def train(train_loader, model, criterion, optimizer, epoch, zoom_factor, batch_size, aux_weight, fcw):
batch_time = AverageMeter()
data_time = AverageMeter()
main_loss_meter = AverageMeter()
aux_loss_meter = AverageMeter()
loss_meter = AverageMeter()
intersection_meter = AverageMeter()
union_meter = AverageMeter()
target_meter = AverageMeter()
model.train()
fcw.train()
world_size = dist.get_world_size()
rank = dist.get_rank()
end = time.time()
for i, (input, target) in enumerate(train_loader):
# to avoid bn problem in ppm module with bin size 1x1, sometimes n may get 1 on one gpu during the last batch, so just discard
# if input.shape[0] < batch_size:
# continue
data_time.update(time.time() - end)
current_iter = (epoch - 1) * len(train_loader) + i + 1
max_iter = args.epochs * len(train_loader)
if args.net_type == 0:
index_split = 4
elif args.net_type in [1, 2, 3]:
index_split = 5
poly_learning_rate(optimizer, args.base_lr, current_iter, max_iter, power=args.power, index_split=index_split)
input = input.cuda()
input_var = torch.autograd.Variable(input)
fcw_input = fcw(input_var)
output, aux = model(input_var, fcw_input)
if zoom_factor != 8:
h = int((target.size()[1]-1) / 8 * zoom_factor + 1)
w = int((target.size()[2] - 1) / 8 * zoom_factor + 1)
# 'nearest' mode doesn't support downsampling operation and while 'bilinear' mode is also fine
target = F.upsample(target.unsqueeze(1).float(), size=(h, w), mode='bilinear').squeeze(1).long()
target = target.data.cuda(async=True)
target_var = torch.autograd.Variable(target)
main_loss = criterion(output, target_var) / world_size
aux_loss = criterion(aux, target_var) / world_size
loss = main_loss + aux_weight * aux_loss
optimizer.zero_grad()
loss.backward()
average_gradients(model)
optimizer.step()
output = output.data.max(1)[1].cpu().numpy()
target = target.cpu().numpy()
intersection, union, target = intersectionAndUnion(output, target, args.classes, args.ignore_label)
intersection_meter.update(intersection)
union_meter.update(union)
target_meter.update(target)
accuracy = sum(intersection_meter.val) / (sum(target_meter.val) + 1e-10)
reduced_loss = loss.data.clone()
reduced_main_loss = main_loss.data.clone()
reduced_aux_loss = aux_loss.data.clone()
dist.all_reduce(reduced_loss)
dist.all_reduce(reduced_main_loss)
dist.all_reduce(reduced_aux_loss)
main_loss_meter.update(reduced_main_loss[0], input.size(0))
aux_loss_meter.update(reduced_aux_loss[0], input.size(0))
loss_meter.update(reduced_loss[0], input.size(0))
batch_time.update(time.time() - end)
end = time.time()
# calculate remain time
remain_iter = max_iter - current_iter
remain_time = remain_iter * batch_time.avg
t_m, t_s = divmod(remain_time, 60)
t_h, t_m = divmod(t_m, 60)
remain_time = '{:02d}:{:02d}:{:02d}'.format(int(t_h), int(t_m), int(t_s))
if rank == 0:
if (i + 1) % args.print_freq == 0:
logger.info('Epoch: [{}/{}][{}/{}] '
'Data {data_time.val:.3f} ({data_time.avg:.3f}) '
'Batch {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Remain {remain_time} '
'MainLoss {main_loss_meter.val:.4f} '
'AuxLoss {aux_loss_meter.val:.4f} '
'Loss {loss_meter.val:.4f} '
'Accuracy {accuracy:.4f}.'.format(epoch, args.epochs, i + 1, len(train_loader),
batch_time=batch_time,
data_time=data_time,
remain_time=remain_time,
main_loss_meter=main_loss_meter,
aux_loss_meter=aux_loss_meter,
loss_meter=loss_meter,
accuracy=accuracy))
writer.add_scalar('loss_train_batch', main_loss_meter.val, current_iter)
writer.add_scalar('mIoU_train_batch', np.mean(intersection / (union + 1e-10)), current_iter)
writer.add_scalar('mAcc_train_batch', np.mean(intersection / (target + 1e-10)), current_iter)
writer.add_scalar('allAcc_train_batch', accuracy, current_iter)
iou_class = intersection_meter.sum / (union_meter.sum + 1e-10)
accuracy_class = intersection_meter.sum / (target_meter.sum + 1e-10)
mIoU = np.mean(iou_class)
mAcc = np.mean(accuracy_class)
allAcc = sum(intersection_meter.sum) / (sum(target_meter.sum) + 1e-10)
if rank == 0:
logger.info('Train result at epoch [{}/{}]: mIoU/mAcc/allAcc {:.4f}/{:.4f}/{:.4f}.'.format(epoch, args.epochs, mIoU, mAcc, allAcc))
return main_loss_meter.avg, mIoU, mAcc, allAcc
def train(train_loader, model, model_ppm, criterion, optimizer, epoch, zoom_factor, batch_size, aux_weight):
batch_time = AverageMeter()
data_time = AverageMeter()
r_loss_meter = AverageMeter()
t_loss_meter = AverageMeter()
loss_meter = AverageMeter()
model.train()
model_ppm.train()
world_size = dist.get_world_size()
rank = dist.get_rank()
