做自己的视频网站,河北网站建设免费推荐,网站开发大赛发言稿,求手机网址对于fcn#xff0c;经常要使用到Deconvolution进行上采样。对于caffe使用者#xff0c;使用Deconvolution上采样#xff0c;其参数往往直接给定#xff0c;不需要通过学习获得。
给定参数的方式很有意思#xff0c;可以通过两种方式实现#xff0c;但是这两种方式并非完…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, dtypestr)for _ in range(25):solver.step(4000)score.seg_tests(solver, False, val, layerscore)
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| 1raiseif h ! w:print filters need to be squareraisefilt 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_upsampletype: Deconvolutionbottom: fc8top: fc8_upsampleparam {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_output1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,:,:]
[[ 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.]] [1,1,:,:]:
[[ 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]] 而第二种方式全部都是[0,0,:,:]这样的矩阵。
以上两种方法应该是第一种对的。因为Deconvolution 其实与卷积类似按照第一种结果才能分别单独地对map上采样而采用第二种则将会得到两个相同的map。因为综合了两个输入map的信息
因此结论 对于多个输入输出的Deconvolution,采用方法1对于单个输入的方法1,2通用。
附上Deconvolution的官方编码 说明 以上的称述有点瑕疵其实caffe已经解决了上述的问题我之前没有好好留意。 关键就在group这个选项。 如果num_output1,则填上group: c 再加上weight_filler: { type: “bilinear” }即可完成初始化。
