def train(self, epoch):
self.model.train()
losses = AverageMeter()
top1 = AverageMeter()
pbar = tqdm(self.train_dataloader)
for batch_idx, data in enumerate(pbar):
self.iters += 1
images, labels = data['images'], data['labels']
if self.use_cuda:
images, labels = images.cuda(), labels.cuda()
outputs = self.model(images)
loss = self.criterion(outputs, labels)
self.optimizer.zero_grad()
loss.backward(loss.data)
self.optimizer.step()
if self.args.optimizer_type == 'sgd':
adjust_learning_rate(self.optimizer, self.args.base_lr, epoch,
self.args.lr_decay_epoch)
prec1 = accuracy(outputs.data, labels.data, topk=(1, ))
loss = loss.view(-1)
losses.update(loss.data[0], images.size(0))
top1.update(prec1[0].detach().item(), images.size(0))
self.writer.add_scalar('train/loss', loss.data[0], self.iters)
self.writer.add_scalar('train/acc', top1.val, self.iters)
if batch_idx % self.args.log_interval == 0:
print('epoch:{}, iter:{}/{}, loss:{}, acc:{}'.format(
epoch, batch_idx, self.iters, losses.avg,
round(top1.avg, 6)))
losses.reset(), top1.reset()
score_train = trainer.get_scores()
precision = score_train['precision']
mAP = score_train['mAP']
batch_time.update(time.time()-start_time)
start_time = time.time()
tbar.set_description('Train loss: %.3f; precision: %.3f; mAP: %.3f; data time: %.3f; batch time: %.3f'%
(train_loss / (i_batch + 1), precision, mAP, data_time.avg, batch_time.avg))
writer.add_scalar('loss', train_loss/len(tbar), epoch)
writer.add_scalar('mAP/train', mAP, epoch)
writer.add_scalar('precision/train', precision, epoch)
writer.add_scalar('distance/train', score_train['distance'], epoch)
trainer.reset_metrics()
data_time.reset()
batch_time.reset()
if epoch % 1 == 0:
with torch.no_grad():
model.eval()
print("evaluating...")
if test:
tbar = tqdm(dataloader_test)
else:
tbar = tqdm(dataloader_val)
start_time = time.time()
for i_batch, sample_batched in enumerate(tbar):
data_time.update(time.time()-start_time)