def train_gcn(param):
torch.cuda.empty_cache()
videos_pkl_train = "/home/lnn/workspace/pygcn/pygcn/ucf_crime_train.pkl"
videos_pkl_test = "/home/lnn/workspace/pygcn/pygcn/ucf_crime_test.pkl"
feature_folder = "/home/lnn/workspace/UCF_Crimes/kinetics_flow5000_feat/"
prediction_folder = "/home/lnn/workspace/UCF_Crimes/kinetics_flow5000/"
test_pred_gt_folder = "/home/lnn/data/UCF_Crimes/test_pred_groundtruth/"
modality = "flow"
gpu_id = 1
iter_size = 32
ucf_crime_train = UCFCrimeSlow(videos_pkl_train,
prediction_folder,
feature_folder,
modality,
graph_generator=soft_uniform_sampling,
graph_generator_param=param)
train_loader = DataLoader(dataset=ucf_crime_train,
batch_size=1,
shuffle=True,
num_workers=16)
model = NoiseFilter(nfeat=1024, nclass=1)
criterion_supervised = SigmoidCrossEntropyLoss()
criterion_unsupervised = SigmoidMAELoss()
if gpu_id != -1:
model = model.cuda(gpu_id)
criterion_supervised = criterion_supervised.cuda(gpu_id)
criterion_unsupervised = criterion_unsupervised.cuda(gpu_id)
optimizer = optim.SGD(model.parameters(),
lr=0.0001,
momentum=0.9,
weight_decay=0.0005)
opt_scheduler = optim.lr_scheduler.StepLR(optimizer, 16, 0.1)
iter_count = 0
avg_loss_train = 0
alpha = 0.5
vid2mean_pred = {}
#model.load_state_dict(torch.load("flow_9_0.6.pth"))
for epoch in range(20):
model.train()
opt_scheduler.step()
for step, data in enumerate(train_loader):
(feat, adj, labeled_index_in_the_graph,
labeled_index), pred, vid = data
feat, adj, pred = Variable(feat), Variable(adj), Variable(pred)
if not vid2mean_pred.has_key(vid[0]):
vid2mean_pred[
vid[0]] = pred.data.cpu().numpy().flatten().copy()
mean_pred = Variable(torch.from_numpy(vid2mean_pred[vid[0]]),
requires_grad=False)
if gpu_id != -1:
feat = feat.cuda(gpu_id)
adj = adj.cuda(gpu_id)
pred = pred.cuda(gpu_id)
mean_pred = mean_pred.cuda(gpu_id)
if iter_count % iter_size == 0:
optimizer.zero_grad()
output = model(feat, adj)
labeled_index_in_the_graph = np.array(
labeled_index_in_the_graph).flatten()
labeled_index = np.array(labeled_index).flatten()
sample_index = get_sample_index(labeled_index, pred)
if "Normal" in vid[0]:
loss_train = criterion_supervised(
output.view(1, -1),
pred.view(1, -1)[:, range(output.shape[1])])
else:
'''
loss_train = criterion_supervised(output.view(1, -1)[:, labeled_index_in_the_graph],
pred.view(1, -1)[:, labeled_index])
'''
loss_train = criterion_supervised(output.view(1, -1)[:, labeled_index_in_the_graph],
pred.view(1, -1)[:, labeled_index])+ \
criterion_unsupervised(output.view(1, -1),
mean_pred.view(1, -1)[:, sample_index])
avg_loss_train += loss_train
iter_count += 1
loss_train.backward()
#torch.nn.utils.clip_grad_norm(model.parameters(), 40)
mean_pred_current = mean_pred.data.cpu().numpy().copy().flatten()
mean_pred_current[sample_index] = sigmoid(
output).data.cpu().numpy().copy().flatten()
vid2mean_pred[vid[0]] = alpha * vid2mean_pred[vid[0]] + (
1 - alpha) * mean_pred_current
if (iter_count + 1) % iter_size == 0:
print("Train loss: %.4f" % (avg_loss_train / iter_size))
avg_loss_train = 0
optimizer.step()
torch.save(model.state_dict(), "flow_%d.pth" % epoch)
# iter_count += 1610
# model.load_state_dict(torch.load("%d.pth" % iter_count))
'''
x = range(len(gt))
plt.scatter(x, gt, color='g')
plt.scatter(x, ans, color='r')
plt.show()
'''
print("Epoch %d done !" % epoch)
gcn_model_path = "/home/lnn/workspace/pygcn/pygcn/models_c3d_2_0.6/c3d_4.pth"
modality = "c3d"
gpu_id = 0
output_folder = "/home/lnn/workspace/pygcn/output_c3d_high_conf_2/"
if __name__ == '__main__':
param = (4, 0.7, 0.1, 2, 2000, -0.1)
ucf_crime_train = UCFCrimeSlow(videos_pkl_train,
prediction_folder,
feature_folder,
modality,
graph_generator=soft_uniform_sampling,
graph_generator_param=param,
random_crop=False)
train_loader = DataLoader(dataset=ucf_crime_train, num_workers=16)
model = NoiseFilter(nfeat=4096, nclass=1)
if gpu_id != -1:
model = model.cuda(gpu_id)
model.load_state_dict(torch.load(gcn_model_path))
model.eval()
vid2ans = {}
for step, data in enumerate(train_loader):
(feat, adj, labeled_index_in_the_graph,
labeled_index), pred, vid = data
feat, adj, pred = Variable(feat), Variable(adj), Variable(pred)
if gpu_id != -1:
feat = feat.cuda(gpu_id)
adj = adj.cuda(gpu_id)
import numpy as np
from sklearn.metrics import roc_auc_score
from utils import graph_generator
from utils import collate_video
from models import NoiseFilter
