deep-person-reid

好用的 reid 訓練框架

取得

git clone https://github.com/KaiyangZhou/deep-person-reid.git
cd deep-person-reid/

環境建置

使用 conda

conda create --name torchreid python=3.7
conda activate torchreid

# install dependencies
# make sure `which python` and `which pip` point to the correct path
pip install -r requirements.txt

# install torch and torchvision (select the proper cuda version to suit your machine)
conda install pytorch torchvision cudatoolkit=9.0 -c pytorch

# install torchreid (don't need to re-build it if you modify the source code)
python setup.py develop

使用 docker

Dockerfile

FROM pytorch/pytorch:0.4.1-cuda9-cudnn7-devel
RUN git clone https://github.com/KaiyangZhou/deep-person-reid.git
RUN cd deep-person-reid/
RUN pip install --upgrade pip
RUN apt update
RUN apt install -y libgl1-mesa-glx
RUN apt-get install -y libglib2.0-0
RUN pip install --upgrade torch
RUN pip install --upgrade tensorboard
ADD ./ ./
RUN pip install -r requirements.txt
RUN python setup.py develop

for 3090

FROM pytorch/pytorch:0.4.1-cuda9-cudnn7-devel
RUN pip install --upgrade pip
RUN apt update
RUN apt install -y libgl1-mesa-glx
RUN apt-get install -y libglib2.0-0
RUN pip install --upgrade torch
RUN pip install --upgrade tensorboard
RUN pip uninstall -y torch torchvision
RUN pip install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu111/torch_nightly.html
ADD ./ ./
RUN pip install -r requirements.txt
RUN python setup.py develop

build

docker build -t torchreid .

run

docker run --gpus all -d \
-it \
-e DISPLAY \
-e QT_X11_NO_MITSHM=1 \
--shm-size="8g" \
-v /tmp/.X11-unix:/tmp/.X11-unix \
-v $HOME/.Xauthority:/root/.Xauthority \
--device=/dev/video0:/dev/video0 \
--name torchreid \
--restart=always \
torchreid bash

下載訓練資料(若 visualize_actmap 無法下載）

mkdir datasets
cd datasets
kaggle datasets download pengcw1/market-1501
unzip market-1501.zip 
rm market-1501.zip 

測試熱區圖

python tools/visualize_actmap.py --root ./datasets/ -d market1501 -m osnet_x1_0 --weights PATH_TO_PRETRAINED_WEIGHTS --save-dir log/visactmap_osnet_x1_0_market1501

或使用 vscode

{
    // 使用 IntelliSense 以得知可用的屬性。
    // 暫留以檢視現有屬性的描述。
    // 如需詳細資訊，請瀏覽: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Visualize activation maps",
            "type": "python",
            "request": "launch",
            "program": "tools/visualize_actmap.py",
            "console": "integratedTerminal",
            "args": [
                "--root",
                "./datasets/",
                "-d",
                "market1501",
                "-m",
                "osnet_x1_0",
                "--weights",
                "PATH_TO_PRETRAINED_WEIGHTS",
                "--save-dir",
                "log/visactmap_osnet_x1_0_market1501"
            ]
        }
    ]
}

訓練

import torchreid
trainmodel = "resnet50"
batch_size = 800

# trainmodel = "osnet_x0_25"
# batch_size = 32

datamanager = torchreid.data.ImageDataManager(
    root='reid-data',
    sources='market1501',
    targets='market1501',
    height=256,
    width=128,
    batch_size_train=batch_size,
    batch_size_test=batch_size,
    transforms=['random_flip', 'random_crop']
)

model = torchreid.models.build_model(
    name=trainmodel,
    num_classes=datamanager.num_train_pids,
    loss='softmax',
    pretrained=True
)

model = model.cuda()

optimizer = torchreid.optim.build_optimizer(
    model,
    optim='adam',
    lr=0.0003
)

scheduler = torchreid.optim.build_lr_scheduler(
    optimizer,
    lr_scheduler='single_step',
    stepsize=20
)
engine = torchreid.engine.ImageSoftmaxEngine(
    datamanager,
    model,
    optimizer=optimizer,
    scheduler=scheduler,
    label_smooth=True
)

engine.run(
    save_dir='log/'+trainmodel,
    max_epoch=60,
    eval_freq=10,
    print_freq=10,
    test_only=False
)