#print('utils.py, class AverageMeter()')
if not os.path.exists('models'):
os.mkdir('models')
epoch = 0
for epoch in range(int(iterations)):
# while i < int(iterations):
for sample in tqdm(data_loader):
images, labels = sample['images'].to(device), sample['labels'].to(device)
logits = model(images)
labels = labels.view(int(bs)*int(seq_length)) ##??
loss = criterion(logits, labels)
optimizer.zero_grad()
loss.backward()
losses.update(loss.item(), images.size(0))
optimizer.step()
print('epoch: {}\tLoss: {loss.val:.4f} ({loss.avg:.4f})'.format(epoch, loss=losses))
epoch += 1
if epoch % it_save == 0:
torch.save({'optimizer_state_dict': optimizer.state_dict(),
'model_state_dict': model.state_dict()}, 'models/swingnet_{}.pth.tar'.format(epoch))
if epoch == iterations:
break
experiment.log_metrics("train_loss", losses, step=epoch)
losses = AverageMeter()
if not os.path.exists("models"):
os.mkdir("models")
i = 0
while i < iterations:
for sample in data_loader:
images, labels = sample["images"].cuda(), sample["labels"].cuda()
logits = model(images)
labels = labels.view(bs * seq_length)
loss = criterion(logits, labels)
optimizer.zero_grad()
loss.backward()
losses.update(loss.item(), images.size(0))
optimizer.step()
print(
"Iteration: {}\tLoss: {loss.val:.4f} ({loss.avg:.4f})".format(
i, loss=losses))
i += 1
if i % it_save == 0:
torch.save(
{
"optimizer_state_dict": optimizer.state_dict(),
"model_state_dict": model.state_dict(),
},
"models/swingnet_{}.pth.tar".format(i),
)
if i == iterations:
break
if not os.path.exists('models'):
os.mkdir('models')
i = 0
while i < iterations:
for sample in data_loader:
images, labels = sample['images'].cuda(), sample['labels'].cuda()
logits = model(images)
labels = labels.view(bs * seq_length)
loss = criterion(logits, labels)
optimizer.zero_grad()
loss.backward()
losses.update(loss.item(), images.size(0))
optimizer.step()
print(
'Iteration: {}\tLoss: {loss.val:.4f} ({loss.avg:.4f})'.format(
i, loss=losses))
i += 1
if i % it_save == 0:
torch.save(
{
'optimizer_state_dict': optimizer.state_dict(),
'model_state_dict': model.state_dict()
}, 'models/{}_{}.pth.tar'.format(version_name, i))
if i == iterations:
break
plt.ylim([np.log(min(np.min(list_loss['loss']), np.min(list_loss['soft']), np.min(list_loss['count']))), np.log(list_loss['loss'][min(500,max(0,i-100))])])
plt.subplot(1, 2, 2)
plt.hist(preds.reshape(-1).cpu().data.numpy())
plt.ylim([0,50])
plt.savefig('loss_' + version_name + '.png')
plt.close('all')
optimizer.zero_grad()
loss.backward()
losses.update(loss.item(), images.size(0))
optimizer.step()
print('Iteration: {}\tLoss: {loss.val:.4f} ({loss.avg:.4f})'.format(i, loss=losses))
i += 1
if i % it_save == 0:
torch.save({'optimizer_state_dict': optimizer.state_dict(),
'model_state_dict': model.state_dict()}, 'models/{}_{}.pth.tar'.format(version_name, i))
if i == iterations:
break
# Run evaluation
import os
os.system("python eval.py " + version_name)