from dataset import UCFCrime
gpu_id = 1
iter_size = 16
frames_per_feat = 1
if __name__ == '__main__':
feature_path = "/home/zjx/data/UCF_Crimes/C3D_features/c3d_fc6_features.hdf5"
ucf_crime = UCFCrime(feature_path, graph_generator)
model = NoiseFilter(nfeat=4096, nclass=2)
criterion = CrossEntropyLoss()
if gpu_id != -1:
model = model.cuda(gpu_id)
criterion = criterion.cuda(gpu_id)
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9, weight_decay=0.0005)
iter_count = 0
avg_loss_train = 0
for epoch in range(8):
model.train()
for step, data in enumerate(train_loader):
(feat, adj), is_normal, vid = data
def train_gcn(param):
torch.cuda.empty_cache()
# videos_pkl_train = "/home/lnn/workspace/pygcn/pygcn/ucf_crime_train.pkl"
videos_pkl_train = "/mmu_ssd/liuchang03/heyuwei/GCN-Anomaly-Detection/pygcn/my_crime_train.pkl"
# videos_pkl_test = "/home/lnn/workspace/pygcn/pygcn/ucf_crime_test.pkl"
# feature_folder = "/home/lnn/workspace/UCF_Crimes/c3d_features_all/_iter_4500/"
feature_folder = "/mmu_ssd/liuchang03/heyuwei/Data/crime_c3d_feature"
prediction_folder = "/mmu_ssd/liuchang03/heyuwei/Data/crime_c3d_score"
# prediction_folder = "/home/lnn/workspace/UCF_Crimes/c3d_features_1_high_conf_sampling_0.6/_iter_1000/"
test_pred_gt_folder = "/home/lnn/data/UCF_Crimes/test_pred_groundtruth/"
modality = "c3d"
gpu_id = 0
iter_size = 32
ucf_crime_train = UCFCrimeSlow(videos_pkl_train, prediction_folder, feature_folder, modality,
graph_generator=soft_uniform_sampling, graph_generator_param=param)
train_loader = DataLoader(dataset=ucf_crime_train, batch_size=1, shuffle=False)
model = NoiseFilter(nfeat=4096, nclass=1)
criterion_supervised = SigmoidCrossEntropyLoss()
criterion_unsupervised = SigmoidMAELoss()
if gpu_id != -1:
model = model.cuda(gpu_id)
criterion_supervised = criterion_supervised.cuda(gpu_id)
criterion_unsupervised = criterion_unsupervised.cuda(gpu_id)
optimizer = optim.SGD(model.parameters(), lr=0.0001, momentum=0.9, weight_decay=0.0005)
opt_scheduler = optim.lr_scheduler.StepLR(optimizer, 5, 0.1)
iter_count = 0
avg_loss_train = 0
alpha = 0.5
vid2mean_pred = {}
for epoch in range(20):
model.train()
opt_scheduler.step()
for step, data in enumerate(train_loader):
(feat, adj, labeled_index_in_the_graph, labeled_index), pred, vid = data
# print('feat shape ' + str(feat.shape))
# print('adj shape ' + str(adj.shape))
# print('labeled_index_in_the_graph: ' + str(labeled_index_in_the_graph))
# print('labeled_index: ' + str(labeled_index))
feat, adj, pred = Variable(feat), Variable(adj), Variable(pred)
if vid[0] not in list(vid2mean_pred.keys()):
# if not vid2mean_pred.has_key(vid[0]):
vid2mean_pred[vid[0]] = pred.data.cpu().numpy().flatten().copy()
mean_pred = Variable(torch.from_numpy(vid2mean_pred[vid[0]]), requires_grad=False)
if gpu_id != -1:
feat = feat.cuda(gpu_id)
adj = adj.cuda(gpu_id)
pred = pred.cuda(gpu_id)
mean_pred = mean_pred.cuda(gpu_id)
if iter_count % iter_size == 0:
optimizer.zero_grad()
output = model(feat, adj)
labeled_index_in_the_graph = np.array(labeled_index_in_the_graph).flatten()
labeled_index = np.array(labeled_index).flatten()
sample_index = get_sample_index(labeled_index, pred)
if "Normal" in vid[0]:
loss_train = criterion_supervised(output.view(1, -1),
pred.view(1, -1)[:, range(output.shape[1])])
else:
# loss_train = criterion_supervised(output.view(1, -1)[:, labeled_index_in_the_graph],
# pred.view(1, -1)[:, labeled_index]) + \
# criterion_unsupervised(output.view(1, -1),
# mean_pred.view(1, -1)[:, sample_index])
loss_train = criterion_supervised(output.view(1, -1)[:, labeled_index_in_the_graph].float(),
pred.view(1, -1)[:, labeled_index]).float() + \
criterion_unsupervised(output.view(1, -1).float(),
mean_pred.view(1, -1)[:, sample_index].float())
avg_loss_train += loss_train
iter_count += 1
loss_train.backward()
mean_pred_current = mean_pred.data.cpu().numpy().copy().flatten()
mean_pred_current[sample_index] = sigmoid(output).data.cpu().numpy().copy().flatten()
vid2mean_pred[vid[0]] = alpha * vid2mean_pred[vid[0]] + (1 - alpha) * mean_pred_current
if (iter_count + 1) % iter_size == 0:
print("Train loss: %.4f" % (avg_loss_train / iter_size))
avg_loss_train = 0
optimizer.step()
torch.save(model.state_dict(), "c3d_%d.pth" % epoch)
# iter_count += 1610
# model.load_state_dict(torch.load("%d.pth" % iter_count))
print("Epoch %d done !" % epoch)