issues of "feature layer" in extract_feature.py

[666zcli]

Hi, yupan I try to use your codes on activitynet. And when I used the extract_feature.py to extract features, I downloaded the pretrained model KnetV3 from the link you provided. In your code, the last layer seemed to be 'global_pool', but I could not find "global_pool" in prototxt file. Could you help me?

[HYPJUDY]

Hi, @666zcli Thank you for pointing this out! I did modify the original prototxt files to correspond with the variable name (e.g., feature_layer) in data/extract_feature.py (and forgot to remind you!). For example, from the line#5906 to the last line in inception_v3_kinetics_pretrained/inception_v3_rgb_deploy.prototxt, I change the layer name of top_cls_pool and top_cls_global_pool to global_pool and remove top_cls_drop layer and fc_action layer for feature extraction as follows:

layer {
  name: "top_cls_pool"
  type: "Pooling"
  bottom: "mixed_10_join"
  top: "top_cls_global_pool"
  pooling_param {
    pool: AVE
    kernel_size: 8
    stride: 1
    pad: 0
  }
}
layer {
  name: "top_cls_drop"
  type: "Dropout"
  bottom: "top_cls_global_pool"
  top: "top_cls_drop"
  dropout_param {
    dropout_ratio: 0.699999988079
  }
}
layer {
  name: "fc_action"
  type: "InnerProduct"
  bottom: "top_cls_drop"
  top: "fc_action"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 1.0
    decay_mult: 2.0
  }
  inner_product_param {
    num_output: 400
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}

to

layer {
  name: "global_pool"
  type: "Pooling"
  bottom: "mixed_10_join"
  top: "global_pool"
  pooling_param {
    pool: AVE
    kernel_size: 8
    stride: 1
    pad: 0
  }
}
#layer {
#  name: "top_cls_drop"
#  type: "Dropout"
#  bottom: "global_pool"
#  top: "top_cls_drop"
#  dropout_param {
#    dropout_ratio: 0.699999988079
#  }
#}
#layer {
#  name: "fc-action"
#  type: "InnerProduct"
#  bottom: "top_cls_drop"
#  top: "fc-action"
#  param {
#    lr_mult: 1.0
#    decay_mult: 1.0
#  }
#  param {
#    lr_mult: 1.0
#    decay_mult: 2.0
#  }
#  inner_product_param {
#    num_output: 400
#    weight_filler {
#      type: "xavier"
#    }
#    bias_filler {
#      type: "constant"
#      value: 0.0
#    }
#  }
#}

I have made a commit to upload the prototxt files I used for your reference. Directly use these files should be fine, otherwise please reopen this issue : )