Caffe---Pycaffe 绘制loss和accuracy曲线

Caffe---Pycaffe 绘制loss和accuracy曲线

　　《Caffe自带工具包---绘制loss和accuracy曲线》：可以看出使用caffe自带的工具包绘制loss曲线和accuracy曲线十分的方便简单，而这种方法看起来貌似只能分开绘制曲线，无法将两种曲线绘制在一张图上。但，我们有时为了更加直观的观察训练loss和测试loss，往往需要将这两种曲线绘制在一张图上。那如何解决呢？python接口，Pycaffe可以实现将这两种曲线绘制在一张图上。

　　目前，我知道的知识面中，Pycaffe有两种方式可以画出loss和accuracy曲线：一种是，根据之前博文里保存的训练日志.log文件，Pycaffe进行绘制；另一种是，Pycaffe自己进行训练，训练完后自动出图。

目录

1，Pycaffe---只绘图（前提已有训练的.log文件）

2，Pycaffe---训练+绘图

正文

1，Pycaffe---只绘图

　　这种方式属于绘制训练后的loss和accuracy曲线，绘图所需的信息，利用python从log日志里面获取。一般步骤：train_xxx_log.sh文件训练，然后保存xxx.log文件，手动将xxx.log文件名改成xxx.txt，然后用Pycaffe绘图。

　　（1）示例，写一个FromLogTxt_draw_LossAccuracy.py文件（参考：https://blog.csdn.net/u014593748/article/details/76152622）：

------------------------------------------------------------------------------------

# -*- coding: utf-8 -*-
#!/usr/bin/env python

import sys
import re
import matplotlib.pyplot as plt
import numpy as np

in_log_path='/home/wp/caffe/myself/road/Log/record_train_road_log.txt'  #输入日志文件的位置
out_fig_path='/home/wp/caffe/myself/road/Log/record_train_road_log.jpg' #输出图片的位置
f=open(in_log_path,'r')
accuracy=[]
train_loss=[]
test_loss=[]

max_iter=0
test_iter=0
test_interval=0
display=0
target_str=['accuracy = ','Test net output #1: loss = ','Train net output #0: loss = ',
            'max_iter: ','test_iter: ','test_interval: ','display: ']
while True:
    line=f.readline()
    # print len(line),line
    if len(line)<1:
        break
    for i in range(len(target_str)):
        str=target_str[i]
        idx = line.find(str)
        if idx != -1:
            num=float(line[idx + len(str):idx + len(str) + 5])
            if(i==0):
                accuracy.append(num)
            elif(i==1):
                test_loss.append(num)
            elif(i==2):
                train_loss.append(num)
            elif(i==3):
                max_iter=float(line[idx + len(str):])
            elif(i==4):
                test_iter=float(line[idx + len(str):])
            elif(i==5):
                test_interval=float(line[idx + len(str):])
            elif(i==6):
                display=float(line[idx + len(str):])
            else:
                pass
f.close()
# print test_iter
# print max_iter
# print test_interval
# print len(accuracy),len(test_loss),len(train_loss)

_,ax1=plt.subplots()
ax2=ax1.twinx()
#绘制train_loss曲线图像，颜色为绿色'g'
ax1.plot(display*np.arange(len(train_loss)),train_loss,color='g',label='train loss',linestyle='-')

#绘制test_loss曲线图像，颜色为黄色'y'
ax1.plot(test_interval*np.arange(len(test_loss)),test_loss,color='y',label='test loss',linestyle='-')

#绘制accuracy曲线图像，颜色为红色'r'
ax2.plot(test_interval*np.arange(len(accuracy)),accuracy,color='r',label='accuracy',linestyle='-')

ax1.legend(loc=(0.7,0.8))  #使用二元组(0.7,0.8)定义标签位置
ax2.legend(loc=(0.7,0.72))
ax1.set_xlabel('iteration')#设置X轴标签
ax1.set_ylabel('loss')     #设置Y1轴标签
ax2.set_ylabel('accuracy') #设置Y2轴标签
plt.savefig(out_fig_path,dpi=100) #将图像保存到out_fig_path路径中，分辨率为100
plt.show()                 #显示图片

------------------------------------------------------------------------------------

# python FromLogTxt_draw_LossAccuracy.py

　　

　　（2）或者，在shell下根据XXX.log文件，提取loss值以及accuracy值，保存到test_loss.txt，train_loss.txt，test_acc.txt。参考https://blog.csdn.net/m0_37477175/article/details/78431717。

终端下，进入相应的目录下：cat train_road_20180525.log | grep "Train net output" | awk '{print $11}'，如下：

python+pandas来间接绘图！首先我们查看一下网络训练参数：

#训练每迭代500次，进行一次预测
test_interval: 500
#每经过100次迭代，在屏幕打印一次运行log
display: 100
#最大迭代次数
max_iter: 10000

#!/usr/bin/env python
# -*- coding:utf-8 -*-

"""
Created on Tue Oct 17  2017

@author: jack wang
This program for visualize the loss and accuracy

"""
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

train_interval = 100  #display = 100
test_interval = 500
max_iter = 10000

def loadData(file):
    dataMat = []
    fr = open(file)
    for line in fr.readlines():
        lineA = line.strip().split()

        dataMat.append(float(lineA[0]))
    return dataMat

trainloss = loadData('trainloss.txt')
testloss = loadData('testloss.txt')

trainLoss = pd.Series(trainloss, index = range(0,max_iter,100))
testLoss = pd.Series(testloss, index = range(0,max_iter+500,500))

fig = plt.figure()
plt.plot(trainLoss)
plt.plot(testLoss)
plt.xlabel(u"iter")
plt.ylabel(u"loss")
plt.title(u"trainloss vs testloss")
plt.legend((u'trainloss', u'testloss'),loc='best')
plt.show()
testacc = loadData('testacc.txt')
testAcc = pd.Series(testacc, index = range(0,max_iter+500,500))
plt.plot(testAcc)
plt.show()

