def build_net():
global net
my_id = multiprocessing.current_process()._identity[0] \
if args.num_worker > 1 else 1
if gpu_list is None:
net = CaffeNet(args.net_proto, args.net_weights, my_id - 1)
else:
net = CaffeNet(args.net_proto, args.net_weights, gpu_list[my_id - 1])
def __init__(self,
models=list(),
total_norm_weights=None,
score_name='fc-action',
dev_id=0,
data_manager=None):
"""
Contruct an action classifier
Args:
models: list of tuples in the form of
(model_proto, model_params, model_fusion_weight, input_type, data_augmentation).
input_type is: 0-RGB, 1-Optical flow.
data_augmentation indicates whether we will augment the data via oversampling
total_norm_weights: sum of all model_fusion_weights when normalization is wanted, otherwise use None
"""
self.__net_vec = [CaffeNet(x[0], x[1], dev_id) for x in models]
self.__net_weights = [float(x[2]) for x in models]
if total_norm_weights is not None:
s = sum(self.__net_weights)
self.__net_weights = [x / s for x in self.__net_weights]
self.__input_type = [x[3] for x in models]
self.__data_aug = [x[4] for x in models]
self.__num_net = len(models)
# whether we should prepare flow stack
self.__need_flow = max(self.__input_type) > 0
# the name in the proto for action classes
self.__score_name = score_name
self.__data_manager = data_manager
def __init__(self,
models=list(),
total_norm_weights=None,
score_name='fc-action',
dev_id=0):
"""
Contruct an action classifier
Args:
models: list of tuples in the form of
(model_proto, model_params, model_fusion_weight, input_type, conv_support, input_size).
input_type is: 0-RGB, 1-Optical flow.
conv_support indicates whether the network supports convolution testing, which is faster. If this is
not supported, we will use oversampling instead
total_norm_weights: sum of all model_fusion_weights when normalization is wanted, otherwise use None
"""
self.__net_vec = [
CaffeNet(x[0],
x[1],
dev_id,
input_size=(340, 256) if x[4] else None) for x in models
]
self.__net_weights = [float(x[2]) for x in models]
if total_norm_weights is not None:
s = sum(self.__net_weights)
self.__net_weights = [x / s for x in self.__net_weights]
self.__input_type = [x[3] for x in models]
self.__conv_support = [x[4] for x in models]
self.__num_net = len(models)
# the input size of the network
self.__input_size = [x[5] for x in models]
# whether we should prepare flow stack
self.__need_flow = max(self.__input_type) > 0
# the name in the proto for action classes
self.__score_name = score_name
# the video downloader
#self.__video_dl = youtube_dl.YoutubeDL(
# {
# 'outtmpl': '%(id)s.%(ext)s'
# }
#)
if self.__need_flow:
self.__flow_extractor = FlowExtractor(dev_id)