Code for pet biometrics challenge
Linux or macOS with python ≥ 3.6
PyTorch ≥ 1.6
Cython (optional to compile evaluation code)
tensorboard (needed for visualization):
pip install tensorboard
gdown (for automatically downloading pre-train model)
pip install faiss-gpu
conda create -n fastreid python=3.8 conda activate fastreid conda install pytorch==1.7.1 torchvision tensorboard -c pytorch pip install -r docs/requirements.txt
We use GPU 3090 for training and testing. The cuda version is 11.1, torch version is 1.7.1, the python version is 3.8.8.
Download the competition datasets pet_biometric_challenge_2022, and then unzip them under the directory like
datasets ├── pet_biometric_challenge_2022 │ └── train │ └── validation │ └── test
Prepare Pre-trained Models
We have trained 8 models, and you can download the pre-trained model form this link: 链接：https://pan.baidu.com/s/1Z3PZLIer8S7U_NdCihgm9A 提取码：bauk. Then you should save it under the path of logs. The file tree should be like as:
logs └── resnet101 └── config.yaml └── model_final.pth
You can get the final submit.csv by runing:
It will generate submit.csv in the root dir, which is the final ensemble result. The test process takes approximately 20mins.
We train our model through three stage. Stage1 train the original dataset with 224 resolution by different losses , backbone and batchsize. Stage2 finetune the trainset with 384 resolution which is inspired by kaggle-landmark-2021-1st-place. Stage3 finetune the model with trainset and validation set which is assigned with pseudo labels. The training process takes approximately 64 hours.