对于fcn，经常要使用到Deconvolution进行上采样。对于caffe使用者，使用Deconvolution上采样，其参数往往直接给定，不需要通过学习获得。

给定参数的方式很有意思，可以通过两种方式实现，但是这两种方式并非完全等价，各有各的价值。

第一种方式：通过net_surgery给定，

这种方式最开始出现在FCN中。https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn32s/solve.py
代码如下：

import caffe
import surgery, scoreimport numpy as np
import os
import systry:import setproctitlesetproctitle.setproctitle(os.path.basename(os.getcwd()))
except:passweights = '../ilsvrc-nets/vgg16-fcn.caffemodel'# init
caffe.set_device(int(sys.argv[1]))
caffe.set_mode_gpu()solver = caffe.SGDSolver('solver.prototxt')
solver.net.copy_from(weights)# surgeries  （这里就是对于反卷积层的参数进行初始化）
interp_layers = [k for k in solver.net.params.keys() if 'up' in k]
surgery.interp(solver.net, interp_layers)# scoring
val = np.loadtxt('../data/segvalid11.txt', dtype=str)for _ in range(25):solver.step(4000)score.seg_tests(solver, False, val, layer='score')

上采样的函数：

 #  make a bilinear interpolation kerneldef upsample_filt(self,size):factor = (size + 1) // 2if size % 2 == 1:center = factor - 1else:center = factor - 0.5og = np.ogrid[:size, :size]return (1 - abs(og[0] - center) / factor) * \(1 - abs(og[1] - center) / factor)# set parameters s.t. deconvolutional layers compute bilinear interpolation# N.B. this is for deconvolution without groupsdef interp_surgery(self,net, layers):for l in layers:print lm, k, h, w = net.params[l][0].data.shape   #仅仅修改w，不需要修改bias，其为0print("deconv shape:\n")print m, k, h, w if m != k and k != 1:print 'input + output channels need to be the same or |output| == 1'raiseif h != w:print 'filters need to be square'raisefilt = self.upsample_filt(h)print(filt)net.params[l][0].data[range(m), range(k), :, :] = filt

第二种方式：直接在Deconvolution中给定参数weight_filler，即：

代码如下:

layer {name: "fc8_upsample"type: "Deconvolution"bottom: "fc8"top: "fc8_upsample"param {lr_mult: 0decay_mult: 0}param {lr_mult: 0decay_mult: 0}convolution_param {num_output: 1kernel_size: 16stride: 8pad: 3weight_filler {   # 这里相当于上面的直接赋值type: "bilinear"}}
}

weight_filler初始化成双线性就等价于直接按照上面的方式赋值。

看起来好像以上两种方法一样，但是实际上有不同。主要区别在对于num_output>1的情形。

比如对于一个输入是2个通道的map，希望对其进行上采样，自然我们希望分别对于map放大即可。如果使用Deconvolution，则shape大小为2,2,16,16（设其大小为16*16）.不考虑bias项。

假设按照上面的方式初始化，则对于第一种方法，得到结果：
[0,0,:,:]:

[[ 0.00390625 0.01171875 0.01953125 0.02734375 0.03515625 0.04296875
0.05078125 0.05859375 0.05859375 0.05078125 0.04296875 0.03515625
0.02734375 0.01953125 0.01171875 0.00390625]
[ 0.01171875 0.03515625 0.05859375 0.08203125 0.10546875 0.12890625
0.15234375 0.17578125 0.17578125 0.15234375 0.12890625 0.10546875
0.08203125 0.05859375 0.03515625 0.01171875]
[ 0.01953125 0.05859375 0.09765625 0.13671875 0.17578125 0.21484375
0.25390625 0.29296875 0.29296875 0.25390625 0.21484375 0.17578125
0.13671875 0.09765625 0.05859375 0.01953125]
[ 0.02734375 0.08203125 0.13671875 0.19140625 0.24609375 0.30078125
0.35546875 0.41015625 0.41015625 0.35546875 0.30078125 0.24609375
0.19140625 0.13671875 0.08203125 0.02734375]
[ 0.03515625 0.10546875 0.17578125 0.24609375 0.31640625 0.38671875
0.45703125 0.52734375 0.52734375 0.45703125 0.38671875 0.31640625
0.24609375 0.17578125 0.10546875 0.03515625]
[ 0.04296875 0.12890625 0.21484375 0.30078125 0.38671875 0.47265625
0.55859375 0.64453125 0.64453125 0.55859375 0.47265625 0.38671875
0.30078125 0.21484375 0.12890625 0.04296875]
[ 0.05078125 0.15234375 0.25390625 0.35546875 0.45703125 0.55859375
0.66015625 0.76171875 0.76171875 0.66015625 0.55859375 0.45703125
0.35546875 0.25390625 0.15234375 0.05078125]
[ 0.05859375 0.17578125 0.29296875 0.41015625 0.52734375 0.64453125
0.76171875 0.87890625 0.87890625 0.76171875 0.64453125 0.52734375
0.41015625 0.29296875 0.17578125 0.05859375]
[ 0.05859375 0.17578125 0.29296875 0.41015625 0.52734375 0.64453125
0.76171875 0.87890625 0.87890625 0.76171875 0.64453125 0.52734375
0.41015625 0.29296875 0.17578125 0.05859375]
[ 0.05078125 0.15234375 0.25390625 0.35546875 0.45703125 0.55859375
0.66015625 0.76171875 0.76171875 0.66015625 0.55859375 0.45703125
0.35546875 0.25390625 0.15234375 0.05078125]
[ 0.04296875 0.12890625 0.21484375 0.30078125 0.38671875 0.47265625
0.55859375 0.64453125 0.64453125 0.55859375 0.47265625 0.38671875
0.30078125 0.21484375 0.12890625 0.04296875]
[ 0.03515625 0.10546875 0.17578125 0.24609375 0.31640625 0.38671875
0.45703125 0.52734375 0.52734375 0.45703125 0.38671875 0.31640625
0.24609375 0.17578125 0.10546875 0.03515625]
[ 0.02734375 0.08203125 0.13671875 0.19140625 0.24609375 0.30078125
0.35546875 0.41015625 0.41015625 0.35546875 0.30078125 0.24609375
0.19140625 0.13671875 0.08203125 0.02734375]
[ 0.01953125 0.05859375 0.09765625 0.13671875 0.17578125 0.21484375
0.25390625 0.29296875 0.29296875 0.25390625 0.21484375 0.17578125
0.13671875 0.09765625 0.05859375 0.01953125]
[ 0.01171875 0.03515625 0.05859375 0.08203125 0.10546875 0.12890625
0.15234375 0.17578125 0.17578125 0.15234375 0.12890625 0.10546875
0.08203125 0.05859375 0.03515625 0.01171875]
[ 0.00390625 0.01171875 0.01953125 0.02734375 0.03515625 0.04296875
0.05078125 0.05859375 0.05859375 0.05078125 0.04296875 0.03515625
0.02734375 0.01953125 0.01171875 0.00390625]]
[0,1,:,:]:

[[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]
[1,0,:,:]：