TRAIN.md

PySOT Training Tutorial

This implements training of SiamRPN with backbone architectures, such as ResNet, AlexNet.

Add PySOT to your PYTHONPATH

export PYTHONPATH=/path/to/pysot:$PYTHONPATH

Prepare training dataset

Prepare training dataset, detailed preparations are listed in training_dataset directory.

• VID
• YOUTUBEBB
• DET
• COCO

Download pretrained backbones

Download pretrained backbones from Google Drive and put them in pretrained_models directory

Training

To train a model (SiamRPN++), run train.py with the desired configs:

cd experiments/siamrpn_r50_l234_dwxcorr_8gpu

Multi-processing Distributed Data Parallel Training

Single node, multiple GPUs:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
python -m torch.distributed.launch \
    --nproc_per_node=8 \
    --master_port=2333 \
    ../../tools/train.py --cfg config.yaml

Multiple nodes:

Node 1: (IP: 192.168.1.1, and has a free port: 2333) master node

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
python -m torch.distributed.launch \
    --nnodes=2 \
    --node_rank=0 \
    --nproc_per_node=8 \
    --master_addr=192.168.1.1 \  # adjust your ip here
    --master_port=2333 \
    ../../tools/train.py

Node 2:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
python -m torch.distributed.launch \
    --nnodes=2 \
    --node_rank=1 \
    --nproc_per_node=8 \
    --master_addr=192.168.1.1 \
    --master_port=2333 \
    ../../tools/train.py

Testing

After training, you can test snapshots on VOT dataset. For AlexNet, you need to test snapshots from 35 to 50 epoch. For ResNet, you need to test snapshots from 10 to 20 epoch.

START=10
END=20
seq $START 1 $END | \
    xargs -I {} echo "snapshot/checkpoint_e{}.pth" | \
    xargs -I {} \ 
    python -u ../../tools/test.py \
        --snapshot {} \
	--config config.yaml \
	--dataset VOT2018 2>&1 | tee logs/test_dataset.log

Evaluation

python ../../tools/eval.py 	 \
	--tracker_path ./results \ # result path
	--dataset VOT2018        \ # dataset name
	--num 4 		 \ # number thread to eval
	--tracker_prefix 'ch*'   # tracker_name