You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Linux:
sudo apt-get install openssh-server
sudo apt-get install openssh-client
cd /home/username/
ssh-keygen -t rsa -P ""
cd ~/.ssh
ssh-copy-id reomte-host-name@romote-host-ip
Note: windows system can not find 'ssh-copy-id' command, you should open the xxx.pub file (/C/xxx/xxx/.ssh/xxx.pub) of your windows computer, then copy the **contents** of the file to the authorized_keys file of the remote host manually. If your remote host does not have authorized_keys, you should manually create it before performing the copy operation:
cd ~/.ssh (Remote host)
touch authorized_keys
vim authorized_keys
Configure the name of the remote host to relize password-free login
cd ~/.ssh
touch config
vim config
Then add tehe following content to the config file:
Host server1(The nick name of the remote host)
HostName ip address(The ip address of remote host)
User root (The user name of remote host)
port 22 (The port of remote host)
10 Remote Debugging
Install Remote development
Display Image:
Install Remote x11 (https://blog.csdn.net/zb12138/article/details/107160825/)
Parallel Train
1 Single GPU
1 Set the visible GPU number
Method 1
CUDA_VISIBLE_DEVICES=0,1 python xxx.py
Method 2
vim xxx.py
os.environ['CUDA_VISIBLE_DEVICES']='0,1'
2 Put the model, data, and loss function on the GPU device
Method 1
model=model.cuda()
Loss=Loss.cuda()
Method 2
device=torch.device('cuda:{}'.format(args.gpu_id))
model=model.to(device)
2 DataParallel (Multi GPU)
model = torch.nn.DataParallel(model.cuda())
3 DistributedDataParallel (Multi GPU)
1 Run in terminal
python -m torch.distributed.launch --nproc_per_node=NUM_GPUS ./bin/dist_train.py
2 Parse parameters
parser.add_argument("--local_rank", type=int,default=0)
2 nodes, 2 GPUs for per node
rank=0,1,2,3
node1: local_rank= 0 or 1 (rank%node_num)
mode2: local_rank= 0 or 1
3 Initialize communication method
torch.distributed.init_process_group(backend='nccl', init_method='env://')
4 Set the GPU number that the current process needs to use
torch.cuda.set_device(args.local_rank)
5 Generate corresponding data labels for each proceess
train_sampler=torch.utils.data.distributed.DistributedSampler(train_data)
train_loader=torch.utils.data.DataLoader(train_data,batch_size=batch_size,shuffle=F
alse,num_workers=2, pin_memory=True, sampler=train_sampler)
Before each epoch, the shuffle effect is achieved by calling the following commands:
train_sampler.set_epoch(epoch)
6 Calculate the loss, summarize the information of each process
Example:
def reduce_tesnor(tensor):
# sum the tensor data across all machines
dist.all_reduce(rt. op=dist.reduce_op.SUM)
return rt
output=model(input)
loss=Loss(output,label)
log_loss = reduce_tensor(loss.clone().detach_())
torch.cuda.synchronize() # wait every process finish above transmission
loss_total += log_loss.item()
7 Avoid conflicts when writing log files or print
if args.local_rank==0:
print('xxxx')
log.info('xxxx')
SiamFC
1 SiamFC-VID
1 Train
python ./bin/my_train.py
2 Test and Evaluate with got10k-toolkit
python ./bin/my_test.py
3 Evaluate
python ./bin/my_eval.py
2 SiamFC-GOT
1 Train
python ./bin/my_train.py
2 Test and Evaluate with got10k-toolkit
python ./bin/my_test.py
3 Batch Test
./bin/cmd_test.sh
4 Evaluate
python ./bin/my_eval.py
5 Hyperparameter
python ./bin/hp_search.py
SiamRPN
1 SiamRPN (YTB&VID)
1 Generate training set
python ./bin/create_dataset_ytbid.py
2 Generate Lmdb file
python ./bin/create_lmdb.py
3 Train
