def AD_sigle_perdiction(model_dir,c3d_features,lengths=16,device=None,network=None):
if device==None:
device = torch.device("cuda" if torch.cuda.is_available()
else "cpu")
if network==None:
print("staring the sigle AD networl")
# pediction of AD with pertrain network
network = AnomalyDetector()
network.to(device)
net = static_model(net=network,
criterion=RegularizedLoss(network, custom_objective).cuda(),
model_prefix=model_dir,
)
model_path = net.get_checkpoint_path(20000)
net.load_checkpoint(pretrain_path=model_path, epoch=20000)
net.net.to(device)
else:
net=network
#no need for anatation or batch loading of the C3D featuers
#from annotation_methods import annotatate_file
#annotation_path=annotatate_file(video_parth,dir_list,normal=True,file_name="Demo_anmotation")
# #runing vedio in to network
# data_loader = FeaturesLoaderVal(features_path=features_dir,
# annotation_path=annotation_path)
#
# data_iter = torch.utils.data.DataLoader(data_loader,
# batch_size=1,
# shuffle=False,
# num_workers=1, # 4, # change this part accordingly
# pin_memory=True)
#print("loading of data done")
#for features, start_end_couples, feature_subpaths, lengths in tqdm(data_iter):
# features is a batch where each item is a tensor of 32 4096D features
c3d_features=torch.from_numpy(c3d_features)
#print(c3d_features.shape)
features = c3d_features.to(device)
with torch.no_grad():
input_var = torch.autograd.Variable(features)
outputs = net.predict(input_var)[0] # (batch_size, 32)
outputs = outputs.reshape(1, 32)#outputs.shape[0]
for vid_len, output in zip([lengths], outputs.cpu().numpy()):
y_true = np.zeros(vid_len)
segments_len = vid_len // 32
# for couple in couples:
# if couple[0] != -1:
# y_true[couple[0]: couple[1]] = 1
y_pred = np.zeros(vid_len)
for i in range(32):
segment_start_frame = i * segments_len
segment_end_frame = (i + 1) * segments_len
y_pred[segment_start_frame: segment_end_frame] = output[i]
#print(y_true)
#print(y_pred)
#print("it is over")
return y_pred
def load_anomaly_detector(ad_model_path, device):
assert path.exists(ad_model_path)
logging.info(f"Loading anomaly detector from {ad_model_path}")
anomaly_detector = pw.System(model=AnomalyDetector(), device=device)
anomaly_detector.load_model_state(ad_model_path)
return anomaly_detector.model.eval()
def load_models(feature_extractor_path, ad_model_path, features_method='c3d', device='cuda'):
"""
Loads both feature extractor and anomaly detector from the given paths
:param feature_extractor_path: path of the features extractor weights to load
:param ad_model_path: path of the anomaly detector weights to load
:param features_method: name of the model to use for features extraction
:param device: device to use for the models
:return: anomaly_detector, feature_extractor
"""
assert path.exists(feature_extractor_path)
assert path.exists(ad_model_path)
feature_extractor, anomaly_detector = None, None
if features_method == 'c3d':
logging.info(f"Loading feature extractor from {feature_extractor_path}")
feature_extractor = C3D(pretrained=feature_extractor_path)
else:
raise NotImplementedError(f"Features extraction method {features_method} not implemented")
logging.info(f"Loading anomaly detector from {ad_model_path}")
feature_extractor = feature_extractor.to(device).eval()
anomaly_detector = pw.System(model=AnomalyDetector(), device=device)
anomaly_detector.load_model_state(ad_model_path)
anomaly_detector = anomaly_detector.model.eval()
return anomaly_detector, feature_extractor
def network_setup(ad_model_dir='/home/peter/Documents/actionID/AnomalyDetectionCVPR2018-Pytorch-master/short_60_low_mem/exps/model'):
device = torch.device("cuda" if torch.cuda.is_available()
else "cpu")
c3d_network = C3D(pretrained=pretrained_3d)
c3d_network.to(device)
print("staring the sigle AD networl")
# pediction of AD with pertrain network
AD_network = AnomalyDetector()
AD_network.to(device)
net = static_model(net=AD_network,
criterion=RegularizedLoss(AD_network, custom_objective).cuda(),
model_prefix=ad_model_dir,
)
model_path = net.get_checkpoint_path(20000)
net.load_checkpoint(pretrain_path=model_path, epoch=20000)
net.net.to(device)
return device,c3d_network,net
"""
train_iter = torch.utils.data.DataLoader(train_loader,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
"""
eval_iter = torch.utils.data.DataLoader(val_loader,
batch_size=args.batch_size,
shuffle=True,
num_workers=1,
pin_memory=True)
"""
iter_seed = torch.initial_seed() + 100
network = AnomalyDetector()
net = model(
net=network,
criterion=RegularizedLoss(network, custom_objective).to(device),
model_prefix=args.model_dir,
step_callback_freq=5,
save_checkpoint_freq=args.save_frequency,
opt_batch_size=args.batch_size, # optional, 60 in the paper
)
if torch.cuda.is_available():
net.net.cuda()
torch.cuda.manual_seed(args.random_seed)
net.net = torch.nn.DataParallel(net.net).cuda()
"""
In the original paper:
help="path to test annotation")
parser.add_argument('--random-seed', type=int, default=1,
help='random seed (default: 1)')
parser.add_argument('--model-dir', type=str, default="./exps/model",
help="set model dir.")
