def main():
global args
args = parser.parse_args()
if args.num_samples == 0:
args.num_samples = None
if args.val_batch_size is None:
args.val_batch_size = args.batch_size
if args.seed:
random.seed(args.seed)
torch.manual_seed(args.seed)
# torch.backends.cudnn.deterministic = True
# warnings.warn('You have chosen to seed training. '
# 'This will turn on the CUDNN deterministic setting, '
# 'which can slow down your training considerably! '
# 'You may see unexpected behavior when restarting from checkpoints.')
# For distributed training
# init_distributed_mode(args)
if not args.no_cuda and not torch.cuda.is_available():
raise Exception("No gpu available for usage")
torch.backends.cudnn.benchmark = args.cudnn
# Init model
channels_in = 1 if args.input_type == 'depth' else 4
model = Models.define_model(mod=args.mod,
in_channels=channels_in,
thres=args.thres)
define_init_weights(model, args.weight_init)
# Load on gpu before passing params to optimizer
if not args.no_cuda:
if not args.multi:
model = model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# model.cuda()
# model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
# model = model.module
save_id = '{}_{}_{}_{}_{}_batch{}_pretrain{}_wlid{}_wrgb{}_wguide{}_wpred{}_patience{}_num_samples{}_multi{}'.\
format(args.mod, args.optimizer, args.loss_criterion,
args.learning_rate,
args.input_type,
args.batch_size,
args.pretrained, args.wlid, args.wrgb, args.wguide, args.wpred,
args.lr_decay_iters, args.num_samples, args.multi)
# INIT optimizer/scheduler/loss criterion
optimizer = define_optim(args.optimizer, model.parameters(),
args.learning_rate, args.weight_decay)
scheduler = define_scheduler(optimizer, args)
# Optional to use different losses
criterion_local = define_loss(args.loss_criterion)
criterion_lidar = define_loss(args.loss_criterion)
criterion_rgb = define_loss(args.loss_criterion)
criterion_guide = define_loss(args.loss_criterion)
# INIT dataset
dataset = Datasets.define_dataset(args.dataset, args.data_path,
args.input_type, args.side_selection)
dataset.prepare_dataset()
train_loader, valid_loader, valid_selection_loader = get_loader(
args, dataset)
# Resume training
best_epoch = 0
lowest_loss = np.inf
args.save_path = os.path.join(args.save_path, save_id)
mkdir_if_missing(args.save_path)
log_file_name = 'log_train_start_0.txt'
args.resume = first_run(args.save_path)
if args.resume and not args.test_mode and not args.evaluate:
path = os.path.join(
args.save_path,
'checkpoint_model_epoch_{}.pth.tar'.format(int(args.resume)))
if os.path.isfile(path):
log_file_name = 'log_train_start_{}.txt'.format(args.resume)
# stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(path)
args.start_epoch = checkpoint['epoch']
lowest_loss = checkpoint['loss']
best_epoch = checkpoint['best epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
log_file_name = 'log_train_start_0.txt'
# stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
print("=> no checkpoint found at '{}'".format(path))
# Only evaluate
elif args.evaluate:
print("Evaluate only")
best_file_lst = glob.glob(os.path.join(args.save_path, 'model_best*'))
if len(best_file_lst) != 0:
best_file_name = best_file_lst[0]
print(best_file_name)
if os.path.isfile(best_file_name):
sys.stdout = Logger(
os.path.join(args.save_path, 'Evaluate.txt'))
print("=> loading checkpoint '{}'".format(best_file_name))
checkpoint = torch.load(best_file_name)
model.load_state_dict(checkpoint['state_dict'])
else:
print("=> no checkpoint found at '{}'".format(best_file_name))
else:
print("=> no checkpoint found at due to empy list in folder {}".