#print(rank)
end = time.time()
for i, (input1, input2, translation, quaternions) in enumerate(train_loader):
# to avoid bn problem in ppm module with bin size 1x1, sometimes n may get 1 on one gpu during the last batch, so just discard
# if input.shape[0] < batch_size:
# continue
data_time.update(time.time() - end)
current_iter = (epoch - 1) * len(train_loader) + i + 1
max_iter = args.epochs * len(train_loader)
index_split = 4
poly_learning_rate(optimizer, args.base_lr, current_iter, max_iter, power=args.power, index_split=index_split)
input1 = input1.cuda()
input_var1 = torch.autograd.Variable(input1)
input2 = input2.cuda()
input_var2 = torch.autograd.Variable(input2)
x1_ICR, x1_PFR, x1_PRP = model(input_var1)
x2_ICR, x2_PFR, x2_PRP = model(input_var2)
x1_ICR = (x1_ICR + x1_PFR + x1_PRP)/3
x2_ICR = (x2_ICR + x2_PFR + x2_PRP)/3
trans, quat = model_ppm(torch.cat([x1_ICR,x2_ICR], 1))
translation = translation.float().cuda(async=True)
translation_var = torch.autograd.Variable(translation)
quaternions = quaternions.float().cuda(async=True)
quaternions_var = torch.autograd.Variable(quaternions)
t_loss = criterion(trans, translation_var) / world_size
r_loss = criterion(quat, quaternions_var) / world_size
loss = r_loss + t_loss
optimizer.zero_grad()
loss.backward()
average_gradients(model)
optimizer.step()
reduced_loss = loss.data.clone()
reduced_t_loss = t_loss.data.clone()
reduced_r_loss = r_loss.data.clone()
dist.all_reduce(reduced_loss)
dist.all_reduce(reduced_t_loss)
dist.all_reduce(reduced_r_loss)
r_loss_meter.update(reduced_r_loss[0], input1.size(0))
t_loss_meter.update(reduced_t_loss[0], input1.size(0))
loss_meter.update(reduced_loss[0], input1.size(0))
batch_time.update(time.time() - end)
end = time.time()
# calculate remain time
remain_iter = max_iter - current_iter
remain_time = remain_iter * batch_time.avg
t_m, t_s = divmod(remain_time, 60)
t_h, t_m = divmod(t_m, 60)
remain_time = '{:02d}:{:02d}:{:02d}'.format(int(t_h), int(t_m), int(t_s))
if rank == 0:
if (i + 1) % args.print_freq == 0:
logger.info('Epoch: [{}/{}][{}/{}] '
'Data {data_time.val:.3f} ({data_time.avg:.3f}) '
'Batch {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Remain {remain_time} '
'rLoss {r_loss_meter.val:.4f} '
'tLoss {t_loss_meter.val:.4f} '
'Loss {loss_meter.val:.4f} '.format(epoch, args.epochs, i + 1, len(train_loader),
batch_time=batch_time,
data_time=data_time,
remain_time=remain_time,
t_loss_meter=t_loss_meter,
r_loss_meter=r_loss_meter,
loss_meter=loss_meter))
writer.add_scalar('loss_train_batch_r', r_loss_meter.val, current_iter)
writer.add_scalar('loss_train_batch_t', t_loss_meter.val, current_iter)
if rank == 0:
logger.info('Train result at epoch [{}/{}].'.format(epoch, args.epochs))
return t_loss_meter.avg, r_loss_meter.avg
def main():
dataset = Train_DataSet("data_road", "image", KITTI_IMAGE_SIZE)
dataset_loader = DataLoader(dataset,
batch_size=2,
shuffle=True,
drop_last=True,
num_workers=0)
dataset_valid = Valid_DataSet("data_road", "image", KITTI_IMAGE_SIZE)
dataset_loader_valid = DataLoader(dataset_valid,
batch_size=1,
num_workers=0)
dataset_test = Test_DataSet("data_road", "image", KITTI_IMAGE_SIZE)
dataset_loader_test = DataLoader(dataset_test, batch_size=1, num_workers=0)
model = PSPNet().to(device)
optimizer = torch.optim.SGD(model.parameters(),
lr=0.01,
momentum=0.9,
weight_decay=0.0001)
writer = SummaryWriter()
epochs = 300
epoch = 0
while True:
running_loss = 0.0
for i, data in enumerate(dataset_loader, 0):
max_iter = epochs * len(dataset_loader)
current_iter = epoch * len(dataset_loader) + i + 1
current_lr = poly_learning_rate(0.01, current_iter, max_iter)
# for index in range(0, 5):
optimizer.param_groups[0]['lr'] = current_lr
images, labels = data
images = images.to(device)
labels = labels.to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
output, main_loss, aux_loss = model(images, labels)
loss = main_loss + 0.4 * aux_loss
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
if i % 10 == 9:
writer.add_scalar("Training Loss", running_loss / 10,
epoch * len(dataset_loader) + i)
print("[%d, %5d] loss: %.3f" %
(epoch + 1, i + 1, running_loss / 10))
running_loss = 0.0
if i % 30 == 0:
grid = torchvision.utils.make_grid(output)
writer.add_image("Output Image", grid, 0)
# Check Accuracy
acc = acc_check(model,
dataset_valid,
dataset_loader_valid,
epoch,
save=1)
writer.add_scalar("Training Accuracy", acc, epoch)
epoch += 1
if acc > 96 or epoch > epochs:
print("Finished Training..!")
