Layers within a layer in Caffe

问题

I have a custom loss layer which I wrote, this layer applies softmax and sigmoid activation to part of the bottom[0] blob.

Ex: `bottom[0]` is of shape (say): `[20, 7, 7, 50]` (`NHWC` format)
I would like to apply `softmax` to `[20, 7, 7, 25]` (first 25 channels) and 
`sigmoid` to `[20, 7, 7, 1]` (just one channel) and the remaining 24 channels are taken in as it is.

How do I effectively allocate memory to the input blobs of these two softmax and sigmoid layers and also free this memory ?

回答1:

Instead of allocating data internally, you can simply use a "Slice" layer externally and slice the input blob using caffe "off-the-shelf" layers:

layer {
  name: "slice"
  type: "Slice"
  bottom: "input_to_loss"
  top: "to_softmax"
  top: "to_sigmoid"
  top: "leftovers"
  slice_param { 
    axis: -1  # slice the last axis
    slice_point: 25
    slice_point: 26
  }
}

回答2:

All the intermediate activations, along with the network's input and output blobs are managed by the net class and set up in the Net<Dtype>::Init function in src/caffe/net.cpp.

You don't need to allocate/deallocate the top and bottom blob memory from within the layer itself.

来源：https://stackoverflow.com/questions/44955498/layers-within-a-layer-in-caffe