format(args.save_path))
validate(valid_selection_loader, model, criterion_lidar, criterion_rgb,
criterion_local, criterion_guide)
return
# Start training from clean slate
else:
# Redirect stdout
sys.stdout = Logger(os.path.join(args.save_path, log_file_name))
# INIT MODEL
print(40 * "=" + "\nArgs:{}\n".format(args) + 40 * "=")
print("Init model: '{}'".format(args.mod))
print("Number of parameters in model {} is {:.3f}M".format(
args.mod.upper(),
sum(tensor.numel() for tensor in model.parameters()) / 1e6))
# Load pretrained state for cityscapes in GLOBAL net
if args.pretrained and not args.resume:
if not args.load_external_mod:
if not args.multi:
target_state = model.depthnet.state_dict()
else:
target_state = model.module.depthnet.state_dict()
check = torch.load('erfnet_pretrained.pth')
for name, val in check.items():
# Exclude multi GPU prefix
mono_name = name[7:]
if mono_name not in target_state:
continue
try:
target_state[mono_name].copy_(val)
except RuntimeError:
continue
print('Successfully loaded pretrained model')
else:
check = torch.load('external_mod.pth.tar')
lowest_loss_load = check['loss']
target_state = model.state_dict()
for name, val in check['state_dict'].items():
if name not in target_state:
continue
try:
target_state[name].copy_(val)
except RuntimeError:
continue
print("=> loaded EXTERNAL checkpoint with best rmse {}".format(
lowest_loss_load))
# Start training
for epoch in range(args.start_epoch, args.nepochs):
print("\n => Start EPOCH {}".format(epoch + 1))
print(datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
print(args.save_path)
# Adjust learning rate
if args.lr_policy is not None and args.lr_policy != 'plateau':
scheduler.step()
lr = optimizer.param_groups[0]['lr']
print('lr is set to {}'.format(lr))
# Define container objects
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
score_train = AverageMeter()
score_train_1 = AverageMeter()
metric_train = Metrics(max_depth=args.max_depth,
disp=args.use_disp,
normal=args.normal)
# Train model for args.nepochs
model.train()
# compute timing
end = time.time()
# Load dataset
for i, (input, gt) in tqdm(enumerate(train_loader)):
# Time dataloader
data_time.update(time.time() - end)
# Put inputs on gpu if possible
if not args.no_cuda:
input, gt = input.cuda(), gt.cuda()
prediction, lidar_out, precise, guide = model(input, epoch)
loss = criterion_local(prediction, gt)
loss_lidar = criterion_lidar(lidar_out, gt)
loss_rgb = criterion_rgb(precise, gt)
loss_guide = criterion_guide(guide, gt)
loss = args.wpred * loss + args.wlid * loss_lidar + args.wrgb * loss_rgb + args.wguide * loss_guide
losses.update(loss.item(), input.size(0))
metric_train.calculate(prediction[:, 0:1].detach(), gt.detach())
score_train.update(metric_train.get_metric(args.metric),
metric_train.num)
score_train_1.update(metric_train.get_metric(args.metric_1),
metric_train.num)
# Clip gradients (usefull for instabilities or mistakes in ground truth)
if args.clip_grad_norm != 0:
nn.utils.clip_grad_norm(model.parameters(),
args.clip_grad_norm)
# Setup backward pass
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Time trainig iteration
batch_time.update(time.time() - end)
end = time.time()
# Print info
if (i + 1) % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Metric {score.val:.4f} ({score.avg:.4f})'.format(
epoch + 1,
i + 1,
len(train_loader),
batch_time=batch_time,
loss=losses,
score=score_train))
print("===> Average RMSE score on training set is {:.4f}".format(
score_train.avg))
print("===> Average MAE score on training set is {:.4f}".format(
score_train_1.avg))
# Evaulate model on validation set
print("=> Start validation set")
score_valid, score_valid_1, losses_valid = validate(
valid_loader, model, criterion_lidar, criterion_rgb,
criterion_local, criterion_guide, epoch)
print("===> Average RMSE score on validation set is {:.4f}".format(
score_valid))
print("===> Average MAE score on validation set is {:.4f}".format(
score_valid_1))
# Evaluate model on selected validation set
if args.subset is None:
print("=> Start selection validation set")