　　注明：这种方法，个人没有顺利的做下来，留作下次继续研究。

2，Pycaffe---训练+绘图

　　这种方式属于绘制训练过程的loss和accuracy曲线。一般步骤：Pycaffe自己写一个文件，里面既能训练网络，又能保存信息，然后绘制图。示例，写一个Pycaffe_TrainTest_then_loss_accuracy.py（参考https://www.cnblogs.com/denny402/p/5686067.html）：

------------------------------------------------------------------------------------

# -*- coding: utf-8 -*-
#!/usr/bin/env python

from pylab import *  
import matplotlib.pyplot as plt  
import caffe

solver = caffe.SGDSolver('/home/wp/caffe/myself/road/prototxt_files/solver.prototxt')

niter = 200  
display= 10    
test_iter = 200  
test_interval =100  
 
train_loss = zeros(ceil(niter * 1.0 / display))   
test_loss = zeros(ceil(niter * 1.0 / test_interval))  
test_acc = zeros(ceil(niter * 1.0 / test_interval))  
 
solver.step(1)  
 
_train_loss = 0; _test_loss = 0; _accuracy = 0   
for it in range(niter):  
    solver.step(1)  
    _train_loss += solver.net.blobs['loss'].data  
    if it % display == 0:  
        train_loss[it // display] = _train_loss / display  
        _train_loss = 0  
 
    if it % test_interval == 0:  
        for test_it in range(test_iter):  
            solver.test_nets[0].forward()  
            _test_loss += solver.test_nets[0].blobs['loss'].data  
            _accuracy += solver.test_nets[0].blobs['accuracy'].data   
        test_loss[it / test_interval] = _test_loss / test_iter   
        test_acc[it / test_interval] = _accuracy / test_iter  
        _test_loss = 0  
        _accuracy = 0  
 
print '\nplot the train loss and test accuracy\n'  
_, ax1 = plt.subplots()  
ax2 = ax1.twinx()  
 
ax1.plot(display * arange(len(train_loss)), train_loss, 'g')  
ax1.plot(test_interval * arange(len(test_loss)), test_loss, 'y')  
ax2.plot(test_interval * arange(len(test_acc)), test_acc, 'r')  
 
ax1.set_xlabel('iteration')  
ax1.set_ylabel('loss')  
ax2.set_ylabel('accuracy')  
plt.show()
plt.pause(0.000001)

------------------------------------------------------------------------------------

# cd caffe

#python Pycaffe_TrainTest_then_loss_accuracy.py

# .py这里放在caffe目录下，不在caffe目录下修改相应的路径即可。

# 代码含义，根据参考文章理解，.py文件中少出现汉字注释，否则会出现[ python: can't open file 'Pycaffe_TrainTest_then_loss_accuracy.py002.py': [Errno 2] No such file or directory ]这样的提示。
最后，训练完后，就会出现loss和accuracy曲线图了。设置niter = 200，快速迭代出图。

附，相关代码说明：

------------------------------------------------------------------------------------

# -*- coding: utf-8 -*-
#加载必要的库  
import matplotlib.pyplot as plt
import caffe
caffe.set_device(0)
caffe.set_mode_gpu()
# 使用SGDSolver，即随机梯度下降算法
solver = caffe.SGDSolver('/home/xxx/mnist/solver.prototxt')  

# 等价于solver文件中的max_iter，即最大解算次数
niter = 9380
# 每隔100次收集一次数据
display= 100  

# 每次测试进行100次解算，10000/100
test_iter = 100
# 每500次训练进行一次测试（100次解算），60000/64
test_interval =938  

#初始化
train_loss = zeros(ceil(niter * 1.0 / display))
test_loss = zeros(ceil(niter * 1.0 / test_interval))
test_acc = zeros(ceil(niter * 1.0 / test_interval))  

# iteration 0，不计入
solver.step(1)  

# 辅助变量
_train_loss = 0; _test_loss = 0; _accuracy = 0
# 进行解算
for it in range(niter):
    # 进行一次解算
    solver.step(1)
    # 每迭代一次，训练batch_size张图片
    _train_loss += solver.net.blobs['SoftmaxWithLoss1'].data
    if it % display == 0:
        # 计算平均train loss
        train_loss[it // display] = _train_loss / display
        _train_loss = 0  

    if it % test_interval == 0:
        for test_it in range(test_iter):
            # 进行一次测试
            solver.test_nets[0].forward()
            # 计算test loss
            _test_loss += solver.test_nets[0].blobs['SoftmaxWithLoss1'].data
            # 计算test accuracy
            _accuracy += solver.test_nets[0].blobs['Accuracy1'].data
        # 计算平均test loss
        test_loss[it / test_interval] = _test_loss / test_iter
        # 计算平均test accuracy
        test_acc[it / test_interval] = _accuracy / test_iter
        _test_loss = 0
        _accuracy = 0  

# 绘制train loss、test loss和accuracy曲线
print '\nplot the train loss and test accuracy\n'
_, ax1 = plt.subplots()
ax2 = ax1.twinx()  

# train loss -> 绿色
ax1.plot(display * arange(len(train_loss)), train_loss, 'g')
# test loss -> 黄色
ax1.plot(test_interval * arange(len(test_loss)), test_loss, 'y')
# test accuracy -> 红色
ax2.plot(test_interval * arange(len(test_acc)), test_acc, 'r')  

ax1.set_xlabel('iteration')
ax1.set_ylabel('loss')
ax2.set_ylabel('accuracy')
plt.show()  
------------------------------------------------------------------------------------