SparseTrack模型使用指导

docker pull registry-cbu.huawei.com/atelier/pytorch_2_1_ascend:pytorch_2.1.0-cann_8.0.rc3-py_3.9-hce_2.0.2409-aarch64-snt9b-20241213131522-aafe527

docker run -itd \
   -u root \
   --privileged \
   --device=/dev/davinci0 \
   --device=/dev/davinci1 \
   --device=/dev/davinci3 \
   --device=/dev/davinci2 \
   --device=/dev/davinci4 \
   --device=/dev/davinci5 \
   --device=/dev/davinci6 \
   --device=/dev/davinci7 \
   --device=/dev/davinci_manager \
   --device=/dev/devmm_svm \
   --device=/dev/hisi_hdc \
   -v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \
   -v /usr/local/dcmi:/usr/local/dcmi \
   -v /etc/ascend_install.info:/etc/ascend_install.info \
   -v /sys/fs/cgroup:/sys/fs/cgroup:ro \
   -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
   -v /usr/bin/hccn_tool:/usr/bin/hccn_tool \
   -v /etc/hccn.conf:/etc/hccn.conf \
   --shm-size 1024g \
   --network=host\
   --name SparseTrack \
   1813d558bbe2\
   /bin/bash

docker exec -it SparseTrack  bash
conda create -n SparseTrack --clone PyTorch-2.1.0
conda activate SparseTrack

pip install 'git+https://github.com/facebookresearch/detectron2.git'

cd /home/HwHiAiUser
git clone https://github.com/hustvl/SparseTrack.git

cd SparseTrack
mkdir pretrain
cd pretrain
wget https://github.com/Megvii-BaseDetection/YOLOX/releases/download/0.1.1rc0/yolox_x.pth

cd /data/datasets
wget https://motchallenge.net/data/MOT17.zip
unzip MOT17.zip

cd /home/HwHiAiUser/SparseTrack
python tools/convert_mot17_to_coco.py

# 修改mot17_train_config.py
data_dir=/data/datasets/MOT17
init_checkpoint=/home/HwHiAiUser/SparseTrack/pretrain/yolox_x.pth

# 修改register_data.py
VAL_JSON="/data/datasets/MOT17/annotations/val_half.json"
VAL_PATH="/data/datasets/MOT17/train"

# 修改train.py，开头插入自适配代码
import torch_npu
from torch_npu.contrib import transfer_to_npu

vi /home/ma-user/anaconda3/envs/SparseTrack/lib/python3.9/site-packages/detectron2/engine/launch.py

dist.init_process_group(
    backend="HCCL" if has_gpu else "GLOO",   #NCCL修改为HCCL
    init_method=dist_url,
    world_size=world_size,
    rank=global_rank,
    timeout=timeout,
)

vi /home/ma-user/anaconda3/envs/SparseTrack/lib/python3.9/site-packages/detectron2/utils/collect_env.py

# 修改前
for k in range(torch.cuda.device_count()):
    cap = ".".join((str(x) for x in torch.cuda.get_device_capability(k)))
# 修改后
for k in range(torch.cuda.device_count()):
    cap_info = torch.cuda.get_device_capability(k)
    if cap_info is not None:
        cap = ".".join((str(x) for x in cap_info))
    else:
        cap = "NPU Device"
                

pip uninstall datasets

vi /home/ma-user/anaconda3/envs/SparseTrack/lib/python3.9/site-packages/detectron2/engine/train_loop.py

if self.zero_grad_before_forward:
    self.optimizer.zero_grad()
with autocast(dtype=torch.bfloat16):   #原来的dtype=self.precision修改为dtype=torch.bfloat16
    loss_dict = self.model(data)

ASCEND_RT_VISIBLE_DEVICES=2,3 python train.py --num-gpus 2 --config-file mot17_train_config.py

# 安装gperftools
wget https://github.com/gperftools/gperftools/releases/download/gperftools-2.16/gperftools-2.16.tar.gz
tar -xf gperftools-2.16.tar.gz && cd gperftools-2.16
./configure --prefix=/usr/local/lib --with-tcmalloc-pagesize=64
make
make install
# 找到包位置
find /usr -name libtcmalloc.so*
# 添加环境变量，开启高性能库
export LD_PRELOAD="$LD_PRELOAD:/usr/local/lib/lib/libtcmalloc.so"

# 开启二级流水
export TASK_QUEUE_ENABLE=2
# 绑核
export CPU_AFFINITY_CONF=1
# 启用hccl ffts+模式
export ASCEND_ENHANCE_enable=1

vi utils/get_optimizer.py

 # 修改前
 optimizer = torch.optim.SGD(   
 # 修改后
 optimizer = torch_npu.optim.NpuFusedSGD(

vi train.py

 # 训练脚本入口添加
torch_npu.npu.set_compile_mode(jit_compile=False)

ASCEND_RT_VISIBLE_DEVICES=2,3 python train.py --num-gpus 2 --config-file mot17_train_config.py