break
writer.close()
model.eval()
correct = 0
total = 0
with torch.no_grad():
for i, data in enumerate(dataset_loader_test):
images, labels = data
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
name = dataset_test.image_dir[i].split("\\")[-1]
output_softmax = outputs.detach().cpu().numpy()[0].transpose(
1, 2, 0)
gt = labels.detach().cpu().numpy()[0].transpose(1, 2, 0)
output_img = np.squeeze(output_softmax)[:, :, 1].reshape(
KITTI_IMAGE_SIZE[1], KITTI_IMAGE_SIZE[0])
segmentation_r = (output_img > 0.8).reshape(
KITTI_IMAGE_SIZE[1], KITTI_IMAGE_SIZE[0], 1)
mask = np.dot(segmentation_r, np.array([[0, 255, 0]]))
cv2.imwrite("data_road/output/{0}".format(name), mask)
total += labels.size()[2] * labels.size()[3]
correct += np.all((output_softmax > 0.8) == gt, axis=2).sum()
correct += (predicted == labels).sum().item()
print("Accuracy of the network on the %d test images: %d %%" %
(dataset_test.__len__(), 100 * correct / total))
summary(model, input_size=(3, 160, 576))
def train(train_loader, model, criterion, optimizer, epoch, zoom_factor,
batch_size, aux_weight):
batch_time = AverageMeter()
data_time = AverageMeter()
main_loss_meter = AverageMeter()
aux_loss_meter = AverageMeter()
loss_meter = AverageMeter()
intersection_meter = AverageMeter()
union_meter = AverageMeter()
target_meter = AverageMeter()
model.train()
world_size = dist.get_world_size()
rank = dist.get_rank()
end = time.time()
for i, (_, input, atts) in enumerate(train_loader):
# to avoid bn problem in ppm module with bin size 1x1, sometimes n may get 1 on one gpu during the last batch, so just discard
# if input.shape[0] < batch_size:
# continue
data_time.update(time.time() - end)
current_iter = (epoch - 1) * len(train_loader) + i + 1
max_iter = args.epochs * len(train_loader)
index_split = 4
poly_learning_rate(optimizer,
args.base_lr,
current_iter,
max_iter,
power=args.power,
index_split=index_split)
input = input.cuda()
input_var = torch.autograd.Variable(input)
output = model(input_var).squeeze(3).squeeze(2) #, aux
# if zoom_factor != 8:
# h = int((target.size()[1] - 1) / 8 * zoom_factor + 1)
# w = int((target.size()[2] - 1) / 8 * zoom_factor + 1)
# target = F.upsample(target.unsqueeze(1).float(), size=(h, w), mode='bilinear').squeeze(1).long()
# target = target.data.cuda(async=True)
# target_var = torch.autograd.Variable(target)
atts_var = torch.autograd.Variable(atts.cuda(async=True))
main_loss = F.multilabel_soft_margin_loss(output,
atts_var) / world_size
# aux_loss = criterion(aux, target_var) / world_size
loss = main_loss # + aux_weight * aux_loss
optimizer.zero_grad()
loss.backward()
average_gradients(model)
optimizer.step()
reduced_loss = loss.data.clone()
reduced_main_loss = main_loss.data.clone()
dist.all_reduce(reduced_loss)
dist.all_reduce(reduced_main_loss)
main_loss_meter.update(reduced_main_loss[0], input.size(0))
loss_meter.update(reduced_loss[0], input.size(0))
batch_time.update(time.time() - end)
end = time.time()
# calculate remain time
remain_iter = max_iter - current_iter
remain_time = remain_iter * batch_time.avg
t_m, t_s = divmod(remain_time, 60)
t_h, t_m = divmod(t_m, 60)
remain_time = '{:02d}:{:02d}:{:02d}'.format(int(t_h), int(t_m),
int(t_s))
if rank == 0:
if (i + 1) % args.print_freq == 0:
logger.info(
'Epoch: [{}/{}][{}/{}] '
'Data {data_time.val:.3f} ({data_time.avg:.3f}) '
'Batch {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Remain {remain_time} '
'MainLoss {main_loss_meter.val:.4f} '
'Loss {loss_meter.val:.4f} '.format(
epoch,
args.epochs,
i + 1,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
