Pow和Eltwise算子推理出错

简单的测试了一下Pow和Eltwise算子，可以caffe转ms, 但在推理和量化的时候都有问题。

输入图片说明
测试的网络结构：

name:"test_power"
layer {
  name: "frame_1" 
  type: "Input"
  top: "frame_1"
  input_param {
    shape {
      dim: 1
      dim: 3
      dim: 272
      dim: 480
    }
  }
}
layer {
  name: "frame_2" 
  type: "Input"
  top: "frame_2"
  input_param {
    shape {
      dim: 1
      dim: 3
      dim: 272
      dim: 480
    }
  }
}
layer {
  name: "concatenate_1"
  type: "Concat"
  bottom: "frame_1"
  bottom: "frame_2"
  top: "concatenate_1"
  concat_param {
    axis: 1
  }
}

layer {
  name: "head_output"
  type: "Convolution"
  bottom: "concatenate_1"
  top: "head_output"
  convolution_param {
    num_output: 2
    bias_term: true
    pad_h: 1
    pad_w: 1
    kernel_h: 3
    kernel_w: 3
    stride_h: 1
    stride_w: 1
    dilation: 1
    weight_filler: {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "fg_output"
  type: "Convolution"
  bottom: "concatenate_1"
  top: "fg_output"
  convolution_param {
    num_output: 2
    bias_term: true
    pad_h: 1
    pad_w: 1
    kernel_h: 3
    kernel_w: 3
    stride_h: 1
    stride_w: 1
    dilation: 1
    weight_filler: {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "flow_output"
  type: "Convolution"
  bottom: "concatenate_1"
  top: "flow_output"
  convolution_param {
    num_output: 2
    bias_term: true
    pad_h: 1
    pad_w: 1
    kernel_h: 3
    kernel_w: 3
    stride_h: 1
    stride_w: 1
    dilation: 1
    weight_filler: {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "head_output_reverse_layer"
  bottom: "head_output"
  top: "head_output_reverse"
  type: "Power"
  power_param {
    power: 1
    scale: -1
    shift: 0
  }
}

layer {
  name: "layer_threshold_head_output"
  bottom: "head_output_reverse"
  top: "argmax_thr_head_output"
  type: "ArgMax"
  argmax_param {
    axis: 1
  }
}

layer {
  name: "fg_output_reverse_layer"
  bottom: "fg_output"
  top: "fg_output_reverse"
  type: "Power"
  power_param {
    power: 1
    scale: -1
    shift: 0
  }
}

layer {
  name: "layer_threshold_fg_output"
  bottom: "fg_output_reverse"
  top: "argmax_thr_fg_output"
  type: "ArgMax"
  argmax_param {
    axis: 1
  }
}

layer {
  name: "flow_output_reverse_layer"
  bottom: "flow_output"
  top: "flow_output_reverse"
  type: "Power"
  power_param {
    power: 1
    scale: -1
    shift: 0
  }
}

layer {
  name: "layer_threshold_flow_output"
  bottom: "flow_output_reverse"
  top: "argmax_thr_flow_output"
  type: "ArgMax"
  argmax_param {
    axis: 1
  }
}

layer {
  name: "argmax_thr_flow_output_power_layer"
  bottom: "argmax_thr_flow_output"
  top: "argmax_thr_flow_output_p"
  type: "Power"
  power_param {
    power: 1
    scale: 4
    shift: 0
  }
}

  layer {
  name: "argmax_thr_head_output_power_layer"
  bottom: "argmax_thr_head_output"
  top: "argmax_thr_head_output_p"
  type: "Power"
  power_param {
    power: 1
    scale: 2
    shift: 0
  }
}

  layer {
  name: "res1"
  type: "Eltwise"
  bottom: "argmax_thr_flow_output_p"
  bottom: "argmax_thr_head_output_p"
  top: "res1"
  eltwise_param {
    operation: SUM
    coeff: 1
    coeff: 1
  }
}

layer {
  name: "final_res"
  type: "Eltwise"
  bottom: "res1"
  bottom: "argmax_thr_fg_output"
  top: "final_res"
  eltwise_param {
    operation: SUM
    coeff: 1
    coeff: 1
  }
}

此外，Eltwise算子只支持两个输入和系数为1的情况

Please assign maintainer to check this issue.
请为此issue分配处理人。
@fangwenyi @chengxiaoli

Please add labels (comp or sig), also you can visit https://gitee.com/mindspore/community/blob/master/sigs/dx/docs/labels.md to find more.
为了让代码尽快被审核，请您为Pull Request打上 组件(comp)或兴趣组(sig) 标签，打上标签的PR可以直接推送给责任人进行审核。
更多的标签可以查看https://gitee.com/mindspore/community/blob/master/sigs/dx/docs/labels.md
以组件相关代码提交为例，如果你提交的是data组件代码，你可以这样评论：
//comp/data
当然你也可以邀请data SIG组来审核代码，可以这样写：
//sig/data
另外你还可以给这个PR标记类型，例如是bugfix或者是特性需求：
//kind/bug or //kind/feature
恭喜你，你已经学会了使用命令来打标签，接下来就在下面的评论里打上标签吧！

