def test(epoch):
global best_accuracy
global best_miou
model.eval()
test_loss_iou = 0
test_loss_vec = 0
hist = np.zeros((config["task1_classes"], config["task1_classes"]))
hist_angles = np.zeros((config["task2_classes"], config["task2_classes"]))
crop_size = config["val_dataset"][args.dataset]["crop_size"]
for i, (inputsBGR, labels, vecmap_angles) in enumerate(val_loader, 0):
inputsBGR = Variable(inputsBGR.float().cuda(),
volatile=True,
requires_grad=False)
outputs, pred_vecmaps = model(inputsBGR)
if args.multi_scale_pred:
loss1 = road_loss(outputs[0],
util.to_variable(labels[0], True, False), True)
num_stacks = model.module.num_stacks if num_gpus > 1 else model.num_stacks
for idx in range(num_stacks - 1):
loss1 += road_loss(outputs[idx + 1],
util.to_variable(labels[0], True, False),
True)
for idx, output in enumerate(outputs[-2:]):
loss1 += road_loss(
output, util.to_variable(labels[idx + 1], True, False),
True)
loss2 = angle_loss(pred_vecmaps[0],
util.to_variable(vecmap_angles[0], True, False))
for idx in range(num_stacks - 1):
loss2 += angle_loss(
pred_vecmaps[idx + 1],
util.to_variable(vecmap_angles[0], True, False))
for idx, pred_vecmap in enumerate(pred_vecmaps[-2:]):
loss2 += angle_loss(
pred_vecmap,
util.to_variable(vecmap_angles[idx + 1], True, False))
outputs = outputs[-1]
pred_vecmaps = pred_vecmaps[-1]
else:
loss1 = road_loss(outputs,
util.to_variable(labels[0], True, False), True)
loss2 = angle_loss(pred_vecmaps,
util.to_variable(labels[0], True, False))
test_loss_iou += loss1.data[0]
test_loss_vec += loss2.data[0]
_, predicted = torch.max(outputs.data, 1)
correctLabel = labels[-1].view(-1, crop_size, crop_size).long()
hist += util.fast_hist(
predicted.view(predicted.size(0), -1).cpu().numpy(),
correctLabel.view(correctLabel.size(0), -1).cpu().numpy(),
config["task1_classes"],
)
_, predicted_angle = torch.max(pred_vecmaps.data, 1)
correct_angles = vecmap_angles[-1].view(-1, crop_size,
crop_size).long()
hist_angles += util.fast_hist(
predicted_angle.view(predicted_angle.size(0), -1).cpu().numpy(),
correct_angles.view(correct_angles.size(0), -1).cpu().numpy(),
config["task2_classes"],
)
p_accu, miou, road_iou, fwacc = util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
test_loss_iou / (i + 1),
test_loss_vec / (i + 1),
is_train=False,
)
p_accu_angle, miou_angle, fwacc_angle = util.performAngleMetrics(
train_loss_angle_file,
val_loss_angle_file,
epoch,
hist_angles,
is_train=False)
viz_util.progress_bar(
i,
len(val_loader),
"Loss: %.6f | VecLoss: %.6f | road miou: %.4f%%(%.4f%%) | angle miou: %.4f%%"
% (
test_loss_iou / (i + 1),
test_loss_vec / (i + 1),
miou,
road_iou,
miou_angle,
),
)
if i % 100 == 0 or i == len(val_loader) - 1:
images_path = "{}/images/".format(experiment_dir)
util.ensure_dir(images_path)
util.savePredictedProb(
inputsBGR.data.cpu(),
labels[-1].cpu(),
predicted.cpu(),
F.softmax(outputs, dim=1).data.cpu()[:, 1, :, :],
predicted_angle.cpu(),
os.path.join(images_path,
"validate_pair_{}_{}.png".format(epoch, i)),
norm_type=config["val_dataset"]["normalize_type"],
)
del inputsBGR, labels, predicted, outputs, pred_vecmaps, predicted_angle
accuracy, miou, road_iou, fwacc = util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
test_loss_iou / len(val_loader),
test_loss_vec / len(val_loader),
is_train=False,
write=True,
)
util.performAngleMetrics(
train_loss_angle_file,
val_loss_angle_file,
epoch,
hist_angles,
is_train=False,
write=True,
)
if miou > best_miou:
best_accuracy = accuracy
best_miou = miou
util.save_checkpoint(epoch, test_loss_iou / len(val_loader), model,
optimizer, best_accuracy, best_miou, config,
experiment_dir)
return test_loss_iou / len(val_loader)
def test(epoch):
global best_accuracy
global best_miou
model.eval()
test_loss_iou = 0