remain_time=remain_time,
main_loss_meter=main_loss_meter,
loss_meter=loss_meter))
writer.add_scalar('loss_train_batch', main_loss_meter.val,
current_iter)
if rank == 0:
logger.info('Train result at epoch [{}/{}]: '.format(
epoch, args.epochs))
return main_loss_meter.avg
def train(train_loader, model, optimizer, epoch, batch_size):
batch_time = AverageMeter()
data_time = AverageMeter()
main_loss_meter = AverageMeter()
loss_meter = AverageMeter()
model.train()
world_size = dist.get_world_size()
rank = dist.get_rank()
end = time.time()
for i, (_, input, target) in enumerate(train_loader):
# to avoid bn problem in ppm module with bin size 1x1, sometimes n may get 1 on one gpu during the last batch, so just discard
# if input.shape[0] < batch_size:
# continue
data_time.update(time.time() - end)
current_iter = (epoch - 1) * len(train_loader) + i + 1
max_iter = args.epochs * len(train_loader)
poly_learning_rate(optimizer,
args.base_lr,
current_iter,
max_iter,
power=args.power,
index_split=4)
input = input.cuda()
input_var = torch.autograd.Variable(input)
output1, output3, output6, output9 = model(input_var)
output1 = output1.squeeze(3).squeeze(2)
output3 = output3.squeeze(3).squeeze(2)
output6 = output6.squeeze(3).squeeze(2)
output9 = output9.squeeze(3).squeeze(2)
target = target.cuda(async=True)
target_var = torch.autograd.Variable(target)
main_loss = (
F.multilabel_soft_margin_loss(output1, target_var) +
F.multilabel_soft_margin_loss(output3, target_var) +
F.multilabel_soft_margin_loss(output6, target_var) +
F.multilabel_soft_margin_loss(output9, target_var)) / world_size
loss = main_loss
optimizer.zero_grad()
loss.backward()
average_gradients(model)
optimizer.step()
reduced_loss = loss.data.clone()
reduced_main_loss = main_loss.data.clone()
dist.all_reduce(reduced_loss)
dist.all_reduce(reduced_main_loss)
main_loss_meter.update(reduced_main_loss[0], input.size(0))
loss_meter.update(reduced_loss[0], input.size(0))
batch_time.update(time.time() - end)
end = time.time()
# calculate remain time
remain_iter = max_iter - current_iter
remain_time = remain_iter * batch_time.avg
t_m, t_s = divmod(remain_time, 60)
t_h, t_m = divmod(t_m, 60)
remain_time = '{:02d}:{:02d}:{:02d}'.format(int(t_h), int(t_m),
int(t_s))
if rank == 0:
if (i + 1) % args.print_freq == 0:
logger.info(
'Epoch: [{}/{}][{}/{}] '
'Data {data_time.val:.3f} ({data_time.avg:.3f}) '
'Batch {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Remain {remain_time} '
'MainLoss {main_loss_meter.val:.4f} '
'Loss {loss_meter.val:.4f} '.format(
epoch,
args.epochs,
i + 1,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
remain_time=remain_time,
main_loss_meter=main_loss_meter,
loss_meter=loss_meter))
if rank == 0:
logger.info('Train result at epoch [{}/{}]'.format(epoch, args.epochs))
return main_loss_meter.avg
def train(train_loader, train_transform, model, model_icr, model_pfr,
model_prp, TripletLoss, criterion, optimizer, epoch, zoom_factor,
batch_size, aux_weight):
batch_time = AverageMeter()
data_time = AverageMeter()
r_loss_meter = AverageMeter()
t_loss_meter = AverageMeter()
proj_loss_meter = AverageMeter()
triple_loss_meter = AverageMeter()
rfine_loss_meter = AverageMeter()
tfine_loss_meter = AverageMeter()
loss_meter = AverageMeter()
model.train()
model_icr.train()
model_pfr.train()
model_prp.train()
world_size = dist.get_world_size()
rank = dist.get_rank()
#print(rank)
end = time.time()
for index, (image_anchor, image1, image2, image3, relative_t1, relative_r1,
relative_t2, relative_r2, relative_t3, relative_r3, image1_r,
image2_r, image3_r, anchor_name, absolute_r1, absolute_t1,
