# d:\git\DeepLearning\caffe\loss.py
# 把loss.py文件放到caffe根目录下运行，否则，要调整相对路径及prototxt里文件路径
# 用python画loos的曲线图
# 源自 http://edu.csdn.net/course/detail/3506 视频
import sys,os;
import numpy as np;
import matplotlib.pyplot as plt;
# 引入caffe的Python接口
caffe_root='D:/git/DeepLearning/caffe/build/x64/install/';
sys.path.insert(0, caffe_root+'python');
import caffe;

# 指定计算设备
# caffe.set_mode_cpu(); # 指定使用CPU计算
caffe.set_mode_gpu(); # 指定使用GPU计算
# caffe.set_device(0) # 明确指定使用那个gpu设备
solver = caffe.SGDSolver('examples/mnist/lenet_solver.prototxt');

niter = 1000;
test_interval = 200;
train_loss = np.zeros(niter);
test_acc = np.zeros(int(np.ceil(niter / test_interval)));

# the main solver loop
for it in range(niter):
    solver.step(1) # SGD by Caffe

    # 保存训练的loss值
    train_loss[it] = solver.net.blobs['loss'].data;
    solver.test_nets[0].forward(start = 'conv1');

    if it % test_interval == 0:
        acc = solver.test_nets[0].blobs['accuracy'].data;
        print('Iteration', it, 'testing...', 'accuracy:', acc);
        test_acc[it // test_interval] = acc;

print(test_acc);
_, ax1 = plt.subplots();
ax2 = ax1.twinx();
ax1.plot(np.arange(niter), train_loss);
ax2.plot(test_interval * np.arange(len(test_acc)), test_acc, 'r');
# 设置曲线标签
ax1.set_xlabel('iteration');
ax1.set_ylabel('train loss');
ax2.set_ylabel('test accuracy');
# 显示曲线
plt.show();