Question

我想在训练时修复一些图层的参数，所以我在这样的图层中使用propagate_down:0：

  layer {
  name: "conv5_2"
  type: "Eltwise"
  bottom: "conv5_1"
  bottom: "conv5_2_1x1_increase"
  top: "conv5_2"
  param{
    propagate_down：0
  }
  eltwise_param {
    operation: SUM
  }
}

我想在此图层之前修复参数，并且只在此图层之后微调参数。

但我遇到了错误：

I0316 15:08:21.036813 20524 solver.cpp:81] Creating training net from train_net file: /home/muses/WSnet/C_coarse2fine/config/pspnet101_ele_8s_train.prototxt
[libprotobuf ERROR google/protobuf/text_format.cc:274] Error parsing text-format caffe.NetParameter: 6595:19: Interpreting non ascii codepoint 239.
[libprotobuf ERROR google/protobuf/text_format.cc:274] Error parsing text-format caffe.NetParameter: 6595:19: Message type "caffe.ParamSpec" has no field named "propagate_down".
F0316 15:08:21.039957 20524 upgrade_proto.cpp:88] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: /home/muses/WSnet/C_coarse2fine/config/pspnet101_ele_8s_train.prototxt

如何解决这个问题？

Answer 1

参数propagate_down用于指示是否将渐变渐变到图层的输入。不意味着是否要更新参数要在caffe中修复参数，您需要为该特定参数设置param { lr_mult: 0 }。

caffe |消息类型＆＃34; caffe.ParamSpec＆＃34;没有名为＆＃34; propagate_down＆＃34;

1 个答案: