Caffe学習シリーズ:モデルの各層データとパラメータの可視化


まずcaffeでcifar 10を訓練し,訓練結果モデルを保存してcaffemodelを得,その後試験画像から1枚を選択して試験を行い,可視化を行った.
In [1]:
# 
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import sys,os,caffe

In [2]:
# , 
caffe_root = '/home/bnu/caffe/' 
sys.path.insert(0, caffe_root + 'python')
os.chdir(caffe_root)
if not os.path.isfile(caffe_root + 'examples/cifar10/cifar10_quick_iter_4000.caffemodel'):
    print("caffemodel is not exist...")

In [3]:
# , 
caffe.set_mode_gpu()
net = caffe.Net(caffe_root + 'examples/cifar10/cifar10_quick.prototxt',
                caffe_root + 'examples/cifar10/cifar10_quick_iter_4000.caffemodel',
                caffe.TEST)

In [4]:
net.blobs['data'].data.shape

Out[4]:
(1, 3, 32, 32)

In [5]:
# , 
im = caffe.io.load_image('examples/images/32.jpg')
print im.shape
plt.imshow(im)
plt.axis('off')

 
(32, 32, 3)

Out[5]:
(-0.5, 31.5, 31.5, -0.5)

 
In [6]:
#  , python 
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 )
binMean=caffe_root+'examples/cifar10/mean.binaryproto'
npyMean=caffe_root+'examples/cifar10/mean.npy'
convert_mean(binMean,npyMean)

In [7]:
# blob , 
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)) #  
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

In [8]:
# 
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')

Out[8]:
(-0.5, 31.5, 31.5, -0.5)

 
In [9]:
# , 
net.forward()
[(k, v.data.shape) for k, v in net.blobs.items()]

Out[9]:
[('data', (1, 3, 32, 32)),
 ('conv1', (1, 32, 32, 32)),
 ('pool1', (1, 32, 16, 16)),
 ('conv2', (1, 32, 16, 16)),
 ('pool2', (1, 32, 8, 8)),
 ('conv3', (1, 64, 8, 8)),
 ('pool3', (1, 64, 4, 4)),
 ('ip1', (1, 64)),
 ('ip2', (1, 10)),
 ('prob', (1, 10))]

In [10]:
# 
[(k, v[0].data.shape) for k, v in net.params.items()]

Out[10]:
[('conv1', (32, 3, 5, 5)),
 ('conv2', (32, 32, 5, 5)),
 ('conv3', (64, 32, 5, 5)),
 ('ip1', (64, 1024)),
 ('ip2', (10, 64))]

In [11]:
#  , 
def show_data(data, padsize=1, padval=0):
    data -= data.min()
    data /= data.max()
    
    # force the number of filters to be square
    n = int(np.ceil(np.sqrt(data.shape[0])))
    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))
    
    # tile the filters into an image
    data = data.reshape((n, n) + data.shape[1:]).transpose((0, 2, 1, 3) + tuple(range(4, data.ndim + 1)))
    data = data.reshape((n * data.shape[1], n * data.shape[3]) + data.shape[4:])
    plt.figure()
    plt.imshow(data,cmap='gray')
    plt.axis('off')
plt.rcParams['figure.figsize'] = (8, 8)
plt.rcParams['image.interpolation'] = 'nearest'
plt.rcParams['image.cmap'] = 'gray'

In [12]:
# (filter)
show_data(net.blobs['conv1'].data[0])
print net.blobs['conv1'].data.shape
show_data(net.params['conv1'][0].data.reshape(32*3,5,5))
print net.params['conv1'][0].data.shape

 
(1, 32, 32, 32)
(32, 3, 5, 5)

 
 
In [13]:
# pooling 
show_data(net.blobs['pool1'].data[0])
net.blobs['pool1'].data.shape

Out[13]:
(1, 32, 16, 16)

 
In [14]:
# (filter)
show_data(net.blobs['conv2'].data[0],padval=0.5)
print net.blobs['conv2'].data.shape
show_data(net.params['conv2'][0].data.reshape(32**2,5,5))
print net.params['conv2'][0].data.shape

 
(1, 32, 16, 16)
(32, 32, 5, 5)

 
 
In [15]:
# (filter), 1024 
show_data(net.blobs['conv3'].data[0],padval=0.5)
print net.blobs['conv3'].data.shape
show_data(net.params['conv3'][0].data.reshape(64*32,5,5)[:1024])
print net.params['conv3'][0].data.shape

 
(1, 64, 8, 8)
(64, 32, 5, 5)

 
 
In [16]:
# 
show_data(net.blobs['pool3'].data[0],padval=0.2)
print net.blobs['pool3'].data.shape

 
(1, 64, 4, 4)

 
In [17]:
#  
feat = net.blobs['prob'].data[0]
print feat
plt.plot(feat.flat)

 
[  5.21440245e-03   1.58397834e-05   3.71246301e-02   2.28459597e-01
   1.08315737e-03   7.17785358e-01   1.91939052e-03   7.67927198e-03
   6.13298907e-04   1.05107691e-04]

Out[17]:
[]

 
 
入力された結果と図示から見ると、最大確率は7.17785358 e-01であり、第5クラス(0から)に属する.cifar 10の10種類の名前と比較します.
airplane、automobile、bird、cat、deer、dog、frog、horse、ship、truck
テスト結果からdogと判断します.テストに間違いありません.