def detect_specific_frame():
args = parse_config()
output_dir = os.path.join(cfg.CODE_DIR, "output", cfg.TAG, args.extra_tag)
os.makedirs(output_dir, exist_ok=True)
eval_output_dir = os.path.join(output_dir, "eval")
os.makedirs(eval_output_dir, exist_ok=True)
log_file = os.path.join(
eval_output_dir,
"log_eval_%s.txt" % datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
logger = common_utils.create_logger(log_file)
test_set, test_dataloader, sampler = build_data_loader(
dist=False,
data_path=cfg.DATA_DIR,
batch_size=args.batch_size,
num_workers=cfg.DATA_CONFIG.NUM_WORKERS,
logger=logger,
split=args.split,
training=False)
model = build_netword(num_class=len(cfg.CLASS_NAMES),
dataset=test_set,
logger=logger)
model.load_params_from_file(args.ckpt, logger=logger)
# for fix,id in enumerate(test_set.sample_idx_list):
# if id == "000137":
# break
for i, data in enumerate(test_dataloader):
if i == 58:
break
input = example_convert_to_torch(data)
model.cuda()
model.eval()
output, _ = model(input)
print(output)
def origin_model():
args = parse_config()
ckpt_dir = args.ckpt
log_file = os.path.join(
args.log_path,
"log_eval_%s.txt" % datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
logger = common_utils.create_logger(log_file)
prepocess_model = PrepocessData()
test_set, test_dataloader, sampler = build_data_loader(
dist=False,
data_path=cfg.DATA_DIR,
batch_size=args.batch_size,
num_workers=cfg.DATA_CONFIG.NUM_WORKERS,
logger=None,
split=args.split,
training=False)
model = build_netword(num_class=len(cfg.CLASS_NAMES),
dataset=test_set,
logger=logger)
model.load_params_from_file(filename=ckpt_dir, logger=logger)
with torch.no_grad():
test_dataloader_iter = iter(test_dataloader)
model.cuda()
model.eval()
data = next(test_dataloader_iter)
input_dict = example_convert_to_torch(data)
points = parpare_point_cloud() #取出点云
fov_flag = get_fov_flag(points, prepocess_model.image_shape,
prepocess_model.get_calib()) #过滤点云
points = points[fov_flag]
inputs = prepocess_model.points2voxel(points)
pred_dict, ret_dict = model(input_dict)
# script_model = torch.jit.trace(model, input_dict)
print("done")
def save_model():
args = parse_config()
ckpt_dir = args.ckpt
log_file = os.path.join(
args.log_path,
"log_eval_%s.txt" % datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
logger = common_utils.create_logger(log_file)
prepocess_model = PrepocessData()
model = DetNet(4)
model.load_params_from_file(filename=ckpt_dir, logger=logger)
points = parpare_point_cloud()
fov_flag = get_fov_flag(points, prepocess_model.image_shape,
prepocess_model.get_calib())
points = points[fov_flag]
with torch.set_grad_enabled(False):
inputs = prepocess_model.points2voxel(points)
model.cuda()
model.eval()
output = model(inputs)
# torch.save(model,"DetNet.pkl")
print("done")
def main():
args = parse_config()
output_dir = os.path.join(cfg.CODE_DIR, "output", cfg.TAG, args.extra_tag)
os.makedirs(output_dir, exist_ok=True)
eval_output_dir = os.path.join(output_dir, "eval")
if not args.eval_all:
# num_list = re.findall(r"\d+",args.ckpt_id) if args.ckpt_id is not None else []
# epoch_id = num_list[-1] if num_list.__len__()>0 else "no number"
epoch_id = args.ckpt_id
eval_output_dir = os.path.join(eval_output_dir, "epoch_%s" % epoch_id)
else:
eval_output_dir = os.path.join(eval_output_dir, "eval_all_default")
os.makedirs(eval_output_dir, exist_ok=True)
log_file = os.path.join(
eval_output_dir,
"log_eval_%s.txt" % datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
logger = common_utils.create_logger(log_file)
#log to file
logger.info("*****************start logging****************")