test_loss_vec = 0
hist = np.zeros((config["task1_classes"], config["task1_classes"]))
crop_size = config["val_dataset"][args.dataset]["crop_size"]
for i, datas in enumerate(val_loader, 0):
inputs, labels, erased_label = data
batch_size = inputs.size(0)
inputs = Variable(inputs.float().cuda(),
volatile=True,
requires_grad=False)
erased_label = Variable(erased_label[-1].float().cuda(),
volatile=True,
requires_grad=False).unsqueeze(dim=1)
temp = erased_label
for k in range(config['refinement']):
in_ = torch.cat((inputs, erased_label, temp), dim=1)
outputs = model(in_)
if args.multi_scale_pred:
loss1 = road_loss(outputs[0], labels[0].long().cuda(), False)
num_stacks = model.module.num_stacks if num_gpus > 1 else model.num_stacks
for idx in range(num_stacks - 1):
loss1 += road_loss(outputs[idx + 1],
labels[0].long().cuda(), False)
for idx, output in enumerate(outputs[-2:]):
loss1 += road_loss(output, labels[idx + 1].long().cuda(),
False)
outputs = outputs[-1]
else:
loss1 = road_loss(outputs, labels[-1].long().cuda(), False)
temp = Variable(torch.max(outputs.data, 1)[1].float(),
volatile=True,
requires_grad=False).unsqueeze(dim=1)
test_loss_iou += loss1.data[0]
_, predicted = torch.max(outputs.data, 1)
correctLabel = labels[-1].view(-1, crop_size, crop_size).long()
hist += util.fast_hist(
predicted.view(predicted.size(0), -1).cpu().numpy(),
correctLabel.view(correctLabel.size(0), -1).cpu().numpy(),
config["task1_classes"],
)
p_accu, miou, road_iou, fwacc = util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
test_loss_iou / (i + 1),
0,
is_train=False,
)
viz_util.progress_bar(
i,
len(val_loader),
"Loss: %.6f | road miou: %.4f%%(%.4f%%)" % (
test_loss_iou / (i + 1),
miou,
road_iou,
),
)
if i % 100 == 0 or i == len(val_loader) - 1:
images_path = "{}/images/".format(experiment_dir)
util.ensure_dir(images_path)
util.savePredictedProb(
inputsBGR.data.cpu(),
labels[-1].cpu(),
predicted.cpu(),
F.softmax(outputs, dim=1).data.cpu()[:, 1, :, :],
None,
os.path.join(images_path,
"validate_pair_{}_{}.png".format(epoch, i)),
norm_type=config["val_dataset"]["normalize_type"],
)
del inputsBGR, labels, predicted, outputs
accuracy, miou, road_iou, fwacc = util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
test_loss_iou / len(val_loader),
0,
is_train=False,
write=True,
)
if miou > best_miou:
best_accuracy = accuracy
best_miou = miou
util.save_checkpoint(epoch, test_loss_iou / len(val_loader), model,
optimizer, best_accuracy, best_miou, config,
experiment_dir)
return test_loss_iou / len(val_loader)
def train(epoch):
train_loss_iou = 0
train_loss_vec = 0
model.train()
optimizer.zero_grad()
hist = np.zeros((config["task1_classes"], config["task1_classes"]))
hist_angles = np.zeros((config["task2_classes"], config["task2_classes"]))
crop_size = config["train_dataset"][args.dataset]["crop_size"]
for i, data in enumerate(train_loader, 0):
inputsBGR, labels, vecmap_angles = data
inputsBGR = Variable(inputsBGR.float().cuda())
outputs, pred_vecmaps = model(inputsBGR)
if args.multi_scale_pred:
loss1 = road_loss(outputs[0], util.to_variable(labels[0]), False)
num_stacks = model.module.num_stacks if num_gpus > 1 else model.num_stacks
for idx in range(num_stacks - 1):
loss1 += road_loss(outputs[idx + 1],
util.to_variable(labels[0]), False)
for idx, output in enumerate(outputs[-2:]):
loss1 += road_loss(output, util.to_variable(labels[idx + 1]),
False)
loss2 = angle_loss(pred_vecmaps[0],
util.to_variable(vecmap_angles[0]))
for idx in range(num_stacks - 1):
loss2 += angle_loss(pred_vecmaps[idx + 1],
util.to_variable(vecmap_angles[0]))
for idx, pred_vecmap in enumerate(pred_vecmaps[-2:]):
loss2 += angle_loss(pred_vecmap,
util.to_variable(vecmap_angles[idx + 1]))
outputs = outputs[-1]
pred_vecmaps = pred_vecmaps[-1]
else:
