def __init__(self, node_name=''):
self.rosparam_(node_name)
self.create_logger_(node_name)
self.logger.info('ROS parameters loaded successfully')
self.logger.info(
'********************** Start logging **********************')
save_config_to_file(cfg, logger=self.logger)
self.logger.info('ROS buffer_len: {}'.format(self.buffer_len))
self.logger.info('ROS lidar_topic: {}'.format(self.lidar_topic))
self.logger.info('ROS debug_flag: {}'.format(self.debug_flag))
self.logger.info(
'ROS depth_threshold: {}'.format(self.depth_threshold))
self.logger.info(
'ROS score_threshold: {}'.format(self.score_threshold))
self.logger.info('ROS model_checkpoint: {}'.format(self.model_checkpoint))
self.model = PointRCNN(num_classes=self.num_class,
use_xyz=True, mode='TEST')
self.model.cuda()
load_checkpoint(
self.model,
filename=str(self.base_dir / self.model_checkpoint),
logger=self.logger)
self.model.eval()
self.logger.info('Model initialization complete')
def repeat_eval_ckpt(root_result_dir, ckpt_dir):
root_result_dir = os.path.join(root_result_dir, 'eval', 'eval_all_' + args.extra_tag)
os.makedirs(root_result_dir, exist_ok=True)
log_file = os.path.join(root_result_dir, 'log_eval_all_%s.txt' % cfg.TEST.SPLIT)
logger = create_logger(log_file)
logger.info('**********************Start logging**********************')
# save config
for key, val in vars(args).items():
logger.info("{:16} {}".format(key, val))
save_config_to_file(cfg, logger=logger)
# create dataloader & network
test_loader = create_dataloader(logger)
model = PointRCNN(num_classes=test_loader.dataset.num_class, use_xyz=True, mode='TEST')
model.cuda()
# copy important files to backup
backup_dir = os.path.join(root_result_dir, 'backup_files')
os.makedirs(backup_dir, exist_ok=True)
os.system('cp *.py %s/' % backup_dir)
os.system('cp ../lib/net/*.py %s/' % backup_dir)
os.system('cp ../lib/datasets/kitti_rcnn_dataset.py %s/' % backup_dir)
# evaluated ckpt record
ckpt_record_file = os.path.join(root_result_dir, 'eval_list_%s.txt' % cfg.TEST.SPLIT)
with open(ckpt_record_file, 'a'):
pass
# tensorboard log
tb_log = SummaryWriter(log_dir=os.path.join(root_result_dir, 'tensorboard_%s' % cfg.TEST.SPLIT))
while True:
# check whether there is checkpoint which is not evaluated
cur_epoch_id, cur_ckpt = get_no_evaluated_ckpt(ckpt_dir, ckpt_record_file)
if cur_epoch_id == -1 or int(float(cur_epoch_id)) < args.start_epoch:
wait_second = 30
print('Wait %s second for next check: %s' % (wait_second, ckpt_dir))
time.sleep(wait_second)
continue
# load checkpoint
train_utils.load_checkpoint(model, filename=cur_ckpt)
# start evaluation
cur_result_dir = os.path.join(root_result_dir, 'epoch_%s' % cur_epoch_id, cfg.TEST.SPLIT)
tb_dict = eval_one_epoch(model, test_loader, cur_epoch_id, cur_result_dir, logger)
step = int(float(cur_epoch_id))
if step == float(cur_epoch_id):
for key, val in tb_dict.items():
tb_log.add_scalar(key, val, step)
# record this epoch which has been evaluated
with open(ckpt_record_file, 'a') as f:
print('%s' % cur_epoch_id, file=f)
logger.info('Epoch %s has been evaluated' % cur_epoch_id)
def load_ckpt_based_on_args(model, logger):
if args.ckpt is not None:
train_utils.load_checkpoint(model, filename=args.ckpt, logger=logger)
total_keys = model.state_dict().keys().__len__()
if cfg.RPN.ENABLED and args.rpn_ckpt is not None:
load_part_ckpt(model, filename=args.rpn_ckpt, logger=logger, total_keys=total_keys)
if cfg.RCNN.ENABLED and args.rcnn_ckpt is not None:
load_part_ckpt(model, filename=args.rcnn_ckpt, logger=logger, total_keys=total_keys)
def load_ckpt_based_on_cfg(config, model, logger):
if config['ckpt'] is not None:
train_utils.load_checkpoint(model,
filename=config['ckpt'],
logger=logger)
total_keys = model.state_dict().keys().__len__()
if cfg.RPN.ENABLED and config['rpn_ckpt'] is not None:
load_part_ckpt(model,
filename=config['rpn_ckpt'],
logger=logger,
total_keys=total_keys)
if cfg.RCNN.ENABLED and config['rcnn_ckpt'] is not None:
load_part_ckpt(model,
filename=config['rcnn_ckpt'],
logger=logger,
total_keys=total_keys)
def load_ckpt_based_on_args(model, logger):
"""
Input: model and logger instance
Output: None
Task: Loads ckpt based on the args --rpn_ckpt and --rcnn_ckpt
"""
if args.ckpt is not None:
train_utils.load_checkpoint(model, filename=args.ckpt, logger=logger)
total_keys = model.state_dict().keys().__len__()
if cfg.RPN.ENABLED and args.rpn_ckpt is not None:
load_part_ckpt(model,
filename=args.rpn_ckpt,
logger=logger,
