TRAINING.md

Training Recipe

Get datasets

MPII Human Pose Dataset

• Download the dataset and create a link to the images directory at ./data/mpii/ images

  ln -s ${MPII_PATH}/images ./data/mpii/images
  

• Download the annotation file in our JSON format, and save it to ./data/mpii/mpii_annotations.json

• You are good to go!

COCO Keypoints 2014/2017

• Download datasets:

  cd ./data/mscoco
  wget http://images.cocodataset.org/zips/train2014.zip
  wget http://images.cocodataset.org/zips/val2014.zip
  wget http://images.cocodataset.org/zips/train2017.zip
  wget http://images.cocodataset.org/zips/val2017.zip
  unzip train2014.zip -d images
  unzip train2014.zip -d images
  unzip train2014.zip -d images
  unzip train2014.zip -d images
  rm -rf *.zip
  

  • Download the coco_annotations_2014.json and coco_annotations_2017.json in our JSON format, and save it to ./data/mscoco

• You are good to go!

Leeds Sports Pose (LSP)

• Download datasets:

  mkdir -p ./data/lsp/images
  cd ./data/lsp/images
  wget http://sam.johnson.io/research/lsp_dataset.zip
  wget http://sam.johnson.io/research/lspet_dataset.zip
  unzip lsp_dataset.zip -d lsp_dataset
  unzip lspet_dataset.zip -d lspet_dataset
  

  • Download the LEEDS_annotations.json in our JSON format, and save it to ./data/lsp

• You are good to go!

Training

• Example 1: Train from scratch - ECCV'16 8-stack hourglass network

CUDA_VISIBLE_DEVICES=0 python ./example/main.py \
--dataset mpii \
--arch hg \
--stack 8 \
--block 1 \
--features 256 \
--checkpoint ./checkpoint/mpii/hg-s8-b1

• Example 2: Train a much faster version of HG (e.g., 1-stack)

CUDA_VISIBLE_DEVICES=0 python ./example/main.py \
--dataset mpii \
--arch hg \
--stack 1 \
--block 1 \
--features 256 \
--checkpoint ./checkpoint/mpii/hg-s1-b1

• Example 3: Train on COCO 2014/2017 (set --year argument )

CUDA_VISIBLE_DEVICES=0 python ./example/main.py \
--dataset mscoco \
--year 2017 \
--arch hg \
--stack 1 \
--block 1 \
--features 256 \
--checkpoint ./checkpoint/mscoco/hg-s1-b1

• Example 4: resume training from a checkpoint

CUDA_VISIBLE_DEVICES=0 python ./example/main.py \
--dataset mpii \
--arch hg \
--stack 8 \
--block 1 \
--features 256 \
--checkpoint ./checkpoint/mpii/hg-s8-b1 \
--resume ./checkpoint/mpii/hg-s8-b1/checkpoint.pth.tar

• Example 5: Simple baselines model (e.g., resnet50)

GPU=0
ARCH=pose_resnet
DSET=mpii
FEAT=256
TRAINB=32
VALB=24
LR=1e-3
WORKERS=8
RESNETLAYERS=50
SOLVER=adam

CHECKPOINT=./checkpoint/${DSET}/${ARCH}-${RESNETLAYERS}-${SOLVER}-lr${LR}

CUDA_VISIBLE_DEVICES=${GPU} python example/main.py \
--arch ${ARCH} \
--dataset ${DSET} \
--stack ${STACK} \
--block ${BLOCK} \
--features ${FEAT} \
--checkpoint ${CHECKPOINT} \
--train-batch ${TRAINB} \
--test-batch ${VALB} \
--lr ${LR} \
--workers ${WORKERS} \
--resnet-layers ${RESNETLAYERS} \
--solver ${SOLVER} \
--epochs 140 \
--schedule 90 120 \
--target-weight # original code uses this trick for training pose_resnet

• Evaluation from an existing model: use -e

CUDA_VISIBLE_DEVICES=0 python ./example/main.py \
--dataset mpii \
--arch hg \
--stack 8 \
--block 1 \
--features 256 \
--checkpoint ./checkpoint/mpii/hg-s8-b1 \
--resume ./checkpoint/mpii/hg-s8-b1/checkpoint.pth.tar \
-e

• Debug: Use -d if you want to visualize the keypoints with images

CUDA_VISIBLE_DEVICES=0 python ./example/main.py \
--dataset mpii \
--arch hg \
--stack 1 \
--block 1 \
--features 256 \
--checkpoint ./checkpoint/mpii/hg-s1-b1 \
--resume ./checkpoint/mpii/hg-s1-b1/checkpoint.pth.tar \
-e \
-d

The visualized images should be like