def get_per_video_mbh_data(split, suffix=''):
""" Returns the 'train' or 'test' video descriptors. The `suffix` argument
can be '_morenull' to load the data with 5296 null samples.
"""
base_path = ('/home/clear/oneata/data/trecvid12/features/'
'dense5.track15mbh.small.skip_1/')
sstats_fn = os.path.join(
base_path, 'statistics_k_256', '%s%s.dat' % (split, suffix))
labels_fn = os.path.join(
base_path, 'statistics_k_256', 'labels_%s%s.info' % (split, suffix))
info_fn = os.path.join(
base_path, 'statistics_k_256', 'info_%s%s.info' % (split, suffix))
gmm_fn = os.path.join(base_path, 'gmm', 'gmm_256')
sstats = np.fromfile(sstats_fn, dtype=np.float32)
labels = np.array([tuple_label[0]
for tuple_label in cPickle.load(open(labels_fn, 'r'))])
video_names = cPickle.load(open(info_fn, 'r'))['video_names']
# Convert sufficient statistics to Fisher vectors.
gmm = gmm_read(open(gmm_fn, 'r'))
data = FVModel.sstats_to_features(sstats, gmm)
return data, labels, video_names
def double_normalization(filenames, sstats_in , sstats_out, N, len_sstats,
gmm):
""" The slices in each sample are converted to Fisher vectors, square-
rooted, L2 normalized, and then aggregated together.
Inputs
------
filenames: list of str
The names of the files to be aggregated. Usually, this should be the
entire dataset, i.e. dataset.get_data('train')[0] +
dataset.get_data('test')[0].
sstats_in: SstatsMap instance
The sufficient statistics that we operate on.
sstats_out: SstatsMap instance
The resulting sufficient statistics.
N: int
The number of slices that are aggregated together. If N is -1 all the
slices in the clip are aggregated together.
len_sstats: int
The length of a typical sufficient statistics vector.
gmm: yael.gmm object
The Gaussian mixture model for the current sufficient statistics.
"""
assert len_sstats == gmm.k + 2 * gmm.k * gmm.d, (
"GMM and len_sstats don't match")
for filename in filenames:
if sstats_out.exists(filename):
continue
if not sstats_in.exists(filename):
print 'Not found ' + filename
continue
if sstats_in.getsize(filename) == 0:
print 'Not computed ' + filename
continue
sstats = sstats_in.read(filename).reshape((-1, len_sstats))
info = sstats_in.read_info(filename)
fv = FVModel.sstats_to_features(sstats, gmm)
fv = power_normalize(fv, 0.5)
fv = L2_normalize(fv)
agg_sstats, agg_info = _aggregate(fv, info, N)
sstats_out.write(filename, agg_sstats, info=agg_info)
def double_normalization(filenames, sstats_in, sstats_out, N, len_sstats, gmm):
""" The slices in each sample are converted to Fisher vectors, square-
rooted, L2 normalized, and then aggregated together.
Inputs
------
filenames: list of str
The names of the files to be aggregated. Usually, this should be the
entire dataset, i.e. dataset.get_data('train')[0] +
dataset.get_data('test')[0].
sstats_in: SstatsMap instance
The sufficient statistics that we operate on.
sstats_out: SstatsMap instance
The resulting sufficient statistics.
N: int
The number of slices that are aggregated together. If N is -1 all the
slices in the clip are aggregated together.
len_sstats: int
The length of a typical sufficient statistics vector.
gmm: yael.gmm object
The Gaussian mixture model for the current sufficient statistics.
"""
assert len_sstats == gmm.k + 2 * gmm.k * gmm.d, (
"GMM and len_sstats don't match")
for filename in filenames:
if sstats_out.exists(filename):
continue
if not sstats_in.exists(filename):
print 'Not found ' + filename
continue
if sstats_in.getsize(filename) == 0:
print 'Not computed ' + filename
continue
sstats = sstats_in.read(filename).reshape((-1, len_sstats))
info = sstats_in.read_info(filename)
fv = FVModel.sstats_to_features(sstats, gmm)
fv = power_normalize(fv, 0.5)
fv = L2_normalize(fv)
agg_sstats, agg_info = _aggregate(fv, info, N)
sstats_out.write(filename, agg_sstats, info=agg_info)
def get_per_video_mbh_data_given_list(list_name, **kwargs):
""" Loads the Fisher vectors corresponding to the samples found in the
list specified by `list_name`.
"""
K = kwargs.get('K', 256)
sstats_path = ('/home/clear/oneata/data/trecvid12/features'
'/dense5.track15mbh.small.skip_1/statistics_k_%d' % K)
# Default base directories.
list_base_path = kwargs.get(
'list_base_path', '/home/lear/douze/src/experiments/trecvid12/data')
cache_base_path = kwargs.get(
'cache_base_path', sstats_path)
double_norm = kwargs.get('double_norm', False)
suffix = '.double_norm' if double_norm else ''
# If this file exists load them directly.
cache_filename = os.path.join(cache_base_path, 'mbh_' + list_name + suffix + '.raw')
if os.path.exists(cache_filename):
print "Loading Fisher vectors from MBH descriptors for list %s..." % (
list_name)
with open(cache_filename, 'r') as ff:
data = np.load(ff)
labels = np.load(ff)
names = cPickle.load(ff)
return data, labels, names
D = 64
data, labels, names = [], [], []
sstats_generic_name = os.path.join(sstats_path, 'stats.tmp' + suffix, '%s.dat')
list_path = os.path.join(list_base_path, list_name + '.list')
gmm_path = os.path.join(sstats_path, 'gmm_%d' % K)
gmm = gmm_read(open(gmm_path, 'r'))
# Get descriptors for the files in list.
print "Creating cache file from list %s..." % list_name
for line in open(list_path, 'r'):
fields = line.split()
sstats_filename = fields[0]
video_name = sstats_filename.split('-')[0]
sys.stdout.write("%s\t\r" % video_name)
try:
class_name = fields[1]
except IndexError:
class_name = 'unknown'
try:
sstats = np.fromfile(sstats_generic_name % sstats_filename,
dtype=np.float32)
if double_norm:
fv = sstats
else:
fv = FVModel.sstats_to_features(sstats, gmm)
except IOError:
print ('Sufficient statistics for video %s are missing;'
'replacing with zeros.') % video_name
fv = np.zeros(K + 2 * K * D)
data.append(fv)
names.append(video_name)
labels.append(rev_label_map[class_name])
data = np.vstack(data)
labels = np.array(labels)
assert data.shape[0] == len(labels), "Data size doesn't match nr of labels"
# Cache results to file.
with open(cache_filename, 'w') as ff:
np.save(ff, data)
np.save(ff, labels)
cPickle.dump(names, ff)
return data, labels, names
