http://blog.csdn.net/zzzzzzz0407/article/details/69831388
网上有许多FCN网络的安装和训练教程,但却没有代码解读的详细教程,这让我这种刚刚入门深度学习的萌新不知所措;为了弄清楚FCN,不知走了多少弯路,想把它记录下来,给自己看看,也希望能帮助到那些和我一样刚刚入门的人。
以下只是小弟的一些拙见,若有错误与不足欢迎指出~
话不多说,上代码:
首先是solve.py文件,fcn没有用sh脚本去写训练文件,应该是和他自己写的surgery.py和score.py有关;里面有两个问题我希望有大牛可以帮忙解决:
1. for _ in range(50):
solver.step(2000)
发现设置完这个以后solver.prototxt中设置的 max_iter失效;
2. score.seg_tests(solver, False, test, layer=’score_sem’, gt=’sem’)
score.seg_tests(solver, False, test, layer=’score_geo’, gt=’geo’)
语义分割和几何分割的解读
#solve.py
import caffe
import surgery, score
import numpy as np
import os #os模块封装了操作系统的目录和文件操作
import sys
try:
import setproctitle
setproctitle.setproctitle(os.path.basename(os.getcwd()#获得当前路径)#返回最后的文件名)
#比如os.getcwd()获得的当前路径为/home/zhangrf/fcn,则os.path.basename()为fcn;
#setproctitle是用来修改进程入口名称,如C++中入口为main()函数
except:
pass
weights = ‘../ilsvrc-nets/vgg16-fcn.caffemodel‘ #用来fine-tune的FCN参数
vgg_weights = ‘../ilsvrc-nets/vgg16-fcn.caffemodel‘ #训练好的VGGNet参数
vgg_proto = ‘../ilsvrc-nets/VGG_ILSVRC_16_layers_deploy.prototxt‘ #VGGNet模型
# init
#caffe.set_device(int(sys.argv[1]))
#获取命令行参数,其中sys.argv[0]为文件名,argv[1]为紧随其后的那个参数
caffe.set_device(0) #GPU型号
caffe.set_mode_gpu() #用GPU模式运行
#solver.net.copy_from(weights)
solver = caffe.SGDSolver(‘solver.prototxt‘) #调用SGD(随即梯度下降)Solver方法,solver.prototxt为所需参数
vgg_net = caffe.Net(vgg_proto,vgg_weights,caffe.TRAIN) #vgg_net是原来的VGGNet模型(包括训练好的参数)
surgery.transplant(solver.net,vgg_net) #FCN模型(参数)与原来的VGGNet模型之间的转化
del vgg_net #删除VGGNet模型
# surgeries
interp_layers = [k for k in solver.net.params.keys() if ‘up‘ in k] #interp_layers为upscore层
surgery.interp(solver.net, interp_layers) #将upscore层中每层的权重初始化为双线性内核插值。
# scoring
test = np.loadtxt(‘../data/sift-flow/test.txt‘, dtype=str) #载入测试图片信息
for _ in range(50):
solver.step(2000) #每2000次训练迭代执行后面的函数
# N.B. metrics on the semantic labels are off b.c. of missing classes;
# score manually from the histogram instead for proper evaluation
score.seg_tests(solver, False, test, layer=‘score_sem‘, gt=‘sem‘) #测试图片语义特征
score.seg_tests(solver, False, test, layer=‘score_geo‘, gt=‘geo‘) #测试图片几何特征
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
surgery.py是精髓,全连接层参数到全卷积层的参数转化是靠它完成的,每一步都很巧妙,如沐春风~
#surgery.py
from __future__ import division #导入python未来支持的语言特征division(精确除法)
import caffe #导入caffe
import numpy as np #导入模块numpy并以np作为别名
def transplant(new_net#FCN, net#VGGNet, suffix=‘‘): #用于将VGGNet的参数转化给FCN(包括全连接层的参数)
"""
Transfer weights by copying matching parameters, coercing parameters of
incompatible shape, and dropping unmatched parameters.
通过复制匹配的参数,强制转换不兼容形状的参数和丢弃不匹配的参数来达到传输(转化)权重的目的;
The coercion is useful to convert fully connected layers to their
equivalent convolutional layers, since the weights are the same and only
the shapes are different.
因为权重的个数是一样的仅仅是Blob的形状不一样,所以强制转换对于将全连接层转换为等效的卷积层是有用的;
In particular, equivalent fully connected and convolution layers have shapes O x I and O x I x H x W respectively for O
outputs channels, I input channels, H kernel height, and W kernel width.
参数数量为O*I*H*W
Both `net` to `new_net` arguments must be instantiated `caffe.Net`s.
参数一对一
"""
for p in net.params: #net.params是字典形式,存放了所有的key-value,p为key
p_new = p + suffix #将p赋给p_new
if p_new not in new_net.params: #用来丢弃fc8(因为FCN中没有fc8)
print ‘dropping‘, p
continue
for i in range(len(net.params[p])):
if i > (len(new_net.params[p_new]) - 1): #感觉没啥用?
print ‘dropping‘, p, i
break
if net.params[p][i].data.shape!= new_net.params[p_new][i].data.shape:
#Blob不一样转换(这边就是全连接层和卷积层的转换,很精髓!!!)
print ‘coercing‘, p, i, ‘from‘, net.params[p][i].data.shape, ‘to‘, new_net.params[p_new][i].data.shape
else: #形状一样则直接copy
print ‘copying‘, p, ‘ -> ‘, p_new, i
new_net.params[p_new][i].data.flat = net.params[p][i].data.flat #将参数按顺序赋值(flat函数只要保证参数个数相同,不用保证数组形状完全一样)
def upsample_filt(size):
"""
Make a 2D bilinear kernel suitable for upsampling of the given (h, w) size.