absolute_r2, absolute_t2, absolute_r3, absolute_t3,
absolute_ranchor, absolute_tanchor) in enumerate(train_loader):
# to avoid bn problem in ppm module with bin size 1x1, sometimes n may get 1 on one gpu during the last batch, so just discard
# if input.shape[0] < batch_size:
# continue
data_time.update(time.time() - end)
current_iter = (epoch - 1) * len(train_loader) + index + 1
max_iter = args.epochs * len(train_loader)
index_split = 4
poly_learning_rate(optimizer,
args.base_lr,
current_iter,
max_iter,
power=args.power,
index_split=index_split)
# print(image_anchor.size())
image_anchor = torch.cat([image_anchor, image_anchor, image_anchor], 0)
image1 = torch.cat([image1, image2, image3], 0)
relative_t1 = torch.cat([relative_t1, relative_t2, relative_t3], 0)
relative_r1 = torch.cat([relative_r1, relative_r2, relative_r3], 0)
image1_r = torch.cat([image1_r, image2_r, image3_r], 0)
# print(image_anchor.size())
image_anchor = image_anchor.cuda()
image_anchor_var = torch.autograd.Variable(image_anchor)
image1 = image1.cuda()
image1_var = torch.autograd.Variable(image1)
x1_ICR, x1_PFR, x1_PRP = model(image_anchor_var)
x2_ICR, x2_PFR, x2_PRP = model(image1_var)
proj_ICR = model_icr(torch.cat([x1_ICR, x2_ICR], 1))
trans, quat = model_pfr(torch.cat([x1_PFR, x2_PFR], 1))
translation = relative_t1.float().cuda(async=True)
translation_var = torch.autograd.Variable(translation)
quaternions = relative_r1.float().cuda(async=True)
quaternions_var = torch.autograd.Variable(quaternions)
proj = image1_r.float().cuda(async=True)
proj_var = torch.autograd.Variable(proj)
triple_loss = TripletLoss(
x1_ICR.squeeze(3).squeeze(2),
x2_ICR.squeeze(3).squeeze(2)) / world_size
t_loss = criterion(trans.squeeze(3).squeeze(2),
translation_var) / world_size
r_loss = criterion(quat.squeeze(3).squeeze(2),
quaternions_var) / world_size
proj_loss = criterion(proj_ICR.squeeze(3).squeeze(2),
proj_var) / world_size
#########################################################################################
absolute_r1 = torch.cat([absolute_r1, absolute_r2, absolute_r3], 0)
absolute_t1 = torch.cat([absolute_t1, absolute_t2, absolute_t3], 0)
course_t = absolute_t1 - trans.squeeze(3).squeeze(2).data.cpu()
course_r = get_coarse_quaternion(
absolute_r1,
quat.squeeze(3).squeeze(2).data.cpu() *
np.array([1, -1, -1, -1], np.float32))
course_rt = torch.cat([course_r, course_t], 1).numpy()
#print(anchor_name, absolute_r1.size(), quat.size(), course_t, course_r)
name_list = []
rt_list = []
for item in anchor_name:
name_info, rt_info = get_retrival_info(
item.replace(
args.data_root,
'/mnt/lustre/dingmingyu/Research/ICCV19/CamNet/scripts/retrival_lists/'
))
name_list.append(name_info)
rt_list.append(rt_info)
#print(len(name_list), name_list)
fine_list = []
fine_rt_list = []
r_fine_loss = 0
t_fine_loss = 0
for i in range(course_rt.shape[0]):
distances = pose_distance(course_rt[i:i + 1],
rt_list[i % int(course_rt.shape[0] / 3)])
num = np.argmin(distances)
#print(num, distances[num])
fine_list.append(name_list[i % int(course_rt.shape[0] / 3)][num])
fine_rt_list.append(rt_list[i % int(course_rt.shape[0] / 3)][num])
for i in range(len(anchor_name)):
#print(anchor_name[i], args.data_root + fine_list[i])
fine_anchor, fine_1, fine_2, fine_3 = cv2.imread(
anchor_name[i].replace('pose.txt', 'color.png')), cv2.imread(
args.data_root +
fine_list[i].replace('pose.txt', 'color.png')), cv2.imread(
args.data_root + fine_list[i + len(anchor_name)].
replace('pose.txt', 'color.png')), cv2.imread(
args.data_root +
fine_list[i + len(anchor_name) + len(anchor_name)].
