def fuzz_config(config):
print config
imgs = lfw.get_images('float32', preproc=config['preproc'])
try:
mm = lfw.slm_memmap(config['slm'], imgs,
name='_test_thoreano_fuzz')
except InvalidDescription:
return
# -- evaluate a single example
a = mm[0]
# -- evaluate a batch
b = mm[10:14]
# --re-evaluate
ra = mm[0]
rb = mm[10:14]
assert np.all(a == ra)
assert np.all(b == rb)
lfw.delete_memmap(mm)
def test(self):
all_test_pair_inds = self.test_pair_inds
try:
x = scipy.io.loadmat(self.matfile)
except:
raise nose.SkipTest()
x_train_fnames_0 = map(string.strip, map(str, x['train_fnames'][::2]))
x_train_fnames_1 = map(string.strip, map(str, x['train_fnames'][1::2]))
x_train_labels = map(int, x['train_labels'])
dataset = skdata.lfw.Aligned()
train_fnames_0, train_fnames_1, train_labels = dataset.raw_verification_task(split='DevTrain')
train_labels = 2*train_labels - 1
train_fnames_0 = ['/'.join(_f.split('/')[-2:]) for _f in map(str, train_fnames_0)]
train_fnames_1 = ['/'.join(_f.split('/')[-2:]) for _f in map(str, train_fnames_1)]
fnames, _l = dataset.raw_classification_task()
fnames = map(str, ['/'.join(_f.split('/')[-2:]) for _f in fnames])
pairs = verification_pairs(split='DevTrain')
for _ind in all_test_pair_inds:
assert train_fnames_0[_ind] == x_train_fnames_0[_ind]
assert train_fnames_1[_ind] == x_train_fnames_1[_ind]
assert fnames[pairs[0][_ind]] == train_fnames_0[_ind]
assert fnames[pairs[1][_ind]] == train_fnames_1[_ind]
assert pairs[2][_ind] == train_labels[_ind], str((pairs[2][_ind], train_labels[_ind]))
assert pairs[2][_ind] == x_train_labels[_ind], str((pairs[2][_ind], x_train_labels[_ind]))
pairs = (pairs[0][all_test_pair_inds],
pairs[1][all_test_pair_inds],
pairs[2][all_test_pair_inds])
x_kern = x['kernel_traintrain'][all_test_pair_inds][:, all_test_pair_inds]
namebase = self.namebase
image_features = slm_memmap(self.desc,
get_images('float32'),
namebase + '_img_feat')
#print np.asarray(image_features[:4])
pf_cache, matches = pairs_memmap(pairs,
image_features,
comparison_name=self.comparison,
name=namebase + '_pairs_DevTrain')
pair_features = np.asarray(pf_cache)
delete_memmap(image_features)
pairs_cleanup((pf_cache, None))
assert (pairs[2] == matches).all()
#XXX: normalize the pair_features here in some way? do we have to compute
#all the features?
def normalize(_f):
_f = _f - _f.mean(0)
fstd = _f.std(0)
fstd[fstd == 0] = 1
_f = _f / fstd
return _f
pair_features = normalize(pair_features)
kern = np.dot(pair_features, pair_features.T)
normdiff = np.abs((kern - x_kern) / x_kern)
normdiffmax = normdiff.max()
if normdiffmax > .1:
print 'kern', kern
print 'x_kern', x_kern
assert 0, ('too much error: %s' % normdiffmax)