import numpy as np
import matplotlib.pyplot as plt
import sys,os
caffe_root = '/home/tyd/caffe/'
sys.path.insert(0,caffe_root+'python')
import caffe
#caffe.set_device(0)
caffe.set_mode_cpu()
solver = caffe.SGDSolver('/home/tyd/caffe/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
#store the train loss
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')
ax1.set_xlabel('iteration')
ax1.set_ylabel('train loss')
ax2.set_ylabel(test accuracy)
plt.show()