def test_vis_graph():
from tensorboardX import SummaryWriter
# 可视化网络结构
dataset = UCSDPed1(path=Config.video_folder) # 所以 Config也要写对
model = LSAUCSD(input_shape=dataset.shape,
code_length=64,
cpd_channels=100).to(device).eval()
model_input = torch.rand([1380, 1, 8, 32, 32]).to(device)
with SummaryWriter(log_dir="./summary/graphs/LSAUCSD") as writer:
writer.add_graph(model, input_to_model=model_input)
def test_get_c_init():
import os, time
from datasets import UCSDPed2_TRAIN
dataset = UCSDPed2_TRAIN(path="data/UCSD_Anomaly_Dataset.v1p2") # 需重写
net = LSAUCSD(input_shape=dataset.shape, code_length=64,
cpd_channels=100).to(device).eval()
# 偷个懒，暂时直接使用作者预训练模型，后面再改成 AE预训练
checkpoint = "checkpoints/ucsd_ped2.pkl"
if os.path.exists(checkpoint):
print("{} load !".format(checkpoint))
net.load_w(checkpoint)
# 只保留 Encoder部分
# net = "" # 不需要，因为直接输出 z (latent vector) 就 OK 了
#
batch_size = 2
c = init_center_c(dataset, net, batch_size)
# Use torch.save(tensor, 'file.pt') and torch.load('file.pt')
torch.save(c, "c_init_ped2.pt")
cl = torch.load("c_init_ped2.pt")
print("c1.shape: ", cl.shape)
def test_ucsdped2():
# type: () -> None
"""
Performs video anomaly detection tests on UCSD Ped2.
"""
# Build dataset and model
dataset = UCSDPed2(path=Config.video_folder)
model = LSAUCSD(input_shape=dataset.shape,
code_length=64,
cpd_channels=100).to(device).eval()
# Set up result helper and perform test
helper = VideoAnomalyDetectionResultHelper(dataset,
model,
checkpoint=Config.model_ckpt,
output_file=Config.output_file)
helper.test_video_anomaly_detection()
def test_ucsdped2():
# type: () -> None
"""
Performs video anomaly detection tests on UCSD Ped2.
"""
# Build dataset and model
dataset = UCSDPed2(path='data/UCSD_Anomaly_Dataset.v1p2')
model = LSAUCSD(input_shape=dataset.shape,
code_length=64,
cpd_channels=100).cuda().eval()
# Set up result helper and perform test
helper = VideoAnomalyDetectionResultHelper(
dataset,
model,
checkpoint='checkpoints/ucsd_ped2.pkl',
output_file='ucsd_ped2.txt')
helper.test_video_anomaly_detection()