total_keys=total_keys)
if cfg.RCNN.ENABLED and args.rcnn_ckpt is not None:
load_part_ckpt(model,
filename=args.rcnn_ckpt,
logger=logger,
total_keys=total_keys)
def train():
print(args.local_rank)
torch.cuda.set_device(args.local_rank)
# create dataloader & network & optimizer
model, model_fn_decorator, net_func = create_model(cfg)
init_weights(model, init_type='kaiming')
model.cuda()
root_result_dir = args.output_dir
os.makedirs(root_result_dir, exist_ok=True)
log_file = os.path.join(root_result_dir, "log_train.txt")
logger = create_logger(log_file, get_rank())
logger.info("**********************Start logging**********************")
logger.info('TRAINED MODEL:{}'.format(net_func))
# log to file
gpu_list = os.environ[
'CUDA_VISIBLE_DEVICES'] if 'CUDA_VISIBLE_DEVICES' in os.environ.keys(
) else 'ALL'
logger.info("CUDA_VISIBLE_DEVICES=%s" % gpu_list)
for key, val in vars(args).items():
logger.info("{:16} {}".format(key, val))
logger.info("***********************config infos**********************")
for key, val in vars(cfg).items():
logger.info("{:16} {}".format(key, val))
# log tensorboard
if get_rank() == 0:
tb_log = SummaryWriter(
log_dir=os.path.join(root_result_dir, "tensorboard"))
else:
tb_log = None
train_loader, test_loader = create_dataloader()
# train_loader, test_loader = create_dataloader_Insensee()
optimizer = create_optimizer(model)
# load checkpoint if it is possible
start_epoch = it = best_res = 0
last_epoch = -1
if args.ckpt is not None:
pure_model = model
it, start_epoch, best_res = load_checkpoint(pure_model, optimizer,
args.ckpt, logger)
last_epoch = start_epoch + 1
lr_scheduler = create_scheduler(optimizer, last_epoch=last_epoch)
# lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.98, last_epoch=-1)
criterion = None
# start training
logger.info('**********************Start training**********************')
ckpt_dir = os.path.join(root_result_dir, "ckpt")
os.makedirs(ckpt_dir, exist_ok=True)
trainer = train_utils.Trainer(model,
model_fn=model_fn_decorator(),
criterion=criterion,
optimizer=optimizer,
ckpt_dir=ckpt_dir,
lr_scheduler=lr_scheduler,
model_fn_eval=model_fn_decorator(),
tb_log=tb_log,
logger=logger,
eval_frequency=1,
cfg=cfg)
trainer.train(start_it=it,
start_epoch=start_epoch,
n_epochs=args.epochs,
train_loader=train_loader,
test_loader=test_loader,
ckpt_save_interval=args.ckpt_save_interval,
best_res=best_res)
logger.info('**********************End training**********************')
use_xyz=True,
mode='TRAIN')
optimizer = create_optimizer(model)
if args.mgpus:
model = nn.DataParallel(model)
model.cuda()
# load checkpoint if it is possible
start_epoch = it = 0
last_epoch = -1
if args.ckpt is not None:
pure_model = model.module if isinstance(
model, torch.nn.DataParallel) else model
it, start_epoch = train_utils.load_checkpoint(pure_model,
optimizer,
filename=args.ckpt,
logger=logger)
last_epoch = start_epoch + 1
lr_scheduler, bnm_scheduler = create_scheduler(
optimizer,
total_steps=len(train_loader) * args.epochs,
last_epoch=last_epoch)
if args.rpn_ckpt is not None:
pure_model = model.module if isinstance(
model, torch.nn.DataParallel) else model
total_keys = pure_model.state_dict().keys().__len__()
train_utils.load_part_ckpt(pure_model,
filename=args.rpn_ckpt,
logger=logger,
# create dataloader & network & optimizer
train_loader, test_loader = create_dataloader(logger)
model = PointRCNN(num_classes=train_loader.dataset.num_class, use_xyz=True, mode='TRAIN')
optimizer = create_optimizer(model)
if args.mgpus:
model = nn.DataParallel(model)
model.cuda()
# load checkpoint if it is possible
start_iter = it = 0
last_iter = -1
if args.pretrain_ckpt is not None:
pure_model = model.module if isinstance(model, torch.nn.DataParallel) else model
_, _ = train_utils.load_checkpoint(pure_model, None, filename=args.pretrain_ckpt, logger=logger)
it = int(args.total_iters*9/30)
if args.ckpt is not None:
pure_model = model.module if isinstance(model, torch.nn.DataParallel) else model
it, _ = train_utils.load_checkpoint(pure_model, optimizer, filename=args.ckpt, logger=logger)
last_iter = it + 1
lr_scheduler, bnm_scheduler = create_scheduler(optimizer, total_steps=args.total_iters, last_iter=last_iter)
lr_warmup_scheduler = None
# start training
logger.info('**********************Start training**********************')
logger.info('experiment ID: %s/%s/%s'%(root_result_dir.split('/')[-3], root_result_dir.split('/')[-2], root_result_dir.split('/')[-1]))
