def run():
statistical_angular_errors = StatisticalValue()
sub_dir = time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime())
print('Test start.')
with torch.no_grad():
for idx, (images, labels, names) in enumerate(testloader):
images, labels = images.to(DEVICE), labels.to(DEVICE)
predictions = model(images)
angular_error = multi_angular_loss(predictions[-1], labels)
statistical_angular_errors.update(angular_error.item(),
names,
sort=True)
view_data = torch.zeros((4, *images.shape[1:]))
view_data[0, :, :, :] = images.squeeze()
view_data[
1, :, :, :] = images.squeeze() / predictions[-1].squeeze()
view_data[2, :, :, :] = predictions[-1].squeeze()
view_data[3, :, :, :] = labels.squeeze()
if not os.path.isdir(os.path.join(TMP_ROOT, 'test', sub_dir)):
os.makedirs(os.path.join(TMP_ROOT, 'test', sub_dir))
torchvision.utils.save_image(
view_data,
os.path.join(TMP_ROOT, 'test/%s/%s' % (sub_dir, names[0])))
print(
'Angular Error: mean: {errors.avg}, mid: {errors.mid}, worst: {errors.max[0]}, best: {errors.min[0]}'
.format(errors=statistical_angular_errors))
print('Test end.')
def run(epoch):
statistical_losses = StatisticalValue()
statistical_angular_errors = StatisticalValue()
optimizer.zero_grad()
for idx, (images, labels, names) in enumerate(trainloader):
images, labels = images.to(DEVICE), labels.to(DEVICE)
predictions = model(images)
loss = torch.zeros(1).to(DEVICE)
for p in predictions:
loss += criterion(p, labels) / ITERATION_SIZE
loss.backward()
angular_error = multi_angular_loss(predictions[-1], labels)
statistical_angular_errors.update(angular_error.item(), names)
statistical_losses.update(loss.item())
iteration_writer.add_scalar(
'Loss/Iteration',
statistical_losses.val[0],
(epoch - 1) * len(trainloader) + idx + 1
)
if (idx + 1) % ITERATION_SIZE == 0:
optimizer.step()
optimizer.zero_grad()
print_training_status(epoch, idx + 1, len(trainloader),
statistical_losses.val[0], statistical_losses.avg)
print(
'Angular Error: mean: {errors.avg}, worst: {errors.max[0]}, best: {errors.min[0]}'.format(
errors=statistical_angular_errors))
scheduler.step()
return statistical_losses