def test(args, net_model, model_path=None):
if model_path is not None:
net_model = torch.load(model_path)
print(">>> [Testing] Load pretrained model from " + model_path)
net_model.eval()
AVEData = AVE_Fully_Dataset(video_dir=args.dir_video,
audio_dir=args.dir_audio,
label_dir=args.dir_labels,
order_dir=args.dir_order_test,
batch_size=402,
status='test')
nb_batch = AVEData.__len__()
audio_inputs, video_inputs, labels, _, _, = AVEData.get_batch(0)
audio_inputs = audio_inputs.to(device)
video_inputs = video_inputs.to(device)
labels = labels.to(device)
fusion, out_prob, cross_att = net_model(audio_inputs, video_inputs,
args.threshold)
labels = labels.cpu().data.numpy()
x_labels = out_prob.cpu().data.numpy()
acc = compute_acc(labels, x_labels, nb_batch)
print('[test]acc: ', acc)
return acc
def test(args, net_model, model_path=None):
if model_path is not None:
net_model = torch.load(model_path)
print(">>> [Testing] Load pretrained model from " + model_path)
net_model.eval()
AVEData = AVE_Weakly_Dataset(video_dir=args.dir_video,
video_dir_bg=args.dir_video_bg,
audio_dir=args.dir_audio,
audio_dir_bg=args.dir_audio_bg,
label_dir=args.dir_labels,
prob_label_dir=args.prob_dir_labels,
label_dir_bg=args.dir_labels_bg,
label_dir_gt=args.dir_labels_gt,
order_dir=args.dir_order_test,
batch_size=402,
status="test")
nb_batch = AVEData.__len__()
audio_inputs, video_inputs, labels = AVEData.get_batch(0)
audio_inputs = audio_inputs.cuda()
video_inputs = video_inputs.cuda()
labels = labels.numpy()
scores_avg, x_labels = net_model(audio_inputs, video_inputs,
args.threshold) # shape:
x_labels = F.softmax(x_labels, dim=-1)
x_labels = x_labels.cpu().data.numpy()
acc = compute_acc(labels, x_labels, nb_batch)
return acc
def val(args, net_model):
net_model.eval()
AVEData = AVE_Weakly_Dataset(video_dir=args.dir_video,
video_dir_bg=args.dir_video_bg,
audio_dir=args.dir_audio,
audio_dir_bg=args.dir_audio_bg,
label_dir=args.dir_labels,
prob_label_dir=args.prob_dir_labels,
label_dir_bg=args.dir_labels_bg,
label_dir_gt=args.dir_labels_gt,
order_dir=args.dir_order_val,
batch_size=402,
status="val")
nb_batch = AVEData.__len__()
audio_inputs, video_inputs, labels = AVEData.get_batch(0)
# labels: [bs, 10, 29]
audio_inputs = audio_inputs.cuda()
video_inputs = video_inputs.cuda()
labels = labels.numpy()
scores_avg, x_labels = net_model(audio_inputs, video_inputs,
args.threshold) # shape:
# x_labels: [bs, 10, 29]
x_labels = F.softmax(x_labels, dim=-1)
x_labels = x_labels.cpu().data.numpy()
acc = compute_acc(labels, x_labels, nb_batch)
print('val accuracy', acc)
return acc
def val(args, net_model):
net_model.eval()
AVEData = AVE_Fully_Dataset(video_dir=args.dir_video,
audio_dir=args.dir_audio,
label_dir=args.dir_labels,
order_dir=args.dir_order_val,
batch_size=402,
status='val')
nb_batch = AVEData.__len__()
audio_inputs, video_inputs, labels, _, _ = AVEData.get_batch(0)
audio_inputs = audio_inputs.to(device)
video_inputs = video_inputs.to(device)
labels = labels.to(device)
fusion, out_prob, cross_att = net_model(audio_inputs, video_inputs,
args.threshold)
labels = labels.cpu().data.numpy()
x_labels = out_prob.cpu().data.numpy()
acc = compute_acc(labels, x_labels, nb_batch)
print('[val]acc: ', acc)
return acc
def train(args, net_model, optimizer):
AVEData = AVE_Fully_Dataset(video_dir=args.dir_video,
audio_dir=args.dir_audio,
label_dir=args.dir_labels,
order_dir=args.dir_order_train,
batch_size=args.batch_size,
status='train')
nb_batch = AVEData.__len__() // args.batch_size
print('nb_batch:', nb_batch)
epoch_l = []
best_val_acc = 0
best_test_acc = 0
best_epoch = 0
for epoch in range(args.nb_epoch):
net_model.train()
epoch_loss = 0
epoch_loss_cls = 0
epoch_loss_avps = 0
n = 0
start = time.time()
SHUFFLE_SAMPLES = True
for i in range(nb_batch):
audio_inputs, video_inputs, labels, segment_label_batch, segment_avps_gt_batch = AVEData.get_batch(
i, SHUFFLE_SAMPLES)
SHUFFLE_SAMPLES = False
audio_inputs = audio_inputs.to(device)
video_inputs = video_inputs.to(device)
labels = labels.to(device)
segment_label_batch = segment_label_batch.to(device)
segment_avps_gt_batch = segment_avps_gt_batch.to(device)
net_model.zero_grad()
fusion, out_prob, cross_att = net_model(audio_inputs, video_inputs,
args.threshold) # shape:
# out_prob: [bs, 10, 29], score_max: [bs, 29]
loss_cls = nn.CrossEntropyLoss()(out_prob.permute(
0, 2, 1), segment_label_batch) # segment_label_batch: [bs, 10]
loss_avps = AVPSLoss(cross_att, segment_avps_gt_batch)
loss = loss_cls + args.LAMBDA * loss_avps
epoch_loss += loss.cpu().data.numpy()
epoch_loss_cls += loss_cls.cpu().data.numpy()
epoch_loss_avps += loss_avps.cpu().data.numpy()
loss.backward()
optimizer.step_lr()
n = n + 1
SHUFFLE_SAMPLES = True
if (epoch + 1) % 60 == 0 and epoch < 170:
optimizer.update_lr()
end = time.time()
epoch_l.append(epoch_loss)
labels = labels.cpu().data.numpy()
x_labels = out_prob.cpu().data.numpy()
acc = compute_acc(labels, x_labels, nb_batch)
print("=== Epoch {%s} lr: {%.6f} | Loss: [{%.4f}] loss_cls: [{%.4f}] | loss_frame: [{%.4f}] | training_acc {%.4f}" \
% (str(epoch), optimizer._optimizer.param_groups[0]['lr'], (epoch_loss) / n, epoch_loss_cls/n, epoch_loss_avps/n, acc))
if epoch % args.save_epoch == 0 and epoch != 0:
val_acc = val(args, net_model)
print('val accuracy:', val_acc, 'epoch=', epoch)
if val_acc >= best_val_acc:
best_val_acc = val_acc
print('best val accuracy:', best_val_acc)
print(
'best val accuracy: {} ***************************************'
.format(best_val_acc))
# torch.save(net_model, model_name + "_" + str(epoch) + "_fully.pt")
if epoch % args.check_epoch == 0 and epoch != 0:
test_acc = test(args, net_model)
print('test accuracy:', test_acc, 'epoch=', epoch)
if test_acc >= best_test_acc:
best_test_acc = test_acc
best_epoch = epoch
print(
'best test accuracy: {} ======================================='
.format(best_test_acc))
torch.save(net_model,
model_name + "_" + str(epoch) + "_fully.pt")
print('[best val accuracy]: ', best_val_acc)
print('[best test accuracy]: ', best_test_acc)
