def set_up(self):
logging.info('Loading jeffnet...')
# We will simply instantiate a jeffnet manually
data_root = '/u/vis/common/deeplearning/models/'
self._net = JeffNet(data_root + 'imagenet.jeffnet.epoch90',
data_root + 'imagenet.jeffnet.meta')
logging.info('Jeffnet loaded.')
self._randproj = None
class DecafnetMapper(mapreducer.BasicMapper):
"""The ImageNet Compute mapper. The input value would be a synset name.
"""
def set_up(self):
logging.info('Loading jeffnet...')
# We will simply instantiate a jeffnet manually
data_root = '/u/vis/common/deeplearning/models/'
self._net = JeffNet(data_root + 'imagenet.jeffnet.epoch90',
data_root + 'imagenet.jeffnet.meta')
logging.info('Jeffnet loaded.')
self._randproj = None
def map(self, key, value):
if type(value) is not str:
value = str(value)
if FLAGS.single:
files = [os.path.join(FLAGS.input_folder, value)]
else:
files = glob.glob(os.path.join(FLAGS.input_folder, value,
'*.JPEG'))
files.sort()
logging.info('%s: %d files', value, len(files))
features = None
for i, f in enumerate(files):
try:
img = io.imread(f)
logging.info(f)
_ = self._net.classify(img)
feat = self._net.feature(FLAGS.feature_name)
dim = np.prod(feat.shape[1:])
if FLAGS.randprojection > 0:
if self._randproj is None:
np.random.seed(1701)
self._randproj = np.random.randn(
dim, FLAGS.randprojection).astype(FEATURE_DTYPE)
self._randproj /= dim
feat = np.dot(feat.reshape(feat.shape[0], dim),
self._randproj)
if features is None:
features = np.zeros((len(files) * 10, ) + feat.shape[1:],
dtype=FEATURE_DTYPE)
features[i * 10:(i + 1) * 10] = feat
except IOError:
# we ignore the exception (maybe the image is corrupted or
# pygist has some bugs?)
print f, Exception, e
outname = str(uuid.uuid4()) + '.npy'
try:
os.makedirs(FLAGS.output_folder)
except OSError:
pass
np.save(os.path.join(FLAGS.output_folder, outname), features)
yield value, outname
class DecafnetMapper(mapreducer.BasicMapper):
"""The ImageNet Compute mapper. The input value would be a synset name.
"""
def set_up(self):
logging.info('Loading jeffnet...')
# We will simply instantiate a jeffnet manually
data_root = '/u/vis/common/deeplearning/models/'
self._net = JeffNet(data_root+'imagenet.jeffnet.epoch90',
data_root+'imagenet.jeffnet.meta')
logging.info('Jeffnet loaded.')
self._randproj = None
def map(self, key, value):
if type(value) is not str:
value = str(value)
if FLAGS.single:
files = [os.path.join(FLAGS.input_folder, value)]
else:
files = glob.glob(os.path.join(FLAGS.input_folder, value, '*.JPEG'))
files.sort()
logging.info('%s: %d files', value, len(files))
features = None
for i, f in enumerate(files):
try:
img = io.imread(f)
logging.info(f)
_ = self._net.classify(img)
feat = self._net.feature(FLAGS.feature_name)
dim = np.prod(feat.shape[1:])
if FLAGS.randprojection > 0:
if self._randproj is None:
np.random.seed(1701)
self._randproj = np.random.randn(dim, FLAGS.randprojection).astype(FEATURE_DTYPE)
self._randproj /= dim
feat = np.dot(feat.reshape(feat.shape[0], dim), self._randproj)
if features is None:
features = np.zeros((len(files) * 10,) + feat.shape[1:],
dtype = FEATURE_DTYPE)
features[i*10:(i+1)*10] = feat
except IOError:
# we ignore the exception (maybe the image is corrupted or
# pygist has some bugs?)
print f, Exception, e
outname = str(uuid.uuid4()) + '.npy'
try:
os.makedirs(FLAGS.output_folder)
except OSError:
pass
np.save(os.path.join(FLAGS.output_folder, outname), features)
yield value, outname
def set_up(self):
logging.info('Loading jeffnet...')
# We will simply instantiate a jeffnet manually
data_root = '/u/vis/common/deeplearning/models/'
self._net = JeffNet(data_root+'imagenet.jeffnet.epoch90',
data_root+'imagenet.jeffnet.meta')
logging.info('Jeffnet loaded.')
self._randproj = None
import argparse import sys import numpy as np try: import matplotlib.pyplot as pylab except: import scipy.misc as pylab from decaf.scripts.jeffnet import JeffNet from decaf.util import smalldata data_root = '/home/rodner/data/deeplearning/models/' if len(sys.argv)<2:\ img = smalldata.lena() else: imgfn = sys.argv[1] img = pylab.imread(imgfn) net = JeffNet(data_root+'imagenet.jeffnet.epoch90', data_root+'imagenet.jeffnet.meta') print "Classify image" for i in range(5): scores = net.classify(img, center_only=True) print "Okay" #print net.top_k_prediction(scores, 10)[1]
parser.add_argument('img_path_patterns',metavar='pattern', nargs='+',action='append',help='A path pattern for the input images. For example, ./lena.jpg will select lena.jpg from the current directory and ./*.jpg will select all images with the extension jpg in the current directory. If --file-list-mode is set, pattern refers to a path to a text file containing the paths to the images.')
parser.add_argument('--model',nargs='?',default=data_root+'imagenet.jeffnet.epoch90',action='store',help='Path to a model of a convulutional network.')
parser.add_argument('--meta',nargs='?',default=data_root+'imagenet.jeffnet.meta',action='store',help='Path to the meta data of the convulutional network.')
parser.add_argument('--out',nargs='?',default='feature_out.csv',action='store',help='File to store the DeCAF features. The output is a comma separated values (CSV) file. Each line corresponds to one input image.')
parser.add_argument('--layer',nargs='?',default='fc7_neuron_cudanet_out',action='store',help='The features of the layer specified here will be used.')
parser.add_argument('--all-features',action='store_true',help='Wheather to store all features or just for the center part of the image.')
parser.add_argument('--file-list-mode',action='store_true',help='If you set this flag, you can use a file list as input.')
parser.add_argument('--silent',action='store_true',help='Suppress output to console.')
parser.add_argument('--warp',action='store_true',help='Warp the input image to 227x227 and use this directly as input.')
args = parser.parse_args()
# Early check, if an image path is passed
args.img_path_patterns
net = JeffNet(args.model, args.meta)
csv_file=open(args.out,"w");
csv_writer = csv.writer(csv_file,delimiter=',', quoting=csv.QUOTE_NONE)
for (i,img_list) in enumerate(args.img_path_patterns):
for img_path in img_list:
if args.file_list_mode:
with open(img_path) as f:
content = f.readlines()
for img_name in content:
writeFeatures(net,img_name.strip('\n'),args.all_features,args.layer,csv_writer,args.silent,args.warp,args.out+'.mat')
else:
writeFeatures(net,img_path,args.all_features,args.layer,csv_writer,args.silent,args.warp,args.out+'.mat')