replace('pose.txt', 'color.png'))
fine_anchor, fine_1, fine_2, fine_3 = train_transform(
fine_anchor, fine_1, fine_2, fine_3)
fine_anchor = torch.cat([
fine_anchor.unsqueeze(0),
fine_anchor.unsqueeze(0),
fine_anchor.unsqueeze(0)
], 0)
fine_1 = torch.cat([
fine_1.unsqueeze(0),
fine_2.unsqueeze(0),
fine_3.unsqueeze(0)
], 0)
fine_anchor_r = absolute_ranchor[i:i + 1]
fine_anchor_r = torch.cat(
[fine_anchor_r, fine_anchor_r, fine_anchor_r], 0)
fine_anchor_t = absolute_tanchor[i:i + 1]
fine_anchor_t = torch.cat(
[fine_anchor_t, fine_anchor_t, fine_anchor_t], 0)
fine_imgs_rt = np.array([
fine_rt_list[i], fine_rt_list[i + len(anchor_name)],
fine_rt_list[i + len(anchor_name) + len(anchor_name)]
]).astype(np.float32)
fine_imgs_r = torch.from_numpy(fine_imgs_rt[:, :4])
fine_imgs_t = torch.from_numpy(fine_imgs_rt[:, 4:])
#print(fine_anchor_r.size(), fine_anchor_t.size(), fine_imgs_r.size(), fine_imgs_t.size())
fine_rela_t = fine_imgs_t - fine_anchor_t
fine_rela_t_var = torch.autograd.Variable(fine_rela_t.cuda())
fine_anchor_r[:, 1:] *= -1
fine_rela_r = get_coarse_quaternion(fine_anchor_r, fine_imgs_r)
fine_rela_r_var = torch.autograd.Variable(fine_rela_r.cuda())
#print(fine_rela_r.size(), fine_rela_t.size(), fine_rela_r, fine_rela_t)
fine_anchor_var = torch.autograd.Variable(fine_anchor.cuda())
fine_imgs_var = torch.autograd.Variable(fine_1.cuda())
_, _, anchor_PRP = model(fine_anchor_var)
_, _, imgs_PRP = model(fine_imgs_var)
trans_PRP, quat_PRP = model_prp(
torch.cat([anchor_PRP, imgs_PRP], 1))
r_fine_loss += criterion(
quat_PRP.squeeze(3).squeeze(2),
fine_rela_r_var) / world_size / len(anchor_name)
t_fine_loss += criterion(
trans_PRP.squeeze(3).squeeze(2),
fine_rela_t_var) / world_size / len(anchor_name)
if rank == 0:
print(anchor_name[i], args.data_root + fine_list[i],
fine_list[i + len(anchor_name)],
fine_list[i + len(anchor_name) + len(anchor_name)],
fine_anchor_r, fine_anchor_t, fine_rela_t, fine_rela_r)
#print(anchor_name[i], args.data_root + fine_list[i], fine_anchor_r[i], fine_anchor_t[i], fine_rt_list[i], fine_rt_list[i+len(anchor_name)], fine_rt_list[i+len(anchor_name)+len(anchor_name)])
loss = r_loss + t_loss + proj_loss + triple_loss + r_fine_loss + t_fine_loss
optimizer.zero_grad()
loss.backward()
average_gradients(model)
optimizer.step()
reduced_loss = loss.data.clone()
reduced_t_loss = t_loss.data.clone()
reduced_r_loss = r_loss.data.clone()
reduced_proj_loss = proj_loss.data.clone()
reduced_triple_loss = triple_loss.data.clone()
reduced_rfine_loss = r_fine_loss.data.clone()
reduced_tfine_loss = t_fine_loss.data.clone()
dist.all_reduce(reduced_loss)
dist.all_reduce(reduced_t_loss)
dist.all_reduce(reduced_r_loss)
dist.all_reduce(reduced_proj_loss)
dist.all_reduce(reduced_triple_loss)
dist.all_reduce(reduced_rfine_loss)
dist.all_reduce(reduced_tfine_loss)
r_loss_meter.update(reduced_r_loss[0], image_anchor.size(0))
t_loss_meter.update(reduced_t_loss[0], image_anchor.size(0))
proj_loss_meter.update(reduced_proj_loss[0], image_anchor.size(0))
triple_loss_meter.update(reduced_triple_loss[0], image_anchor.size(0))
rfine_loss_meter.update(reduced_rfine_loss[0], image_anchor.size(0))
tfine_loss_meter.update(reduced_tfine_loss[0], image_anchor.size(0))
loss_meter.update(reduced_loss[0], image_anchor.size(0))
batch_time.update(time.time() - end)
end = time.time()
# calculate remain time
remain_iter = max_iter - current_iter
remain_time = remain_iter * batch_time.avg
t_m, t_s = divmod(remain_time, 60)
t_h, t_m = divmod(t_m, 60)
remain_time = '{:02d}:{:02d}:{:02d}'.format(int(t_h), int(t_m),
int(t_s))
if rank == 0:
if (index + 1) % args.print_freq == 0:
logger.info(
'Epoch: [{}/{}][{}/{}] '
'Data {data_time.val:.3f} ({data_time.avg:.3f}) '
'Batch {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Remain {remain_time} '
'rLoss {r_loss_meter.val:.4f} '
'tLoss {t_loss_meter.val:.4f} '
'projLoss {proj_loss_meter.val:.4f} '
'tripleLoss {triple_loss_meter.val:.4f} '
'rfineLoss {rfine_loss_meter.val:.4f} '
'tfineLoss {tfine_loss_meter.val:.4f} '
'Loss {loss_meter.val:.4f} '.format(
epoch,
args.epochs,
index + 1,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
remain_time=remain_time,
t_loss_meter=t_loss_meter,
r_loss_meter=r_loss_meter,
proj_loss_meter=proj_loss_meter,
triple_loss_meter=triple_loss_meter,
rfine_loss_meter=rfine_loss_meter,
tfine_loss_meter=tfine_loss_meter,
loss_meter=loss_meter))
writer.add_scalar('loss_train_batch_r', r_loss_meter.val,
current_iter)
writer.add_scalar('loss_train_batch_t', t_loss_meter.val,
current_iter)
if rank == 0:
logger.info('Train result at epoch [{}/{}].'.format(
epoch, args.epochs))
return t_loss_meter.avg, r_loss_meter.avg
def train(train_loader, model, criterion, optimizer, epoch, zoom_factor, batch_size, ins_weight):
batch_time = AverageMeter()
data_time = AverageMeter()
main_loss_meter = AverageMeter()
# aux_loss_meter = AverageMeter()
loss_meter = AverageMeter()
intersection_meter = AverageMeter()
union_meter = AverageMeter()
target_meter = AverageMeter()
model.train()
world_size = dist.get_world_size()
rank = dist.get_rank()
end = time.time()
for i, (pre_image_patch, pre_aug_mask, pre_ins_mask, flow_patch, inverse_flow_patch, pre_image_patch_2, pre_ins_mask_2, flow_patch_2, inverse_flow_patch_2, image_patch, ins_mask) in enumerate(train_loader):
#abandon the two objs.