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)
#非分布式训练
dist_test = False
if dist_test:
total_gpus = dist.get_world_size()
logger.info("total_batch_size:%d" % (total_gpus * args.batch_size))
for key, val in vars(args).items():
logger.info("{:16} {}".format(key, val))
log_cfg_to_file(cfg, logger=logger)
ckpt_dir = args.ckpt
test_set, test_dataloader, sampler = build_data_loader(
dist=dist_test,
data_path=cfg.DATA_DIR,
batch_size=args.batch_size,
num_workers=cfg.DATA_CONFIG.NUM_WORKERS,
logger=logger,
split=args.split,
training=False,
args=args)
model = build_netword(num_class=len(cfg.CLASS_NAMES),
dataset=test_set,
logger=logger)
with torch.no_grad():
if args.eval_all:
repeat_eval_ckpt(model, test_dataloader, cfg, eval_output_dir,
logger, ckpt_dir)
else:
eval_single_ckpt(model, test_dataloader, eval_output_dir, logger,
ckpt_dir, epoch_id, args)
def main():
args = parse_config()
dist_train = False # 非分布式运行
if args.launcher=="pytorch":
args.batch_size, cfg.LOCAL_RANK = getattr(common_utils, 'init_dist_%s' % args.launcher)(
args.batch_size, args.tcp_port, args.local_rank, backend='nccl'
)
dist_train = True
output_dir = os.path.join(cfg.CODE_DIR,"output",args.TAG,args.split,args.extra_tag)
os.makedirs(output_dir,exist_ok=True)
ckpt_dir = args.ckpt
os.makedirs(ckpt_dir,exist_ok=True)
log_file = os.path.join(output_dir,("log_train_%s.txt" % datetime.datetime.now().strftime("%Y%m%d-%H%M%S")))
logger = common_utils.create_logger(log_file)
logger.info("*********************start logging********************************")
gpu_list = os.environ["CUDA_VISIBLE_DEVICES"] if "CUDA_VISIBLE_DEVICES" in os.environ.keys() else "ALL"
log_cfg_to_file(cfg,logger=logger)
tb_log_path = os.path.join(output_dir,"tensorboard")
os.makedirs(tb_log_path,exist_ok=True)
tb_log = SummaryWriter(log_dir=tb_log_path if args.local_rank ==0 else None)
dataset,data_loader,sampler= build_data_loader(dist=dist_train,
batch_size=args.batch_size,
num_workers=args.workers,
training=True,
logger=logger,
split=args.split,
args = args,
train_all=args.train_all)
if args.detector_name=="LZnet":
model = FPVdet(dataset=dataset,logger=logger)
elif args.detector_name=="pvrcnn":
model = Part2net(num_class=len(cfg.CLASS_NAMES),dataset=dataset)
else:
raise NotImplementedError
model.cuda()
optimizer = build_optimizer(model, cfg.MODEL.OPTIMIZATION)
start_epoch = it = 0
last_epoch = -1
if args.pretrained_model is not None:
model.load_params_from_file(filename=args.pretrained_model,to_cpu=dist_train)
ckpt_list = glob.glob(os.path.join(ckpt_dir, "*checkpoint_epoch_*.pth"))
if len(ckpt_list) > 0:
ckpt_list.sort(key=os.path.getatime)
it, start_epoch = model.load_params_with_optimizer(ckpt_list[-1], to_cpu=dist,
optimizer=optimizer, logger=logger)
last_epoch = start_epoch + 1
model.train()
#logger.info(model)
if dist_train:
model = nn.parallel.DistributedDataParallel(model,device_ids=[args.local_rank])
lr_scheduler,lr_warmup_scheduler = build_scheduler(optimizer,
total_iters_each_epoch=len(data_loader),
total_epochs=args.epochs,
last_epoch=last_epoch,
optim_cfg=cfg.MODEL.OPTIMIZATION,
)
logger.info("************start training*************")
train_model(
model,
optimizer,
data_loader,
model_func=model_fn_decorator(),
lr_scheduler=lr_scheduler,
optim_cfg=cfg.MODEL.OPTIMIZATION,
start_epoch=start_epoch,
total_epochs=args.epochs,
start_iter=it,
rank=args.local_rank,
tb_log=tb_log,
ckpt_save_dir=ckpt_dir,
train_sampler=sampler,
lr_warmup_scheduler = lr_warmup_scheduler,
ckpt_save_interval=args.ckpt_save_interval ,
max_ckpt_save_num=args.max_ckpt_save_num
)
logger.info("**************End training********************")