caffe下两种方法测试mnist

1. 运行%caffe_root%\examples\mnist\test_lenet.sh

caffe.exe test -model=examples/mnist/lenet_train_test.prototxt -weights=examples/mnist/lenet_iter_10000.caffemodel

其中caffe.exe在目录%caffe_root%\build\tools\Release目录下。

2. 用python写脚本，通过predict函数预测

import sys
import caffe
import struct
import time

caffe_root = r"D:\caffe"
sys.path.insert(0, caffe_root+r'\python')
test_MNIST_label = caffe_root + r"\data\mnist\t10k-labels-idx1-ubyte"

# prepare label
index = 0
print("TEST LABEL DATA:" + test_MNIST_label)
binfile = open(test_MNIST_label, 'rb')
buf = binfile.read()
magic, numImages = struct.unpack_from('>II', buf, index)
print(magic, numImages)
index += struct.calcsize('>II')

label = []
for i in range(numImages):
    tmp, = struct.unpack_from('>B', buf, index)
    index += struct.calcsize('>B')
    label.append(tmp)

# prepare image data
TEST_DATA_DIR = caffe_root + r"\examples\mnist\myTest\data\mnist_test"
input_image = []
for i in range(numImages):
    TEST_DATA_FILE = TEST_DATA_DIR + r"\test_%s.bmp"%i
    input_image.append(caffe.io.load_image(TEST_DATA_FILE, color=False))
    if 0 == i % 1000:
        print (len(input_image), TEST_DATA_FILE)

# prediction
start = time.clock()
MODEL_FILE = caffe_root + r"\examples\mnist\lenet.prototxt"
PRETRAINED = caffe_root + r"\examples\mnist\lenet_iter_10000.caffemodel"
net = caffe.Classifier(MODEL_FILE, PRETRAINED, image_dims = [28, 28])

predi = []
prediction = net.predict(input_image, oversample=False)
end = time.clock()
print("Done in time:%f(s)" % (end - start))
for i in range(numImages):
    #predi.append(prediction[i].argmax())
    predi.append(prediction[i].flatten().argsort())
    if 0 == i%1000:
        print (i, prediction[i], predi[i])

# get accelerate
count = 0
wrong = []
for i in range(numImages):
    if label[i] in predi[i][-1:]:
        count += 1
    else:
        wrong.append(i)

print ("accelerate:%.2f[count = %d]%%"%(float(count)/numImages*100, count))
print ("wrong:", wrong[:20])

* 测试结果

用1方法测试得accuracy=98.68%，用2方法测试得accuracy=99.09%。两种方法精度并不相同，而训练结束时测试得accuracy=99.08%也与之不同。难到是数据输入时有转换误差？或是train_test网络与deploy网络有所不同？或是用不同的patch会影响测试结果？原因还有待调查。