python ./bin/my_train.py
4 Test and Evaluate
python ./bin/my_test.py
2 SiamRPN-GOT
1 Train
python ./bin/my_train.py
2 Test and Evaluate
python ./bin/my_test.py
3 Batch Test
./bin/cmd_test.sh
4 Hyperparameter
python ./bin/hp_search.py
5 DDP Train
./bin/cmd_dist_train.sh
3 SiamRPNpp-UP
Note that you should first build region by run the follow command:
python setup.py build_ext —-inplace
1 Train
python ./bin/my_train.py
2 Test
python ./bin/my_test.py
3 Batch Test
./bin/cmd_test.sh
4 Batch Evaluate
./bin/cmd_eval.sh
5 Demo
python ./bin/my_demo.py
6 Hyperparameter
python ./bin/hp_search.py
7 DDP Train
./bin/cmd_dist_train.sh
DaSiamRPN
1 DaSiamRPN
1 Test
python ./bin/my_test.py
2 DaSiamRPN-GOT
1 Train
python ./bin/my_train.py
2 Test
python ./bin/my_test.py
3 Batch Test
./bin/cmd_test.sh
4 Hyperparameter
python ./bin/hp_search.py
3 SiamRPNpp-DA
Note that you should first build region by run the follow command:
python setup.py build_ext —-inplace
1 Train
python ./bin/my_train.py
2 Test
python ./bin/my_test.py
3 Batch Test
./bin/cmd_test.sh
4 Batch Evaluate
./bin/cmd_eval.sh
5 Demo
python ./bin/my_demo.py
6 Hyperparameter
python ./bin/hp_search.py
7 DDP Train
./bin/cmd_dist_train.sh
UpdateNet
1 UpdateNet-FC
1 Generate training set
python ./updatenet/create_template.py
2 Train UpdateNet (Note thae you should change the stage value )
python ./updatenet/train_upd.py
3 Test UpdateNet (Note that you should set udpatenet path and stage value)
python ./bin/my_test.py
2 UpdateNet-DA
1 Generate training set
python ./updatenet/create_template.py
2 Train UpdateNet (Note that you should change the stage value
)
python ./updatenet/train_upd.py
3 Test UpdateNet (Note that you should set udpatenet path and stage value)
python ./bin/my_test.py
3 UpdateNet-UP
1 Generate training set
python ./updatenet/create_template.py
2 Train UpdateNet
python ./updatenet/train_upd.py
3 Test UpdateNet
python ./bin/my_test.py
4 UpdateNet-DW
Note that you should first build region by run the follow command:
python setup.py build_ext —-inplace
1 Generate training set
python ./updatenet/create_template.py
2 Train UpdateNet
Note that you should change the stage value
python ./updatenet/train_upd.py
3 Test UpdateNet (Note that you should set udpatenet path and stage value)
python ./bin/my_test.py
SiamDW
1 SiamDW-FC
1 Train
python ./bin/my_train.py
2 Teat and Evaluate
python ./bin/my_test.py
3 Batch Test
python ./bin/cmd_test.py
3 Hyperparameters
python ./bin/hp_search.py
2 SiamDW-UP
1 Train
python ./bin/my_train.py
2 Teat and Evaluate
python ./bin/my_test.py
3 Batch Test
python ./bin/cmd_test.py
3 Hyperparameters
python ./bin/hp_search.py
SiamRPNpp
1 SiamRPNpp-DW
Note that you should first build region by run the follow command:
python setup.py build_ext —-inplace
1 Train
python ./bin/my_train.py
2 Test
python ./bin/my_test.py
3 Batch Test
./bin/cmd_test.sh
4 Batch Evaluate
./bin/cmd_eval.sh
5 Demo
python ./bin/my_demo.py
6 Hyperparameter
python ./bin/hp_search.py
7 DDP Train
./bin/cmd_dist_train.sh
2 SiamRPNpp-ResNet
SiamFCpp
1 SiamFCpp-GOT
1 first, you should run compile.sh
sh ./compile.sh
2 Train
python ./bin/my_train.py
3 Test
python ./bin/my_test.py
Experiments
OTB 2015
"success_score": 0.6289266117015362,
"precision_score": 0.830571693318284,
"success_rate": 0.7891486658050533,
"speed_fps": 84.38537836344958
official
"success_score": 0.6797143434600249,
"precision_score": 0.8841645010368359,
"success_rate": 0.8551268591684209,
"speed_fps": 144.9084986738754