How do I reduce 4096-dimensional feature vector to 1024-dimensional vector in CNN Caffemodel?

问题

I used 16-layers VGGnet to extract features from an image. It outputs a 4096-dimensional feature vector. However, I need a 1024-dimensional vector. How do I further reduce this 4096-vector into 1024-vector? Do I need to add a new layer on top of fc7?

回答1:

Yes, you need to add another layer on top of fc7. This is how your last few layers should be like

layers {
  bottom: "fc7"
  top: "fc7"
  name: "relu7"
  type: RELU
}
layers {
  bottom: "fc7"
  top: "fc7"
  name: "drop7"
  type: DROPOUT
  dropout_param {
    dropout_ratio: 0.5
  }
}
layers {
  name: "fc8"
  bottom: "fc7"
  top: "fc8"
  type: INNER_PRODUCT
  inner_product_param {
    num_output: 1024
  }
  blobs_lr: 0
  blobs_lr: 0
}
layers {
  name: "loss"
  type: SOFTMAX_LOSS
  bottom: "fc8"
  bottom: "label"
  top: "loss/loss"
}
layers {
  name: "accuracy/top1"
  type: ACCURACY
  bottom: "fc8"
  bottom: "label"
  top: "accuracy@1"
  include: { phase: TEST }
  accuracy_param {
    top_k: 1
  }
}