pre_image_patch = None
pre_ins_mask = None
data_time.update(time.time() - end)
current_iter = (epoch - 1) * len(train_loader) + i + 1
max_iter = args.epochs * len(train_loader)
if args.net_type == 0:
index_split = 4
poly_learning_rate(optimizer, args.base_lr, current_iter, max_iter, power=args.power, index_split=index_split)
# pre image1.
# print (np.squeeze(pre_aug_mask.cpu().numpy()).shape) # (1, 1, 433, 433)
pre_tmp_out = np.squeeze(pre_aug_mask.cpu().numpy(), axis=0).transpose((1, 2, 0))
pre_tmp_out = np.squeeze(label_to_prob(pre_tmp_out, 1))
flow_patch = np.squeeze(flow_patch.cpu().numpy()).transpose((1, 2, 0))
inverse_flow_patch_numpy = np.squeeze(inverse_flow_patch.cpu().numpy()).transpose((1, 2, 0))
warp_pred = prob_to_label(flo.get_warp_label(flow_patch, inverse_flow_patch_numpy, pre_tmp_out))
pre = torch.from_numpy(warp_pred).contiguous().float()
warped_pred_aug_mask_var = torch.autograd.Variable(torch.unsqueeze(torch.unsqueeze(pre, dim=0), dim=0).cuda(async=True))
inverse_flow_patch_var = torch.autograd.Variable(inverse_flow_patch.cuda(async=True))
pre_input_var = torch.autograd.Variable(pre_image_patch_2.cuda(async=True))
# input model
pre_output_ins = model(pre_input_var, warped_pred_aug_mask_var, inverse_flow_patch_var)
seg_loss = 0
pre_ins_mask222 = torch.autograd.Variable(pre_ins_mask_2).squeeze(1).long()
pre_ins_mask_var = torch.autograd.Variable(pre_ins_mask222.cuda(async=True))
# #debug1
# import cv2
# # pre_image_patch_2_debug = np.squeeze(pre_image_patch_2.cpu().numpy()).transpose((1, 2, 0))
# warped_pred_aug_mask_var_debug = torch.unsqueeze(pre, dim=0).cpu().numpy().transpose((1, 2, 0))
# pre_ins_mask_var_debug = np.squeeze(pre_ins_mask_2.cpu().numpy())
# warped_pred_aug_mask_var_debug[warped_pred_aug_mask_var_debug == 255] = 0
# pre_ins_mask_var_debug[pre_ins_mask_var_debug == 255] = 0
# # cv2.imwrite("debug/"+str(rank)+ "__" + str(i) +"pre_patch.jpg", pre_image_patch_2_debug)
# cv2.imwrite("debug/"+str(rank)+ "__" + str(i) +"pre_warped_patch.png", warped_pred_aug_mask_var_debug * 255)
# cv2.imwrite("debug/"+str(rank)+ "__" + str(i) +"pre_ins_mask.png", pre_ins_mask_var_debug * 255)
# #debug1 over
ins_loss = criterion(pre_output_ins, pre_ins_mask_var) / world_size
last_ins_loss = ins_loss
loss = seg_loss + ins_loss
loss = loss * 0.8 # tow loss weight.
# current image.
image_patch_var = torch.autograd.Variable(image_patch.cuda(async=True))
# pre_output_ins (1, 2 ,433, 433)
# pre_output_ins_var = torch.argmax(pre_output_ins, dim=1, keepdim=True).float()
pre_output_ins_var = torch.argmax(pre_output_ins[0], dim=0, keepdim=True).float()
# tmp_out = pre_output_ins_var.data.max(1)[1].cpu().numpy().transpose((1,2,0))
tmp_out = pre_output_ins_var.cpu().numpy().transpose((1,2,0))
# print (np.unique(tmp_out))
# cv2.imwrite("debug/"+str(rank)+ "__" + str(i) +"pre_out.png", tmp_out * 255)
tmp_prob = np.squeeze(label_to_prob(tmp_out, 1))
flow_patch_2 = np.squeeze(flow_patch_2.cpu().numpy()).transpose((1, 2, 0))
inverse_flow_patch_2_numpy = np.squeeze(inverse_flow_patch_2.cpu().numpy()).transpose((1, 2, 0))
#warp
warp_pred = prob_to_label(flo.get_warp_label(flow_patch_2, inverse_flow_patch_2_numpy, tmp_prob))
pre = torch.from_numpy(warp_pred).contiguous().float()
warped_pred_aug_mask_var = torch.autograd.Variable(torch.unsqueeze(torch.unsqueeze(pre, dim=0), dim=0).cuda(async=True))
