def evalEpoch(epoch, dataloader, model, criterion, evaluation, opt, writer):
model.eval()
total_loss, total_accuracy = Meter(), Meter()
for batch in dataloader:
if not opt.use_rcnn:
img, ques, ques_mask, _, ans_idx = batch
else:
img, ques, img_mask, ques_mask, _, ans_idx = batch
img = Variable(img, volatile=True)
img_mask = Variable(img_mask, volatile=True) if opt.use_rcnn else None
ques = Variable(ques, volatile=True)
ques_mask = Variable(ques_mask, volatile=True)
ans_idx = Variable(ans_idx, volatile=True)
img, img_mask, ques, ques_mask, ans_idx = \
move_to_cuda((img, img_mask, ques, ques_mask, ans_idx), devices=opt.gpus)
ques = model.word_embedded(ques)
score = model(img, ques, img_mask, ques_mask) if opt.use_rcnn else \
model(img, ques, img_mask, ques_mask, is_train=False)
accuracy = evaluation(score, ans_idx)
loss = criterion(score, ans_idx)
total_loss.update((loss.data[0] / opt.batch_size))
total_accuracy.update(accuracy.data[0])
writer.add_scalar("epoch/val_loss", total_loss.avg, global_step=epoch)
writer.add_scalar("epoch/val_accuracy",
total_accuracy.avg,
global_step=epoch)
return total_loss.avg, total_accuracy.avg
def evalEpoch(epoch, dataloader, model, criterion, evaluation, opt, writer):
model.eval()
total_loss, total_accuracy = Meter(), Meter()
for i, batch in enumerate(dataloader):
img, ques, ques_mask, _, ans_idx = batch
img = torch.tensor(img, requires_grad=False)
img_mask = None
ques = torch.tensor(ques, requires_grad=False)
ques_mask = torch.tensor(ques_mask, requires_grad=False)
ans_idx = torch.tensor(ans_idx, requires_grad=False)
img, img_mask, ques, ques_mask, ans_idx = move_to_cuda(
(img, img_mask, ques, ques_mask, ans_idx), devices=opt['gpus'])
ques = model.word_embedded(ques)
score = model(img, ques, img_mask, ques_mask, is_train=False)
accuracy = evaluation(score, ans_idx)
loss = criterion(score, ans_idx)
total_loss.update((loss.item() / opt['batch_size']))
total_accuracy.update(accuracy.item())
if (i + 1) % opt['log_interval'] == 0:
print("VALIDATION: Epoch %5d; iter %6i" % (epoch, i + 1))
writer.add_scalar("epoch/val_loss", total_loss.avg, global_step=epoch)
writer.add_scalar("epoch/val_accuracy",
total_accuracy.avg,
global_step=epoch)
return total_loss.avg, total_accuracy.avg
def trainEpoch(epoch, dataloader, model, criterion, evaluation, optim, opt,
writer):
model.train()
loss_record = [Meter() for _ in range(3)]
accuracy_record = [Meter() for _ in range(3)]
timer = Timer()
timer.tic()
optim.step_epoch()
for i, batch in enumerate(dataloader):
if not opt.use_rcnn:
img, ques, ques_mask, _, ans_idx = batch
else:
img, ques, img_mask, ques_mask, _, ans_idx = batch
img = Variable(img) if opt.use_rcnn else Variable(img, volatile=True)
img_mask = Variable(img_mask) if opt.use_rcnn else None
ques = Variable(ques, volatile=True)
ques_mask = Variable(ques_mask)
ans_idx = Variable(ans_idx)
img, img_mask, ques, ques_mask, ans_idx = \
move_to_cuda((img, img_mask, ques, ques_mask, ans_idx), devices=opt.gpus)
ques = model.word_embedded(ques)
ques = Variable(ques.data)
optim.zero_grad()
score = model(img, ques, img_mask, ques_mask) if opt.use_rcnn else \
model(img, ques, img_mask, ques_mask, is_train=True)
loss = criterion(score, ans_idx)
loss.backward()
accuracy = evaluation(Variable(score.data, volatile=True),
Variable(ans_idx.data, volatile=True))
_, ratio, updates, params = optim.step()
for j in range(3):
loss_record[j].update((loss.data[0] / opt.batch_size))
accuracy_record[j].update(accuracy.data[0])
if ratio is not None:
writer.add_scalar("statistics/update_to_param_ratio",
ratio,
global_step=(epoch * len(dataloader) + i))
writer.add_scalar("statistics/absolute_updates",
updates,
global_step=(epoch * len(dataloader) + i))
writer.add_scalar("statistics/absolute_params",
params,
global_step=(epoch * len(dataloader) + i))
if (i + 1) % 10 == 0:
writer.add_scalar("iter/train_loss",
loss_record[0].avg,
global_step=(epoch * len(dataloader) + i))
writer.add_scalar("iter/train_accuracy",
accuracy_record[0].avg,
global_step=(epoch * len(dataloader) + i))
loss_record[0].reset()
accuracy_record[0].reset()
if (i + 1) % opt.log_interval == 0:
print(
"Epoch %5d; iter %6i; loss: %8.2f; accuracy: %8.2f; %6.0fs elapsed"
% (epoch, i + 1, loss_record[1].avg, accuracy_record[1].avg,
timer.toc(average=False)))
loss_record[1].reset()
accuracy_record[1].reset()
timer.tic()
writer.add_scalar("epoch/train_loss",
loss_record[2].avg,
global_step=epoch)
writer.add_scalar("epoch/train_accuracy",
accuracy_record[2].avg,
global_step=epoch)
return loss_record[2].avg, accuracy_record[2].avg