def test(logger, epoch):
model.eval()
test_loss = 0
meters = logger.reset_meters('val')
begin = time.time()
for batch_idx, data in enumerate(test_loader):
batch_size = data['answer'].size(0)
if args.cuda:
question, image, target = data['question'].cuda(
async=True), data['image'].float().cuda(
async=True), data['answer'].cuda(async=True)
else:
question, image, target = data['question'], data['image'].float(
), data['answer']
question, image, target = Variable(question, volatile=True), Variable(
image, volatile=True), Variable(target, volatile=True)
# Compute output and loss
output = model(question, image)
loss = F.nll_loss(output, target).data[0]
test_loss += loss.item() # sum up batch loss
meters['loss'].update(loss.item(), n=batch_size)
acc1, acc5 = utils.accuracy(output.data, target.data, topk=(1, 5))
meters['acc1'].update(acc1[0], n=batch_size)
meters['acc5'].update(acc5[0], n=batch_size)
tensorboard_writer.add_scalar("Test loss", loss.item(),
epoch * len(test_loader) + batch_idx)
tensorboard_writer.add_scalar("Test top1 acc", acc1[0],
epoch * len(test_loader) + batch_idx)
tensorboard_writer.add_scalar("Test top 5 acc ", acc5[0],
epoch * len(test_loader) + batch_idx)
meters['batch_time'].update(time.time() - begin, n=batch_size)
test_loss /= len(test_loader.dataset)
print(
'\n Test set: Average loss: {:.4f}, Acc@1 {acc1.avg:.3f} Acc@5 {acc5.avg:.3f}'
.format(test_loss, acc1=meters['acc1'], acc5=meters['acc5']))
logger.log_meters('val', n=epoch)
return meters['acc1'].avg
def train(epoch, logger):
begin = time.time()
model.train()
meters = logger.reset_meters('train')
start = time.time()
for batch_idx, data in enumerate(train_loader):
batch_size = data['question'].size(0)
# Measures the data loading time
meters['data_time'].update(time.time() - begin, n=batch_size)
if args.cuda:
question, image, target = data['question'].cuda(
), data['image'].float().cuda(), data['answer'].cuda(async=True)
else:
question, image, target = data['question'], data['image'].float(
), data['answer']
question, image, target = Variable(question), Variable(
image), Variable(target)
# Compute output and loss
output = model(question, image)
if args.cuda:
torch.cuda.synchronize()
loss = F.nll_loss(output, target)
# Log the loss
meters['loss'].update(loss.data[0], n=batch_size)
# Measure accuracy
acc1, acc5 = utils.accuracy(output.data, target.data, topk=(1, 5))
meters['acc1'].update(acc1[0], n=batch_size)
meters['acc5'].update(acc5[0], n=batch_size)
#print(loss)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm(model.parameters(), 5)
if args.cuda:
torch.cuda.synchronize()
# optimizer.step()
#if args.cuda:
# torch.cuda.synchronize()
meters['batch_time'].update(time.time() - begin, n=batch_size)
begin = time.time()
optimizer.step()
if batch_idx % args.log_interval == 0:
print("Time since last print : {}".format(time.time() - start))
start = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {acc1.val:.3f} ({acc1.avg:.3f})\t'
'Acc@5 {acc5.val:.3f} ({acc5.avg:.3f})'.format(
epoch,
batch_idx,
len(train_loader),
batch_time=meters['batch_time'],
data_time=meters['data_time'],
loss=meters['loss'],
acc1=meters['acc1'],
acc5=meters['acc5']))
sys.stdout.flush()
logger.log_meters('train', n=epoch)