parser.add_argument('--classid_path', type=str, default="ClassIDs.txt",
help="path to ClassIDs")
if __name__ == "__main__":
args = parser.parse_args()
# load network
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
network = AnomalyDetector()
network.to(device)
net = static_model(net=network,
criterion=RegularizedLoss(network, custom_objective).to(device),
model_prefix=args.model_dir)
model_path = net.get_checkpoint_path(20000)
net.load_checkpoint(pretrain_path=model_path, epoch=20000)
net.net.to(device)
# enable cudnn tune
cudnn.benchmark = True
# load test data
data_loader = FeaturesLoader(features_path=args.features_path,
annotation_path=args.annotation_path_test,
classid_path=args.classid_path)
device = get_torch_device()
cudnn.benchmark = True # enable cudnn tune
# Data loader
train_loader = FeaturesDatasetWrapper(features_path=args.features_path,
annotation_path=args.annotation_path)
train_iter = torch.utils.data.DataLoader(train_loader,
batch_size=args.batch_size,
num_workers=0,
pin_memory=True)
# Model
if args.checkpoint is not None and path.exists(args.checkpoint):
model = TorchModel.load_model(args.checkpoint)
else:
network = AnomalyDetector(args.feature_dim)
model = TorchModel(network)
# Training parameters
"""
In the original paper:
lr = 0.01
epsilon = 1e-8
"""
optimizer = torch.optim.Adadelta(network.parameters(),
lr=args.lr_base,
eps=1e-8)
criterion = RegularizedLoss(network, custom_objective)
# Callbacks
def AD_perdiction(model_dir, dir_list, device=None):
if device == None:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#pediction of AD with pertrain network
network = AnomalyDetector()
network.to(device)
net = static_model(
net=network,
criterion=RegularizedLoss(network, custom_objective).cuda(),
model_prefix=model_dir,
)
model_path = net.get_checkpoint_path(20000)
net.load_checkpoint(pretrain_path=model_path, epoch=20000)
net.net.to(device)
from annotation_methods import annotatate_file
annotation_path = annotatate_file(video_path,
dir_list,
normal=[True, True],
file_name="Demo_anmotation")
#runing vedio in to network
data_loader = FeaturesLoaderVal(features_path=features_dir,
annotation_path=annotation_path)
data_iter = torch.utils.data.DataLoader(
data_loader,
batch_size=1,
shuffle=False,
num_workers=1, # 4, # change this part accordingly
pin_memory=True)
print("it is over")
for features, start_end_couples, feature_subpaths, lengths in tqdm(
data_iter):
# features is a batch where each item is a tensor of 32 4096D features
features = features.to(device)
with torch.no_grad():
input_var = torch.autograd.Variable(features)
outputs = net.predict(input_var)[0] # (batch_size, 32)
outputs = outputs.reshape(outputs.shape[0], 32)
for vid_len, couples, output in zip(lengths, start_end_couples,
outputs.cpu().numpy()):
y_true = np.zeros(vid_len)
segments_len = vid_len // 32
for couple in couples:
if couple[0] != -1:
y_true[couple[0]:couple[1]] = 1
y_pred = np.zeros(vid_len)
print()
for i in range(32):
segment_start_frame = i * segments_len
segment_end_frame = (i + 1) * segments_len
y_pred[segment_start_frame:segment_end_frame] = output[i]
print(y_true)
print(y_pred)
print("it is over")
return y_pred
if __name__ == "__main__":
args = get_args()
register_logger(log_file=args.log_file)
os.makedirs(args.exps_dir, exist_ok=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cudnn.benchmark = True # enable cudnn tune
train_loader = FeaturesDatasetWrapper(features_path=args.features_path, annotation_path=args.annotation_path)
train_iter = torch.utils.data.DataLoader(train_loader,
batch_size=args.batch_size,
num_workers=0, # 4, # change this part accordingly
pin_memory=True)
network = AnomalyDetector()
system = pw.System(model=network, device=device)
"""
In the original paper:
lr = 0.01
epsilon = 1e-8
"""
optimizer = torch.optim.Adadelta(network.parameters(), lr=args.lr_base, eps=1e-8)
loss_wrapper = pw.loss_wrappers.GenericPointWiseLossWrapper(RegularizedLoss(network, custom_objective))
system.train(
loss_wrapper,
optimizer,
train_data_loader=train_iter,
# Optimizations
device = get_torch_device()
cudnn.benchmark = True # enable cudnn tune
# Data loader
train_loader = FeaturesLoader(features_path=args.features_path, annotation_path=args.annotation_path)
train_iter = torch.utils.data.DataLoader(train_loader,
batch_size=args.batch_size,
num_workers=8,
pin_memory=True)
# Model
if args.checkpoint is not None and path.exists(args.checkpoint):
model = TorchModel.load_model(args.checkpoint)
else:
network = AnomalyDetector(args.feature_dim)
model = TorchModel(network)
model = model.to(device).train()
# Training parameters
"""
In the original paper:
lr = 0.01
epsilon = 1e-8
"""
optimizer = torch.optim.Adadelta(model.parameters(), lr=args.lr_base, eps=1e-8)
criterion = RegularizedLoss(network, custom_objective).to(device)
# Callbacks
tb_writer = SummaryWriter(log_dir=tb_dir)
