如何繪制Caffe網絡訓練曲線

當我們設計好網絡結構後，在神經網絡訓練的過程中，迭代輸出的log信息中，一般包括，迭代次數，訓練損失代價，測試損失代價，測試精度等。本文提供一段示例，簡單講述如何繪制訓練曲線（training curve）。

首先看一段訓練的log輸出，網絡結構參數的那段忽略，直接跳到訓練迭代階段：

I0627 21:30:06.004370 15558 solver.cpp:242] Iteration 0, loss = 21.6953
I0627 21:30:06.004420 15558 solver.cpp:258]     Train net output #0: loss = 21.6953 (* 1 = 21.6953 loss)
I0627 21:30:06.004426 15558 solver.cpp:571] Iteration 0, lr = 0.01
I0627 21:30:28.592690 15558 solver.cpp:242] Iteration 100, loss = 13.6593
I0627 21:30:28.592730 15558 solver.cpp:258]     Train net output #0: loss = 13.6593 (* 1 = 13.6593 loss)
I0627 21:30:28.592733 15558 solver.cpp:571] Iteration 100, lr = 0.01

...

I0627 21:37:47.926597 15558 solver.cpp:346] Iteration 2000, Testing net (#0)
I0627 21:37:48.588079 15558 blocking_queue.cpp:50] Data layer prefetch queue empty
I0627 21:40:40.575474 15558 solver.cpp:414]     Test net output #0: loss = 13.07728 (* 1 = 13.07728 loss)
I0627 21:40:40.575477 15558 solver.cpp:414]     Test net output #1: loss/top-1 = 0.00226
I0627 21:40:40.575487 15558 solver.cpp:414]     Test net output #2: loss/top-5 = 0.01204
I0627 21:40:40.708261 15558 solver.cpp:242] Iteration 2000, loss = 13.1739
I0627 21:40:40.708302 15558 solver.cpp:258]     Train net output #0: loss = 13.1739 (* 1 = 13.1739 loss)
I0627 21:40:40.708307 15558 solver.cpp:571] Iteration 2000, lr = 0.01

...

I0628 01:28:47.426129 15558 solver.cpp:242] Iteration 49900, loss = 0.960628
I0628 01:28:47.426177 15558 solver.cpp:258]     Train net output #0: loss = 0.960628 (* 1 = 0.960628 loss)
I0628 01:28:47.426182 15558 solver.cpp:571] Iteration 49900, lr = 0.01
I0628 01:29:10.084050 15558 solver.cpp:449] Snapshotting to binary proto file train_net/net_iter_50000.caffemodel
I0628 01:29:10.563587 15558 solver.cpp:734] Snapshotting solver state to binary proto filetrain_net/net_iter_50000.solverstate
I0628 01:29:10.692239 15558 solver.cpp:346] Iteration 50000, Testing net (#0)
I0628 01:29:13.192075 15558 blocking_queue.cpp:50] Data layer prefetch queue empty
I0628 01:31:00.595120 15558 solver.cpp:414]     Test net output #0: loss = 0.6404232 (* 1 = 0.6404232 loss)
I0628 01:31:00.595124 15558 solver.cpp:414]     Test net output #1: loss/top-1 = 0.953861
I0628 01:31:00.595127 15558 solver.cpp:414]     Test net output #2: loss/top-5 = 0.998659
I0628 01:31:00.727577 15558 solver.cpp:242] Iteration 50000, loss = 0.680951
I0628 01:31:00.727618 15558 solver.cpp:258]     Train net output #0: loss = 0.680951 (* 1 = 0.680951 loss)
I0628 01:31:00.727623 15558 solver.cpp:571] Iteration 50000, lr = 0.0096

這是一個普通的網絡訓練輸出，含有1個loss，可以看出solver.prototxt的部分參數為：

test_interval: 2000
base_lr: 0.01
lr_policy: "step" # or "multistep"
gamma: 0.96
display: 100
stepsize: 50000 # if is "multistep", the first stepvalue is set as 50000
snapshot_prefix: "train_net/net"

當然，上面的分析，即便不理會，對下面的代碼也沒什麼影響，繪制訓練曲線本質就是文件操作，從上面的log文件中，我們可以看出：

• 對於每個出現字段] Iteration和loss =的文本行，含有訓練的迭代次數以及損失代價；
• 對於每個含有字段] Iteration和Testing net (#0)的文本行，含有測試的對應的訓練迭代次數；
• 對於每個含有字段#2:和loss/top-5的文本行，含有測試top-5的精度。

根據這些分析，就可以對文本進行如下處理：

import os
import sys
import numpy as np
import matplotlib.pyplot as plt
import math
import re
import pylab
from pylab import figure, show, legend
from mpl_toolkits.axes_grid1 import host_subplot

# read the log file
fp = open('log.txt', 'r')

train_iterations = []
train_loss = []
test_iterations = []
test_accuracy = []

for ln in fp:
  # get train_iterations and train_loss
  if '] Iteration ' in ln and 'loss = ' in ln:
    arr = re.findall(r'ion \b\d+\b,',ln)
    train_iterations.append(int(arr[0].strip(',')[4:]))
    train_loss.append(float(ln.strip().split(' = ')[-1]))
  # get test_iteraitions
  if '] Iteration' in ln and 'Testing net (#0)' in ln:
    arr = re.findall(r'ion \b\d+\b,',ln)
    test_iterations.append(int(arr[0].strip(',')[4:]))
  # get test_accuracy
  if '#2:' in ln and 'loss/top-5' in ln:
    test_accuracy.append(float(ln.strip().split(' = ')[-1]))

fp.close()

host = host_subplot(111)
plt.subplots_adjust(right=0.8) # ajust the right boundary of the plot window
par1 = host.twinx()
# set labels
host.set_xlabel("iterations")
host.set_ylabel("log loss")
par1.set_ylabel("validation accuracy")

# plot curves
p1, = host.plot(train_iterations, train_loss, label="training log loss")
p2, = par1.plot(test_iterations, test_accuracy, label="validation accuracy")

# set location of the legend, 
# 1->rightup corner, 2->leftup corner, 3->leftdown corner
# 4->rightdown corner, 5->rightmid ...
host.legend(loc=5)

# set label color
host.axis["left"].label.set_color(p1.get_color())
par1.axis["right"].label.set_color(p2.get_color())
# set the range of x axis of host and y axis of par1
host.set_xlim([-1500, 160000])
par1.set_ylim([0., 1.05])

plt.draw()
plt.show()

示例代碼中，添加了簡單的注釋，如果網絡訓練的log輸出與本中所列出的不同，只需要略微修改其中的一些參數設置，就能繪制出訓練曲線圖。

最後附上繪制出的訓練曲線圖：