00_ex160307.md

ex160307

see http://takatakamanbou.hatenablog.com/entry/2016/03/07/221032

source

• ex160307.py using caffe.Classifier
• ex160307b.py using caffe.Net
• mnist.py is in https://gist.github.com/takatakamanbou/35d12aaa81b91d8d7e2c

experimental results

%run ex160307.py
# fnModel   =  lenet.prototxt
# fnTrained =  lenet_iter_10000.caffemodel
# conv1 (20, 1, 5, 5) (20,)
# conv2 (50, 20, 5, 5) (50,)
# ip1 (500, 800) (500,)
# ip2 (10, 500) (10,)
# useGPU =  True
# X.shape =  (10000, 28, 28, 1)
# recognition rate =  0.9902
# time =  4.34732294083

%run ex160307.py
# useGPU =  False
# time =  7.81033301353

%run ex160307b.py
# useGPU =  True
# time =  0.200743913651

%run ex160307b.py
# useGPU =  False
# time =  4.46612715721

ex160307.py

import numpy as np
import mnist
import caffe
import time


### reading the pre-trained CNN

fnModel   = 'lenet.prototxt'
fnTrained = 'lenet_iter_10000.caffemodel'
cnn = caffe.Classifier( fnModel, fnTrained )

print '# fnModel   = ', fnModel
print '# fnTrained = ', fnTrained
for ln, params in cnn.params.items():
    print '#', ln, params[0].data.shape, params[1].data.shape

useGPU = False
if useGPU:
    caffe.set_mode_gpu()
else:
    caffe.set_mode_cpu()
print '# useGPU = ', useGPU


### reading the test data

mnist = mnist.MNIST( pathMNIST = '../../150117-mnist' )
LT = 'T'
dat = mnist.getImage( LT )
lab = mnist.getLabel( LT )

w, h = mnist.ncol, mnist.nrow
nch = 1
#ndat = 100
ndat = dat.shape[0]


### predicting the class labels for the data

X = dat[:ndat, :, np.newaxis].reshape( ( -1, h, w, nch ) ) / 256
print '# X.shape = ', X.shape

ts = time.time()
Z = cnn.predict( X, oversample = False )
te = time.time()

pred = np.argmax( Z, axis = 1 )
cr = np.sum( pred == lab[:ndat] ) / float( ndat )

print '# recognition rate = ', cr
print '# time = ', te - ts

ex160307b.py

import numpy as np
import mnist
import caffe
import time


### reading the pre-trained CNN

fnModel   = 'lenet.prototxt'
fnTrained = 'lenet_iter_10000.caffemodel'
cnn = caffe.Net( fnModel, fnTrained, caffe.TEST )

print '# fnModel   = ', fnModel
print '# fnTrained = ', fnTrained
for ln, params in cnn.params.items():
    print '#', ln, params[0].data.shape, params[1].data.shape

useGPU = False
if useGPU:
    caffe.set_mode_gpu()
else:
    caffe.set_mode_cpu()
print '# useGPU = ', useGPU


### reading the test data

mnist = mnist.MNIST( pathMNIST = '../../150117-mnist' )
LT = 'T'
dat = mnist.getImage( LT )
lab = mnist.getLabel( LT )

w, h = mnist.ncol, mnist.nrow
nch = 1
#ndat = 100
ndat = dat.shape[0]


### predicting the class labels for the data

X = dat[:ndat, np.newaxis, :].reshape( ( -1, nch, h, w ) ) / 256
print '# X.shape = ', X.shape

ts = time.time()
#rv = cnn.forward_all( data = X )
rv = cnn.forward_all( **{cnn.inputs[0]:X} )
#Z = rv['prob']
Z = rv[cnn.outputs[0]]
te = time.time()

pred = np.argmax( Z, axis = 1 )
cr = np.sum( pred == lab[:ndat] ) / float( ndat )

print '# recognition rate = ', cr
print '# time = ', te - ts