# pred_aug_mask_var warp to this image.
inverse_flow_patch_var_2 = torch.autograd.Variable(inverse_flow_patch_2.cuda(async=True))
# input model
output_ins = model(image_patch_var, warped_pred_aug_mask_var, inverse_flow_patch_var_2)
ins_mask = torch.autograd.Variable(ins_mask).squeeze(1).long()
ins_mask_var = torch.autograd.Variable(ins_mask.cuda(async=True))
#debug2
# pre_image_patch_2_debug = np.squeeze(image_patch.cpu().numpy()).transpose((1, 2, 0))
# warped_pred_aug_mask_var_debug = torch.unsqueeze(pre, dim=0).cpu().numpy().transpose((1, 2, 0))
# pre_ins_mask_var_debug = np.squeeze(ins_mask.cpu().numpy())
# warped_pred_aug_mask_var_debug[warped_pred_aug_mask_var_debug == 255] = 0
# pre_ins_mask_var_debug[pre_ins_mask_var_debug == 255] = 0
# # cv2.imwrite("debug/"+str(rank)+ "__" + str(i) +"now_patch.jpg", pre_image_patch_2_debug)
# cv2.imwrite("debug/"+str(rank)+ "__" + str(i) +"now_warped_patch.png", warped_pred_aug_mask_var_debug * 255)
# cv2.imwrite("debug/"+str(rank)+ "__" + str(i) +"now_ins_mask.png", pre_ins_mask_var_debug * 255)
# #debug2 over
seg_loss = 0
ins_loss = criterion(output_ins, ins_mask_var) / world_size
loss = ins_loss + seg_loss
optimizer.zero_grad()
loss.backward()
average_gradients(model)
optimizer.step()
output = output_ins.data.max(1)[1].cpu().numpy()
target = ins_mask.cpu().numpy()
intersection, union, target = intersectionAndUnion(output, target, 2, args.ignore_label) # 1 = args.classes
intersection_meter.update(intersection)
union_meter.update(union)
target_meter.update(target)
accuracy = sum(intersection_meter.val) / (sum(target_meter.val) + 1e-10)
reduced_loss = last_ins_loss.data.clone()
reduced_main_loss = ins_loss.data.clone()
# reduced_aux_loss = ins_loss.data.clone() # ins_loss replace here.
dist.all_reduce(reduced_loss)
dist.all_reduce(reduced_main_loss)
# dist.all_reduce(reduced_aux_loss)
main_loss_meter.update(reduced_main_loss[0], image_patch.size(0))
# aux_loss_meter.update(reduced_aux_loss[0], input.size(0))
loss_meter.update(reduced_loss[0], image_patch.size(0))
batch_time.update(time.time() - end)
end = time.time()
# calculate remain time
remain_iter = max_iter - current_iter
remain_time = remain_iter * batch_time.avg
t_m, t_s = divmod(remain_time, 60)
t_h, t_m = divmod(t_m, 60)
remain_time = '{:02d}:{:02d}:{:02d}'.format(int(t_h), int(t_m), int(t_s))
if rank == 0:
if (i + 1) % args.print_freq == 0:
logger.info('Epoch: [{}/{}][{}/{}] '
'Data {data_time.val:.3f} ({data_time.avg:.3f}) '
'Batch {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Remain {remain_time} '
'cur ins Loss {main_loss_meter.val:.4f} '
'pre ins Loss {loss_meter.val:.4f} '
'Accuracy {accuracy:.4f}.'.format(epoch, args.epochs, i + 1, len(train_loader),
batch_time=batch_time,
data_time=data_time,
remain_time=remain_time,
main_loss_meter=main_loss_meter,
loss_meter=loss_meter,
accuracy=accuracy))
writer.add_scalar('loss_train_batch', main_loss_meter.val, current_iter)
writer.add_scalar('mIoU_train_batch', np.mean(intersection / (union + 1e-10)), current_iter)
writer.add_scalar('mAcc_train_batch', np.mean(intersection / (target + 1e-10)), current_iter)
writer.add_scalar('allAcc_train_batch', accuracy, current_iter)
iou_class = intersection_meter.sum / (union_meter.sum + 1e-10)
accuracy_class = intersection_meter.sum / (target_meter.sum + 1e-10)
mIoU = np.mean(iou_class)
mAcc = np.mean(accuracy_class)
allAcc = sum(intersection_meter.sum) / (sum(target_meter.sum) + 1e-10)
if rank == 0:
logger.info(
'Train result at epoch [{}/{}]: mIoU/mAcc/allAcc {:.4f}/{:.4f}/{:.4f}.'.format(epoch, args.epochs, mIoU,
mAcc, allAcc))
return main_loss_meter.avg, mIoU, mAcc, allAcc