Adding Deconvolution Support

[RishabhMalviya]

I needed to convert the Deconvolution layers in some Caffe models for a paper I was implementing in Torch. I thought it would be a good idea to put in a pull request for the added support.

I've also updated the caffe.proto to match the one currently at https://github.com/BVLC/caffe/blob/master/src/caffe/proto/caffe.proto.

[RishabhMalviya]

Okay wait, this actually might not be working perfectly. I tried loading a fully convolutionalized version of AlexNet from here (using train.prototxt and the model linked to in the caffemodel-url file) by running:

model = loadcaffe.load('train.prototxt', 'fcn-alexnet-pascal.caffemodel', 'cudnn')

The output is:

Successfully loaded fcn-alexnet-pascal.caffemodel
warning: module 'data [type Input]' not found
warning: module 'data_data_0_split [type Split]' not found
conv1: 96 3 11 11
conv2: 256 48 5 5
conv3: 384 256 3 3
conv4: 384 192 3 3
conv5: 256 192 3 3
fc6: 4096 256 6 6
fc7: 4096 4096 1 1
score_fr: 21 4096 1 1

There's no mention of the deconvolution layer named "upscore", present towards the end of the net (in the train.prototxt file);

...

layer {
  name: "upscore"
  type: "Deconvolution"
  bottom: "score_fr"
  top: "upscore"
  param {
    lr_mult: 0
  }
  convolution_param {
    num_output: 21
    bias_term: false
    kernel_size: 63
    stride: 32
  }
}
layer {
  name: "score"
  type: "Crop"
  bottom: "upscore"
  bottom: "data"
  top: "score"
  crop_param {
    axis: 2
    offset: 18
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "score"
  bottom: "label"
  top: "loss"
  loss_param {
    ignore_label: 255
    normalize: true
  }
}

Could this be because it's followed by another unrecognized layer type ('Crop')?

The only good news is that it's no longer complaining about not finding 'module 'data [type Deconvolution]'' anymore.

[Naruto-Sasuke]

In loadcaffe.cpp, line 203, break should be removed.
Also your pull cannot work for deconv layer as the bias copy
inconsistent. (I don't know how to directly comment the exact line..)

[Naruto-Sasuke]

Hope can help you, we badly needs it now.

[Naruto-Sasuke]

It should be removed

[Naruto-Sasuke]

I don't know whether it is right to exchange nInputPlane with nOutputPlane. However it actually avoid mistakes in https://github.com/szagoruyko/loadcaffe/blob/master/loadcaffe.lua#L39 . However it also gives wrong loading info in
https://github.com/szagoruyko/loadcaffe/blob/master/loadcaffe.cpp#L692 . (Specially for FullConvolution, then printf("%s: %d %d %d %d\n", name, nInputPlane, nOutputPlane, kW, kH) is right....)