This folder includes Torch code for training of the single person pose estimation network.
To run this code, the following must be installed:
- Torch7
- hdf5 (and the torch-hdf5 package)
- cudnn
We finetune our model based on the pre-trained pyraNet model.
# First link the images dir
cd $SPPE_ROOT/train/src
ln -s /data/MPII/images ../data/mpii/images
ln -s /data/COCO/images ../data/coco/images
# Then finetune the model using PGPG
th main.lua -expID finetune -usePGPG -loadModel $MODEL_DIR -LR 0.5e-4 -nEpochs 50
# Or start a new training with multi-GPU
th main.lua -expID coco -nGPU 8 -trainBatch 16 -validBatch 16
For more options, you can read the src/opt.lua
file.