def __init__(self,
batch_size,
n_classes,
feature_dim,
feature_name,
is_training,
is_shuffle=True):
"""
Initialization
"""
self.batch_size = batch_size
self.is_training = is_training
self.n_classes = n_classes
self.feature_dim = feature_dim
self.feature_name = feature_name
self.is_shuffle = is_shuffle
self.dataset_name = 'charades'
# load annotation
root_path = '/content/drive/My Drive/Charades_timeception/Charades'
annotation_path = '%s/video_annotation.pkl' % (root_path)
if self.is_training:
(video_names, y, _, _) = utils.pkl_load(annotation_path)
else:
(_, _, video_names, y) = utils.pkl_load(annotation_path)
# in case of single label classification, debinarize the labels
if config.cfg.MODEL.CLASSIFICATION_TYPE == 'sl':
y = utils.debinarize_label(y)
# in any case, make sure target is float
y = y.astype(np.float32)
# convert relative to root pathes
feats_path = np.array([
'%s/%s/%s.pkl' % (root_path, feature_name, p) for p in video_names
])
n_samples = len(y)
self.n_samples = n_samples
self.n_batches = utils.calc_num_batches(n_samples, batch_size)
self.feats_path = feats_path
self.y = y
# shuffle the data
if self.is_shuffle:
self.__shuffle()
def __init__(self,
batch_size,
n_classes,
feature_dim,
feature_name,
is_training,
is_shuffle=True):
"""
Initialization
"""
self.batch_size = batch_size
self.is_training = is_training
self.n_classes = n_classes
self.feature_dim = feature_dim
self.feature_name = feature_name
self.is_shuffle = is_shuffle
self.dataset_name = 'charades'
# load annotation
root_path = '/content/drive/My Drive/Charades_timeception/Charades'
annotation_path = '%s/video_annotation.pkl' % (root_path)
#/content/timeception/data/charades/annotation/video_annotation.pkl
if self.is_training:
(video_names, y, _, _) = utils.pkl_load(annotation_path)
else:
(_, _, video_names, y) = utils.pkl_load(annotation_path)
# convert relative to root pathes
feats_path = np.array([
'%s/%s/%s.pkl' % (root_path, feature_name, p) for p in video_names
])
n_samples = len(y)
self.n_samples = n_samples
self.n_batches = utils.calc_num_batches(n_samples, batch_size)
self.feats_path = feats_path
self.y = y
# shuffle the data
if self.is_shuffle:
self.__shuffle()
def __getitem__(self, index):
"""
Generate one batch of data
"""
y = self.y[index]
p = self.feats_path[index]
x = utils.pkl_load(p) # (T, H, W, C)
# convert to channel last
x = np.transpose(x, (3, 0, 1, 2)) # (T, H, W, C)
return x, y
def __load_features(self, params):
idx_video = params[0]
feats_path = params[1]
video_name = feats_path.split('/')[-1]
try:
# load feature from file
feats = utils.pkl_load(feats_path)
n_feats = len(feats)
assert n_feats == self.__n_frames_per_video, 'Sorry, wrong number of frames, expected: %d, got: %d' % (
self.__n_frames_per_video, n_feats)
self.__batch_features[idx_video] = feats
except Exception as exp:
print('\nSorry, error in loading feature %s' % (feats_path))
print(exp)
def __getitem__(self, index):
"""
Generate one batch of data.
"""
idx_start = index * self.batch_size
idx_stop = (index + 1) * self.batch_size
y = self.y[idx_start:idx_stop]
feats_path = self.feats_path[idx_start:idx_stop]
n_items = len(feats_path)
x_shape = tuple([n_items] + list(self.feature_dim))
x = np.zeros(x_shape, dtype=np.float32)
# loop of feature pathes and load them
for idx, p in enumerate(feats_path):
x[idx] = utils.pkl_load(p)
return x, y