score_selection, score_selection_1, losses_selection = validate(
valid_selection_loader, model, criterion_lidar, criterion_rgb,
criterion_local, criterion_guide, epoch)
total_score = score_selection
print("===> Average RMSE score on selection set is {:.4f}".format(
score_selection))
print("===> Average MAE score on selection set is {:.4f}".format(
score_selection_1))
else:
total_score = score_valid
print("===> Last best score was RMSE of {:.4f} in epoch {}".format(
lowest_loss, best_epoch))
# Adjust lr if loss plateaued
if args.lr_policy == 'plateau':
scheduler.step(total_score)
lr = optimizer.param_groups[0]['lr']
print('LR plateaued, hence is set to {}'.format(lr))
# File to keep latest epoch
with open(os.path.join(args.save_path, 'first_run.txt'), 'w') as f:
f.write(str(epoch))
# Save model
to_save = False
if total_score < lowest_loss:
to_save = True
best_epoch = epoch + 1
lowest_loss = total_score
save_checkpoint(
{
'epoch': epoch + 1,
'best epoch': best_epoch,
'arch': args.mod,
'state_dict': model.state_dict(),
'loss': lowest_loss,
'optimizer': optimizer.state_dict()
}, to_save, epoch)
if not args.no_tb:
writer.close()
def main():
global args
args = parser.parse_args()
if args.num_samples == 0:
args.num_samples = None
if args.cuda and not torch.cuda.is_available():
raise Exception("No gpu available for usage")
# Init model
channels_in = 1 if args.input_type == 'depth' else 4
model = Models.define_model(mod=args.mod, in_channels=channels_in)
if args.mod == 'mod':
define_init_weights(model, args.weight_init)
# Load on gpu before passing params to optimizer
if args.cuda:
model = model.cuda()
save_id = '{}_{}_{}_{}_batch{}_pretrain{}_wlid{}_wrgb{}_wguide{}_wpred{}_num_samples{}'.\
format(args.mod, args.loss_criterion_source,
args.learning_rate,
args.input_type,
args.batch_size,
args.load_path!='', args.wlid, args.wrgb, args.wguide, args.wpred,
args.num_samples)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.learning_rate,
weight_decay=args.weight_decay)
# Optional to use different losses
criterion_source = define_loss(args.loss_criterion_source)
criterion_target = define_loss(args.loss_criterion_target)
# INIT KITTI dataset
print('Load KITTI')
dataset = Datasets.define_dataset('kitti', args.data_path_target,
args.input_type)
dataset.prepare_dataset()
train_loader = get_loader(args, dataset, only_train=True)
# INIT Carla dataset
print('Load Carla')
dataset = Datasets.define_dataset('carla', args.data_path_source,
args.input_type)
dataset.prepare_dataset()
# The sparsification of the data and projection from the LiDAR reference
# frame to the RGB camera explained in the paper happens in the dataloader
train_loader_carla = get_loader(args,
dataset,
is_carla=True,
only_train=True)
train_loader_iter = iter(train_loader)
# Resume training
if args.save_name == '':
args.save_path = os.path.join(args.save_path, save_id)
else:
args.save_path = os.path.join(args.save_path, args.save_name)
if os.path.exists(args.save_path):
raise Exception('Save path already exists')
mkdir_if_missing(args.save_path)
# INIT MODEL
print(40 * "=" + "\nArgs:{}\n".format(args) + 40 * "=")
print("Init model: '{}'".format(args.mod))
print("Number of parameters in model {} is {:.3f}M".format(
args.mod.upper(),
sum(tensor.numel() for tensor in model.parameters()) / 1e6))
# Load pretrained state
if args.load_path != '':
print("=> loading checkpoint {:s}".format(args.load_path))
check = torch.load(
args.load_path,
map_location=lambda storage, loc: storage)['state_dict']
model.load_state_dict(check)
if args.use_image_translation:
image_trans_net = ResnetGeneratorCycle(3, 3, 64, n_blocks=9)
state_dict = torch.load('./image_translation_weights.pth')
image_trans_net.load_state_dict(state_dict)
image_trans_net.eval()
if args.cuda:
image_trans_net = image_trans_net.cuda()
# Start training
global_step = 0
for epoch in range(args.start_epoch, args.nepochs):
print("\n => Start EPOCH {}".format(epoch + 1))
# Define container objects
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
score_train_rmse = AverageMeter()
score_train_mae = AverageMeter()
