Caffe的Python接口进行Cifar10可视化

由于我上一篇想调可视化的内容，不知道为什么调不起来，这个先转载过来保存一下。地址

根据训练好的cifar10数据的model，从测试图片中选出一张进行测试，并进行网络模型、卷积结果及参数可视化
注意：本文中代码运行在windows+ipython notebook下，已事先配置好caffe的python接口

导入必需的包

123456789

import numpy as npimport matplotlib.pyplot as pltimport matplotlib.image as mpimgimport caffe%matplotlib inlineplt.rcParams['figure.figsize'] = (8, 8)plt.rcParams['image.interpolation'] = 'nearest'plt.rcParams['image.cmap'] = 'gray'

载入网络模型

123456

# 载入模型，并显示各层数据信息caffe.set_mode_gpu()net = caffe.Net('examples/cifar10/cifar10_quick.prototxt',                'examples/cifar10/cifar10_quick_iter_5000.caffemodel.h5',                caffe.TEST)[(k, v.data.shape) for k, v in net.blobs.items()]

[('data', (1L, 3L, 32L, 32L)), ('conv1', (1L, 32L, 32L, 32L)), ('pool1', (1L, 32L, 16L, 16L)), ('conv2', (1L, 32L, 16L, 16L)), ('pool2', (1L, 32L, 8L, 8L)), ('conv3', (1L, 64L, 8L, 8L)), ('pool3', (1L, 64L, 4L, 4L)), ('ip1', (1L, 64L)), ('ip2', (1L, 10L)), ('prob', (1L, 10L))]

可视化网络模型

使用GraphViz+Caffe的draw_net.py来可视化网络模型

123

Rem 运行以下命令前必需先安装配置GraphVizRem --rankdir参数为网络方向,BT代表图片上网络从底至顶绘出python ./Build/x64/Release/pycaffe/draw_net.py examples/cifar10/cifar10_quick_train_test.prototxt examples/cifar10/cifar-quick.png --rankdir=BT

Drawing net to examples/cifar10/cifar-quick.png

1234

#显示模型图片net_im = mpimg.imread('examples/cifar10/cifar-quick.png')plt.imshow(net_im)plt.axis('off')

(-0.5, 904.5, 2079.5, -0.5)

加载测试图片

12345

#加载测试图片，并显示im = caffe.io.load_image('examples/cifar10/cat.jpg')print im.shapeplt.imshow(im)plt.axis('off')

(1200L, 1600L, 3L)(-0.5, 1599.5, 1199.5, -0.5)

转换均值

12345678

#　编写一个函数，将二进制的均值转换为python的均值def convert_mean(binMean,npyMean):    blob = caffe.proto.caffe_pb2.BlobProto()    bin_mean = open(binMean, 'rb' ).read()    blob.ParseFromString(bin_mean)    arr = np.array( caffe.io.blobproto_to_array(blob) )    npy_mean = arr[0]    np.save(npyMean, npy_mean )

1234

# 调用函数转换均值binMean='examples/cifar10/mean.binaryproto'npyMean='examples/cifar10/mean.npy'convert_mean(binMean,npyMean)

将图片载入Blob

12345678

#将图片载入blob中,并减去均值transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})transformer.set_transpose('data', (2,0,1))transformer.set_mean('data', np.load(npyMean).mean(1).mean(1)) # 减去均值transformer.set_raw_scale('data', 255)  transformer.set_channel_swap('data', (2,1,0))net.blobs['data'].data[...] = transformer.preprocess('data',im)inputData=net.blobs['data'].data

12345678

#显示减去均值前后的数据plt.figure()plt.subplot(1,2,1),plt.title("origin")plt.imshow(im)plt.axis('off')plt.subplot(1,2,2),plt.title("subtract mean")plt.imshow(transformer.deprocess('data', inputData[0]))plt.axis('off')

(-0.5, 31.5, 31.5, -0.5)

编写用于参数/卷积结果可视化的函数

12345678910111213141516171819

#　编写一个函数，用于显示各层数据def show_data(data, padsize=1, padval=0):    # data归一化    data -= data.min()    data /= data.max()        # 根据data中图片数量data.shape[0]，计算最后输出时每行每列图片数n    n = int(np.ceil(np.sqrt(data.shape[0])))    # padding = ((图片个数维度的padding),(图片高的padding), (图片宽的padding), ....)    padding = ((0, n ** 2 - data.shape[0]), (0, padsize), (0, padsize)) + ((0, 0),) * (data.ndim - 3)    data = np.pad(data, padding, mode='constant', constant_values=(padval, padval))        # 先将padding后的data分成n*n张图像    data = data.reshape((n, n) + data.shape[1:]).transpose((0, 2, 1, 3) + tuple(range(4, data.ndim + 1)))    # 再将（n, W, n, H）变换成(n*w, n*H)    data = data.reshape((n * data.shape[1], n * data.shape[3]) + data.shape[4:])    plt.figure()    plt.imshow(data,cmap='gray')    plt.axis('off')

可视化各层数据

123456

# 运行模型并显示第一个卷积层的输出数据和权值（filter）net.forward()print net.blobs['conv1'].data[0].shapeshow_data(net.blobs['conv1'].data[0])print net.params['conv1'][0].data.shapeshow_data(net.params['conv1'][0].data.reshape(32*3,5,5))

(32L, 32L, 32L)(32L, 3L, 5L, 5L)

123

# 显示第一次pooling后的输出数据show_data(net.blobs['pool1'].data[0])net.blobs['pool1'].data.shape

(1L, 32L, 16L, 16L)

12345

# 显示第二次卷积后的输出数据以及相应的权值（filter）show_data(net.blobs['conv2'].data[0],padval=0.5)print net.blobs['conv2'].data.shapeshow_data(net.params['conv2'][0].data.reshape(32**2,5,5))print net.params['conv2'][0].data.shape

(1L, 32L, 16L, 16L)(32L, 32L, 5L, 5L)

12345

# 显示第三次卷积后的输出数据以及相应的权值（filter）,取前１024个进行显示show_data(net.blobs['conv3'].data[0],padval=0.5)print net.blobs['conv3'].data.shapeshow_data(net.params['conv3'][0].data.reshape(64*32,5,5)[:1024])print net.params['conv3'][0].data.shape

(1L, 64L, 8L, 8L)(64L, 32L, 5L, 5L)

123

# 显示第三次池化后的输出数据show_data(net.blobs['pool3'].data[0],padval=0.2)print net.blobs['pool3'].data.shape

(1L, 64L, 4L, 4L)

1234

# 最后一层输入属于某个类的概率feat = net.blobs['prob'].data[0]print featplt.plot(feat.flat)

[ 0.00170287  0.00115923  0.0225699   0.60395384  0.00453733  0.14171894  0.00307363  0.01260873  0.15008588  0.05858969][<matplotlib.lines.Line2D at 0x3bd38080>]

与cifar10中的10种类型名称进行对比：

airplane、automobile、bird、cat、deer、dog、frog、horse、ship、truck

根据测试结果，判断为Cat。