You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

correlation.h 1.6 kB

3 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152
  1. /**
  2. * Copyright 2020 Huawei Technologies Co., Ltd
  3. *
  4. * Licensed under the Apache License, Version 2.0 (the "License");
  5. * you may not use this file except in compliance with the License.
  6. * You may obtain a copy of the License at
  7. *
  8. * http://www.apache.org/licenses/LICENSE-2.0
  9. *
  10. * Unless required by applicable law or agreed to in writing, software
  11. * distributed under the License is distributed on an "AS IS" BASIS,
  12. * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. * See the License for the specific language governing permissions and
  14. * limitations under the License.
  15. */
  16. /*!
  17. * \file correlation.h
  18. * \brief
  19. */
  20. #ifndef GE_OP_CORRELATION_OPS_H
  21. #define GE_OP_CORRELATION_OPS_H
  22. #include "graph/operator_reg.h"
  23. namespace ge {
  24. /**
  25. *@brief Computes a 2D Correlation given 4D "x" and "filter" tensors.
  26. *
  27. *@par Inputs:
  28. * @li filter: A 4D tensor of filters.
  29. * @li x: A 4D tensor of input images, batch number must equal to batch
  30. * number of "filter", and channel must equal to channel of "filter".
  31. *
  32. *@par Attributes:
  33. * @li groups: set correlation mode, must be 1 or channel.
  34. *
  35. *@par Outputs:
  36. *y: A Tensor. Has the same type as "x".
  37. *@par Third-party framework compatibility
  38. * Compatible with caffe correlation custom operator.
  39. */
  40. REG_OP(Correlation)
  41. .INPUT(filter, TensorType({DT_FLOAT16, DT_INT8}))
  42. .INPUT(x, TensorType({DT_FLOAT16, DT_INT8}))
  43. .OUTPUT(y, TensorType({DT_FLOAT16, DT_INT32}))
  44. .ATTR(groups, Int, 1)
  45. .OP_END_FACTORY_REG(Correlation)
  46. } // namespace ge
  47. #endif // GE_OP_NN_CALCULATION_OPS_H

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示