你好，问题收到了，已经安排人员专门处理您的这个问题，请随时关注码云上信息

Eltwise报错情况：
输入图片说明

name:"test_power"
layer {
  name: "frame_1" 
  type: "Input"
  top: "frame_1"
  input_param {
    shape {
      dim: 1
      dim: 3
      dim: 272
      dim: 480
    }
  }
}
layer {
  name: "frame_2" 
  type: "Input"
  top: "frame_2"
  input_param {
    shape {
      dim: 1
      dim: 3
      dim: 272
      dim: 480
    }
  }
}
layer {
  name: "concatenate_1"
  type: "Concat"
  bottom: "frame_1"
  bottom: "frame_2"
  top: "concatenate_1"
  concat_param {
    axis: 1
  }
}

layer {
  name: "head_output"
  type: "Convolution"
  bottom: "concatenate_1"
  top: "head_output"
  convolution_param {
    num_output: 2
    bias_term: true
    pad_h: 1
    pad_w: 1
    kernel_h: 3
    kernel_w: 3
    stride_h: 1
    stride_w: 1
    dilation: 1
    weight_filler: {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "fg_output"
  type: "Convolution"
  bottom: "concatenate_1"
  top: "fg_output"
  convolution_param {
    num_output: 2
    bias_term: true
    pad_h: 1
    pad_w: 1
    kernel_h: 3
    kernel_w: 3
    stride_h: 1
    stride_w: 1
    dilation: 1
    weight_filler: {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "flow_output"
  type: "Convolution"
  bottom: "concatenate_1"
  top: "flow_output"
  convolution_param {
    num_output: 2
    bias_term: true
    pad_h: 1
    pad_w: 1
    kernel_h: 3
    kernel_w: 3
    stride_h: 1
    stride_w: 1
    dilation: 1
    weight_filler: {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}

layer {
  name: "head_output_reverse_layer"
  bottom: "head_output"
  top: "head_output_reverse"
  type: "Power"
  power_param {
    power: 1
    scale: -1
    shift: 0
  }
}

layer {
  name: "layer_threshold_head_output"
  bottom: "head_output_reverse"
  top: "argmax_thr_head_output"
  type: "ArgMax"
  argmax_param {
    axis: 1
  }
}

layer {
  name: "fg_output_reverse_layer"
  bottom: "fg_output"
  top: "fg_output_reverse"
  type: "Power"
  power_param {
    power: 1
    scale: -1
    shift: 0
  }
}

layer {
  name: "layer_threshold_fg_output"
  bottom: "fg_output_reverse"
  top: "argmax_thr_fg_output"
  type: "ArgMax"
  argmax_param {
    axis: 1
  }
}

layer {
  name: "flow_output_reverse_layer"
  bottom: "flow_output"
  top: "flow_output_reverse"
  type: "Power"
  power_param {
    power: 1
    scale: -1
    shift: 0
  }
}

layer {
  name: "layer_threshold_flow_output"
  bottom: "flow_output_reverse"
  top: "argmax_thr_flow_output"
  type: "ArgMax"
  argmax_param {
    axis: 1
  }
}

  layer {
  name: "res1"
  type: "Eltwise"
  bottom: "argmax_thr_flow_output"
  bottom: "argmax_thr_head_output"
  top: "res1"
  eltwise_param {
    operation: SUM
    coeff: 1
    coeff: 1
  }
}

layer {
  name: "final_res"
  type: "Eltwise"
  bottom: "res1"
  bottom: "argmax_thr_fg_output"
  top: "final_res"
  eltwise_param {
    operation: SUM
    coeff: 1
    coeff: 1
  }
}

这个eltwise确实只支持两个输入的操作,是一个arithmetic算子,认为两个就够用了.onnx标准也只有两个输入.
量化的时候没有找到pow算子的backend,是这个不支持量化,不要用全量化了,用部分量化就可已

没有量化，直接推理也不行呢，build model就会出错，上面的这两个情况，都是没有做量化，在前向推理的时候出现找不到backend的问题

你用我上面发的这两个model，测试看看呢

你用我上面发的这两个model，测试看看呢

@突tututu兔发一下你生成的ms模型吧,放到这个帖子附件里,我定位一下

貌似添加不了附件，我怎么发给你

貌似添加不了附件，我怎么发给你

@突tututu兔 1183489276@qq.com

邮箱在上面发一下吧

你好，问题解决了吗？如果没解决的话，可以找zhaodezan支撑下
同时由于长时间没有反馈，此ISSUE先关闭，如有问题，可以反馈下具体信息，并将ISSUE状态修改为WIP，我们这边会进一步跟踪，谢谢

GVP MindSpore / mindspore

内容风险标识

评论 (12)

GVPMindSpore / mindspore

内容风险标识