HR-NAS: Searching Efficient High-Resolution Neural Architectures with Lightweight Transformers (CVPR21 Oral)

Environment

Require Python3, CUDA>=10.1, and torch>=1.4, all dependencies are as follows:

pip3 install torch==1.4.0 torchvision==0.5.0 opencv-python tqdm tensorboard lmdb pyyaml packaging Pillow==6.2.2 matplotlib yacs pyarrow==0.17.1
pip3 install cityscapesscripts  # for Cityscapes segmentation
pip3 install mmcv-full==latest+torch1.4.0+cu101 -f https://openmmlab.oss-accelerate.aliyuncs.com/mmcv/dist/index.html  # for Segmentation data loader
pip3 install pycocotools shapely==1.6.4 Cython pandas pyyaml json_tricks scikit-image  # for COCO keypoint estimation

or pip3 install requirements.txt

Setup

Optionally configure NCCL before running:

export NCCL_IB_DISABLE=1 
export NCCL_IB_HCA=mlx5_0 
export NCCL_IB_GID_INDEX=3 
export NCCL_SOCKET_IFNAME=eth0 
export HOROVOD_MPI_THREADS_DISABLE=1
export OMP_NUM_THREADS=56
export KMP_AFFINITY=granularity=fine,compact,1,0

Set the following ENV variable:

$MASTER_ADDR: IP address of the node 0 (Not required if you have only one node (machine))
$MASTER_PORT: Port used for initializing distributed environment
$NODE_RANK: Index of the node
$N_NODES: Number of nodes 
$NPROC_PER_NODE: Number of GPUs (NOTE: should exactly match local GPU numbers with `CUDA_VISIBLE_DEVICES`)

Example1 (One machine with 8 GPUs):

Node 1: 
>>> python -m torch.distributed.launch --nproc_per_node=8
--nnodes=1 --node_rank=0 --master_port=1234 train.py

Example2 (Two machines, each has 8 GPUs):

Node 1: (IP: 192.168.1.1, and has a free port: 1234)
>>> python -m torch.distributed.launch --nproc_per_node=8
--nnodes=2 --node_rank=0 --master_addr="192.168.1.1"
--master_port=1234 train.py

Node 2:
>>> python -m torch.distributed.launch --nproc_per_node=8
--nnodes=2 --node_rank=1 --master_addr="192.168.1.1"
--master_port=1234 train.py

Datasets

1. ImageNet

   • Prepare ImageNet data following pytorch example.
   • Optional: Generate lmdb dataset by utils/lmdb_dataset.py. If not, please overwrite dataset:imagenet1k_lmdb in yaml to dataset:imagenet1k.
   • The directory structure of $DATA_ROOT should look like this:

   ${DATA_ROOT}
   ├── imagenet
   └── imagenet_lmdb
   

   • Link the data:

   ln -s YOUR_LMDB_DIR data/imagenet_lmdb

2. Cityscapes

   • Download data from Cityscapes.
   • unzip gtFine_trainvaltest.zip leftImg8bit_trainvaltest.zip
   • Link the data:

   ln -s YOUR_DATA_DIR data/cityscapes

   • preprocess the data:

   python3 tools/convert_cityscapes.py data/cityscapes --nproc 8

3. ADE20K

   • Download data from ADE20K.
   • unzip ADEChallengeData2016.zip
   • Link the data:

   ln -s YOUR_DATA_DIR data/ade20k

4. COCO keypoints

   • Download data from COCO.
   • build tools

   git clone https://github.com/cocodataset/cocoapi.git
   cd cocoapi/PythonAPI
   python3 setup.py build_ext --inplace
   python3 setup.py build_ext install
   make  # for nms

   • Unzip and Link the data:

   ln -s YOUR_DATA_DIR data/coco

   • We also provide person detection result of COCO val2017 and test-dev2017 to reproduce our multi-person pose estimation results. Please download from OneDrive or GoogleDrive.
   • Download and extract them under data/coco/person_detection_results, and make them look like this:

   ${POSE_ROOT}
   |-- data
   `-- |-- coco
       `-- |-- annotations
           |   |-- person_keypoints_train2017.json
           |   `-- person_keypoints_val2017.json
           |-- person_detection_results
           |   |-- COCO_val2017_detections_AP_H_56_person.json
           |   |-- COCO_test-dev2017_detections_AP_H_609_person.json
           `-- images
               |-- train2017
               |   |-- 000000000009.jpg
               |   |-- 000000000025.jpg
               |   |-- 000000000030.jpg
               |   |-- ... 
               `-- val2017
                   |-- 000000000139.jpg
                   |-- 000000000285.jpg
                   |-- 000000000632.jpg
                   |-- ... 
   

Running (train & evaluation)

• Search for NAS models.

  python3 -m torch.distributed.launch --nproc_per_node=${NPROC_PER_NODE} --nnodes=${N_NODES} \
      --node_rank=${NODE_RANK} --master_addr=${MASTER_ADDR} --master_port=${MASTER_PORT} \
      --use_env train.py app:configs/YOUR_TASK.yml

  Supported tasks:

  • cls_imagenet
  • seg_cityscapes
  • seg_ade20k
  • keypoint_coco

  The super network is constructed using model_kwparams in YOUR_TASK.yml.
  To enable the searching of Transformer, set prune_params.use_transformer=True in YOUR_TASK.yml, the token numbers of each Transformer will be printed during training.
  The searched architecture can be found in best_model.json in the output dir.

• Retrain the searched models.

  • For retraining the searched classification model, please use best_model.json to overwrite the checkpoint.json in root dir of this project.

  • Modify models/hrnet.py to set checkpoint_kwparams = json.load(open('checkpoint.json')) and class InvertedResidual(InvertedResidualChannelsFused)

  • Retrain the model.

  python3 -m torch.distributed.launch --nproc_per_node=${NPROC_PER_NODE} --nnodes=${N_NODES} \
      --node_rank=${NODE_RANK} --master_addr=${MASTER_ADDR} --master_port=${MASTER_PORT} \
      --use_env train.py app:configs/cls_retrain.yml

Miscellaneous

1. Plot keypoint detection results.

   python3 tools/plot_coco.py --prediction output/results/keypoints_val2017_results_0.json --save-path output/vis/

2. About YAML config

• The codebase is a general ImageNet training framework using yaml config with several extension under apps dir, based on PyTorch.
  • YAML config with additional features
    • ${ENV} in yaml config.
    • _include for hierachy config.
    • _default key for overwriting.
    • xxx.yyy.zzz for partial overwriting.
  • --{{opt}} {{new_val}} for command line overwriting.

3. Any questions regarding HR-NAS, feel free to contact the author (mingyuding@hku.hk).
4. If you find our work useful in your research please consider citing our paper:

   @inproceedings{ding2021hrnas,
     title={HR-NAS: Searching Efficient High-Resolution Neural Architectures with Lightweight Transformers},
     author={Ding, Mingyu and Lian, Xiaochen and Yang, Linjie and Wang, Peng and Jin, Xiaojie and Lu, Zhiwu and Luo, Ping},
     booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
     year={2021}
   }