def train_1epoch(_model, _train_loader, _optimizer, _loss_func, _epoch,
_nb_epochs):
print('==> Epoch:[{0}/{1}][training stage]'.format(_epoch, _nb_epochs))
accuracy_list = []
loss_list = []
_model.train()
for i, (data, label) in enumerate(_train_loader):
label = label.cuda(async=True)
input_var = Variable(data).cuda()
target_var = Variable(label).cuda().long()
output = _model(input_var)
loss = _loss_func(output, target_var)
loss_list.append(loss)
# measure accuracy and record loss
prec1 = accuracy(output.data, label) # , topk=(1, 5))
accuracy_list.append(prec1)
# print('Top1:', prec1)
# compute gradient and do SGD step
_optimizer.zero_grad()
loss.backward()
_optimizer.step()
return float(sum(accuracy_list) / len(accuracy_list)), float(
sum(loss_list) / len(loss_list)), _model
def train_1epoch(_img_feature_extract_model, _dcnn_model, _train_loader,
_optimizer, _loss_func, _epoch, _nb_epochs):
print('==> Epoch:[{0}/{1}][training stage]'.format(_epoch, _nb_epochs))
accuracy_list = []
loss_list = []
_img_feature_extract_model.train()
_dcnn_model.train()
for i, (dat, label, video_name) in enumerate(_train_loader):
label = label.cuda()
# input_var = Variable(dat).cuda()
target_var = Variable(label).cuda().long()
print(dat.size())
for batch_idx in range(batch_size):
for frame_idx in range(max_frame_num):
output = _img_feature_extract_model(
dat[batch_idx, :, frame_idx, :].unsqueeze_(0).cuda())
print(output.size())
# output = _model(input_var)
loss = _loss_func(output, target_var)
loss_list.append(loss.data)
accuracy_list.append(accuracy(output.data, label))
# compute gradient and do SGD step
_optimizer.zero_grad()
loss.backward()
_optimizer.step()
return float(sum(accuracy_list) / len(accuracy_list)), float(
sum(loss_list) /
len(loss_list)), _img_feature_extract_model, _dcnn_model
def train_tsn_1epoch(_model, _train_loader, _optimizer, _loss_func, _epoch,
_nb_epochs):
print('==> Epoch:[{0}/{1}][training stage]'.format(_epoch, _nb_epochs))
accuracy_list = []
loss_list = []
_model.train()
for i, (data, label) in enumerate(_train_loader):
label = label.cuda()
# input_var = Variable(data).cuda()
target_var = Variable(label).cuda().long()
for i, dat in enumerate(data):
input_var = Variable(dat).cuda()
if i == 0:
output = _model(input_var)
else:
output = output + _model(input_var)
loss = _loss_func(output / 3, target_var)
# measure accuracy and record loss
prec = accuracy(output.data, label)
loss_list.append(loss)
accuracy_list.append(float(prec))
# compute gradient and do SGD step
_optimizer.zero_grad()
loss.backward()
_optimizer.step()
return float(sum(accuracy_list) / len(accuracy_list)), float(
sum(loss_list) / len(loss_list)), _model
def all_frmae_validation_epoch(_model, _val_loader, _optimizer, _loss_func,
_epoch, _nb_epochs):
print('==> Epoch:[{0}/{1}][validation stage]'.format(_epoch, _nb_epochs))
accuracy_list = []
loss_list = []
_model.eval()
with torch.no_grad():
for i, (dat, label) in enumerate(_val_loader):
label = label.cuda()
input_var = Variable(dat).cuda()
target_var = Variable(label).cuda()
# compute output
output = _model(input_var)
loss = _loss_func(output, target_var)
loss_list.append(loss.data)
accuracy_list.append(accuracy(output.data, label))
return float(sum(accuracy_list) / len(accuracy_list)), float(
sum(loss_list) / len(loss_list)), _model