loss1 = road_loss(outputs, util.to_variable(labels[-1]), False)
loss2 = angle_loss(pred_vecmaps,
util.to_variable(vecmap_angles[-1]))
train_loss_iou += loss1.data[0]
train_loss_vec += loss2.data[0]
_, predicted = torch.max(outputs.data, 1)
correctLabel = labels[-1].view(-1, crop_size, crop_size).long()
hist += util.fast_hist(
predicted.view(predicted.size(0), -1).cpu().numpy(),
correctLabel.view(correctLabel.size(0), -1).cpu().numpy(),
config["task1_classes"],
)
_, predicted_angle = torch.max(pred_vecmaps.data, 1)
correct_angles = vecmap_angles[-1].view(-1, crop_size,
crop_size).long()
hist_angles += util.fast_hist(
predicted_angle.view(predicted_angle.size(0), -1).cpu().numpy(),
correct_angles.view(correct_angles.size(0), -1).cpu().numpy(),
config["task2_classes"],
)
p_accu, miou, road_iou, fwacc = util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
train_loss_iou / (i + 1),
train_loss_vec / (i + 1),
)
p_accu_angle, miou_angle, fwacc_angle = util.performAngleMetrics(
train_loss_angle_file, val_loss_angle_file, epoch, hist_angles)
viz_util.progress_bar(
i,
len(train_loader),
"Loss: %.6f | VecLoss: %.6f | road miou: %.4f%%(%.4f%%) | angle miou: %.4f%% "
% (
train_loss_iou / (i + 1),
train_loss_vec / (i + 1),
miou,
road_iou,
miou_angle,
),
)
torch.autograd.backward([loss1, loss2])
if i % config["trainer"]["iter_size"] == 0 or i == len(
train_loader) - 1:
optimizer.step()
optimizer.zero_grad()
del (
outputs,
pred_vecmaps,
predicted,
correct_angles,
correctLabel,
inputsBGR,
labels,
vecmap_angles,
)
util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
train_loss_iou / len(train_loader),
train_loss_vec / len(train_loader),
write=True,
)
util.performAngleMetrics(train_loss_angle_file,
val_loss_angle_file,
epoch,
hist_angles,
write=True)
def train(epoch):
train_loss_iou = 0
train_loss_vec = 0
model.train()
optimizer.zero_grad()
hist = np.zeros((config["task1_classes"], config["task1_classes"]))
crop_size = config["train_dataset"][args.dataset]["crop_size"]
for i, data in enumerate(train_loader, 0):
inputs, labels, erased_label = data
batch_size = inputs.size(0)
inputs = Variable(inputs.float().cuda())
erased_label = Variable(
erased_label[-1].float().cuda()).unsqueeze(dim=1)
temp = erased_label
for k in range(config['refinement']):
in_ = torch.cat((inputs, erased_label, temp), dim=1)
outputs = model(in_)
if args.multi_scale_pred:
loss1 = road_loss(outputs[0], labels[0].long().cuda(), False)
num_stacks = model.module.num_stacks if num_gpus > 1 else model.num_stacks
for idx in range(num_stacks - 1):
loss1 += road_loss(outputs[idx + 1],
labels[0].long().cuda(), False)
for idx, output in enumerate(outputs[-2:]):
loss1 += road_loss(output, labels[idx + 1].long().cuda(),
False)
outputs = outputs[-1]
else:
loss1 = road_loss(outputs, labels[-1].long().cuda(), False)
loss1.backward()
temp = Variable(torch.max(outputs.data,
1)[1].float()).unsqueeze(dim=1)
train_loss_iou += loss1.data[0]
_, predicted = torch.max(outputs.data, 1)
correctLabel = labels[-1].view(-1, crop_size, crop_size).long()
hist += util.fast_hist(
predicted.view(predicted.size(0), -1).cpu().numpy(),
correctLabel.view(correctLabel.size(0), -1).cpu().numpy(),
config["task1_classes"],
)
p_accu, miou, road_iou, fwacc = util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
train_loss_iou / (i + 1),
0,
)
viz_util.progress_bar(
i,
len(train_loader),
"Loss: %.6f | road miou: %.4f%%(%.4f%%)" % (
train_loss_iou / (i + 1),
miou,
road_iou,
),
)
if i % config["trainer"]["iter_size"] == 0 or i == len(
train_loader) - 1:
optimizer.step()
optimizer.zero_grad()
del (
outputs,
predicted,
correctLabel,
inputs,
labels,
)
util.performMetrics(
train_loss_file,
val_loss_file,
epoch,
hist,
train_loss_iou / len(train_loader),
0,
write=True,
)