad. Datatype support float32, float16, bfloat16. Format support ND.
* @li mean: A Tensor. Mean of each group. Datatype support float32, float16, bfloat16. Format support ND.
* @li rstd: A Tensor. Reciprocal standard deviation of each group. Datatype support float32, float16, bfloat16. Format
support ND.
* @li x: A Tensor. Specifies the offset. Datatype support float32, float16, bfloat16. Format support ND. Same shape as
mean.
* @li gamma: A Tensor. Specifies the scaling factor. Datatype support float32, float16, bfloat16. Format support ND.
Same shape as dy.
* @li beta: A Tensor. Specifies the intercept. Datatype support float32, float16, bfloat16. Format support ND. Same
shape as gamma.
* @par Attributes:
* @li num_groups: Int. Number specifying the number of group.
* @li data_format: An optional String, Defaults to NCHW.
* @li swish_scale: An optional float. Defaults to "1.0".
* @li dgamma_is_require: An optional bool, controls whether to return weight.grad. Defaults to true.
* @li dbeta_is_require: An optional bool, controls whether to return beta.grad. Defaults to true.
* @par Outputs:
* Three output, including:
* @li dx: A Tensor. x factor grad. Datatype is the same as the input Datatype. Format support ND.
* @li dgamma: A Tensor. scale factor grad. Datatype is the same as the input Datatype. Format support ND.
* @li dbeta: A Tensor. offset factor grad. Datatype is the same as the input Datatype. Format support ND.
* @par Third-party framework compatibility
* @li Compatible with the backward of PyTorch operator GroupNorm and Swish.
*/
REG_OP(GroupNormSwishGrad)
.INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.INPUT(mean, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.INPUT(rstd, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.INPUT(gamma, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.INPUT(beta, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.OUTPUT(dx, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.OUTPUT(dgamma, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.OUTPUT(dbeta, TensorType({DT_FLOAT16, DT_FLOAT, DT_BF16}))
.REQUIRED_ATTR(num_groups, Int)
.ATTR(data_format, String, "NCHW")
.ATTR(swish_scale, Float, 1.0)
.ATTR(dgamma_is_require, Bool, true)
.ATTR(dbeta_is_require, Bool, true)
.OP_END_FACTORY_REG(GroupNormSwishGrad)
}
#endif