上采样卷积核的制作
"""
factor = (size + 1) // 2
if size % 2 == 1:
center = factor - 1
else:
center = factor - 0.5
og = np.ogrid[:size, :size] #生成一列向量和一行向量
return (1 - abs(og[0] - center) / factor) * \ #(64*1)的列向量和(1*64)行向量相乘则得到一个64*64的数组
(1 - abs(og[1] - center) / factor)
def interp(net, layers):
"""
Set weights of each layer in layers to bilinear kernels for interpolation.
将每一层的权重设置为双线性内核插值。
"""
for l in layers:
m, k, h, w = net.params[l][0].data.shape
if m != k and k != 1:
print ‘input + output channels need to be the same or |output| == 1‘
raise
if h != w:
print ‘filters need to be square‘
raise
filt = upsample_filt(h) #初始化卷积核的参数(64*64*1)
net.params[l][0].data[range(m), range(k), :, :] = filt #这边很关键!!!只有对于对应层的那层filter有参数,其余都为0,而且有filter参数的那层还都是相等的~
#因为前一层已经是个分类器了,对分类器进行特征组合没有任何意义!所以这一层的上采样效果上而言只是对应的上采样(属于猴子还是属于猴子)
def expand_score(new_net, new_layer, net, layer):#这个函数干啥用的没看懂- -貌似solve.py里没有这个函数的调用
"""
Transplant an old score layer‘s parameters, with k < k‘ classes, into a new
score layer with k classes s.t. the first k‘ are the old classes.
"""
old_cl = net.params[layer][0].num
new_net.params[new_layer][0].data[:old_cl][...] = net.params[layer][0].data
new_net.params[new_layer][1].data[0,0,0,:old_cl][...] = net.params[layer][1].data
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
- 58
- 59
- 60
- 61
- 62
- 63
- 64
- 65
- 66
- 67
- 68
- 69
- 70
- 71
- 72
- 73
- 74
- 75
- 76
- 77
- 78
score.py是评估文件,fcn不用caffe框架自带的test测试,而是自己写了个评估文件,里面有许多评估指标。不过小弟才疏学浅,对于很多评估指标的概念完全没接触过,比方说IU的概念,overall accuracy和mean accuracy的区别,希望有大牛可以科普,这部分由于对概念的不知以及数据量的巨大,没有很好的解读~实在有愧
#score.py
from __future__ import division
import caffe
import numpy as np
import os
import sys
from datetime import datetime
from PIL import Image
def fast_hist(a, b, n):
k = (a >= 0) & (a < n)
return np.bincount(n * a[k].astype(int) + b[k], minlength=n**2).reshape(n, n)
def compute_hist(net, save_dir#False, dataset, layer=‘score‘, gt=‘label‘):
n_cl = net.blobs[layer].channels #3通道的图
if save_dir:
os.mkdir(save_dir)
hist = np.zeros((n_cl, n_cl)) #创建一个二维数组hist[3][3],元素都为0
loss = 0
for idx in dataset:
net.forward()
hist += fast_hist(net.blobs[gt].data[0, 0].flatten() #将数据拉为1列,
net.blobs[layer].data[0].argmax(0).flatten(),
n_cl)
if save_dir: #是否需要保存图片
im = Image.fromarray(net.blobs[layer].data[0].argmax(0).astype(np.uint8), mode=‘P‘)
im.save(os.path.join(save_dir, idx + ‘.png‘))
# compute the loss as well
loss += net.blobs[‘loss‘].data.flat[0]
return hist, loss / len(dataset)
def seg_tests(solver#配置文件, save_format#False, dataset#test文件, layer=‘score‘#实验输出, gt=‘label‘#真实输出):
print ‘>>>‘, datetime.now(), ‘Begin seg tests‘
solver.test_nets[0].share_with(solver.net) #将solver.net复制给solver.test_net[0]
do_seg_tests(solver.test_nets[0], solver.iter, save_format, dataset, layer, gt)
def do_seg_tests(net, iter#累计迭代次数, save_format, dataset, layer=‘score‘, gt=‘label‘):
n_cl = net.blobs[layer].channels
if save_format:
save_format = save_format.format(iter) #format函数用来格式化数据;如果save_format为TRUE,则为1
hist, loss = compute_hist(net, save_format, dataset, layer, gt)
# mean loss
print ‘>>>‘, datetime.now(), ‘Iteration‘, iter, ‘loss‘, loss #平均误差
# overall accuracy
acc = np.diag(hist).sum() / hist.sum()
print ‘>>>‘, datetime.now(), ‘Iteration‘, iter, ‘overall accuracy‘, acc
# per-class accuracy
acc = np.diag(hist) / hist.sum(1)
print ‘>>>‘, datetime.now(), ‘Iteration‘, iter, ‘mean accuracy‘, np.nanmean(acc)
# per-class IU
iu = np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist))
print ‘>>>‘, datetime.now(), ‘Iteration‘, iter, ‘mean IU‘, np.nanmean(iu)
freq = hist.sum(1) / hist.sum()
print ‘>>>‘, datetime.now(), ‘Iteration‘, iter, ‘fwavacc‘, (freq[freq > 0] * iu[freq > 0]).sum()
return hist
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
以上是我对fcn代码的一点拙见,这次经历也让我明白,未接触过的代码并不可怕,只要沉下心来,一条一条的进行调试,你也可以知其所以然,发现其中奥秘。在研究生生涯开始之际写下我的第一篇博客,希望自己能在今后的学习生涯中保持这份初心,送给自己,也送给看过这篇博客的有缘人。
时间: 2024-10-10 07:03:45