metric_train = Metrics(max_depth=args.max_depth)
# Train model for args.nepochs
model.train()
# compute timing
end = time.time()
for i, (input, gt, filepath) in tqdm(enumerate(train_loader_carla)):
# Time dataloader
data_time.update(time.time() - end)
loss_extra = 0
# Put inputs on gpu if possible
if args.cuda:
input, gt = input.cuda(), gt.cuda()
# The LiDAR depths have large regions where no input depth is given
# We remove all of the GT in the synthetic data where no input information is given
# in a NxN window around the GT point (we set N=41) to avoid the model trying to estimate
# depth for areas without any input guidance
input_depth = input[:, 0:1]
input_depth, gt = filter_data(input_depth,
gt,
max_depth=args.max_depth)
input[:, 0:1] = input_depth
### Load target set (KITTI) data
if args.train_target:
try:
input_target, gt_target, filepath_t = next(
train_loader_iter)
except:
train_loader_iter = iter(train_loader)
input_target, gt_target, filepath_t = next(
train_loader_iter)
if args.cuda:
input_target, gt_target = input_target.cuda(
), gt_target.cuda()
if args.use_image_translation:
# The CycleGAN model was trained with inputs in the range of [-1, 1]
with torch.no_grad():
rgb_trans = image_trans_net(input[:, 1:] / 128.5 - 1)
rgb_trans = 128.5 * (rgb_trans + 1)
rgb_trans = rgb_trans.clamp(0, 255)
input = torch.cat([input[:, :1], rgb_trans], 1)
if args.train_target:
input_joint = torch.cat([input, input_target])
prediction, lidar_out, precise, guide = model(
input_joint, epoch)
# We separate predictions from the target domain and source domain
prediction_target, lidar_out_target, precise_target, guide_target = prediction[
args.batch_size:], lidar_out[args.batch_size:], precise[
args.batch_size:], guide[args.batch_size:]
prediction, lidar_out, precise, guide = prediction[:args.
batch_size], lidar_out[:
args
.
batch_size], precise[:
args
.
batch_size], guide[:
args
.
batch_size]
else:
prediction, lidar_out, precise, guide = model(input, epoch)
# We compute the loss for the source domain data
loss = criterion_source(prediction, gt)
loss_lidar = criterion_source(lidar_out, gt)
loss_rgb = criterion_source(precise, gt)
loss_guide = criterion_source(guide, gt)
loss = args.wpred * loss + args.wlid * loss_lidar + args.wrgb * loss_rgb + args.wguide * loss_guide
if args.train_target:
loss_target = 0
# We filter the input data for supervision as explained in the paper
filtered_sparse_data = filter_sparse_guidance(
input_target[:, :1], args.filter_window, args.filter_th)
# We compute the loss for the target domain data
loss_target += args.wpred * (criterion_target(
prediction_target, filtered_sparse_data))
loss_target += args.wlid * (criterion_target(
lidar_out_target, filtered_sparse_data))
loss_target += args.wrgb * (criterion_target(
precise_target, filtered_sparse_data))
loss_target += args.wguide * (criterion_target(
guide_target, filtered_sparse_data))
loss = loss + loss_target
metric_train.calculate(prediction[:, 0:1].detach(), gt.detach())
score_train_rmse.update(metric_train.get_metric('rmse'),
metric_train.num)
score_train_mae.update(metric_train.get_metric('mae'),
metric_train.num)
losses.update(loss.item(), input.size(0))
# Optimization step
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time.update(time.time() - end)
end = time.time()
global_step += 1
# Print info
if (i + 1) % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'RMSE Train {score.val:.4f} ({score.avg:.4f})'.format(
epoch + 1,
i + 1,
len(train_loader_carla),
batch_time=batch_time,
loss=losses,
score=score_train_rmse))
if global_step == args.n_training_iterations:
dict_save = {
'epoch': epoch + 1,
'arch': args.mod,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(dict_save, False, epoch + 1, global_step)
return 1
print("===> Average RMSE score on training set is {:.4f}".format(
score_train_rmse.avg))
print("===> Average MAE score on training set is {:.4f}".format(
score_train_mae.avg))
dict_save = {
'epoch': epoch + 1,
'arch': args.mod,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(dict_save, False, epoch + 1)