NPPNet

Neural Architecture Search for Joint Human Parsing and Pose Estimation (ICCV2021)

Preparation

1. Download the LIP dataset from https://www.sysuhcp.com/lip
2. Put the prepare_file.zip to your root path of LIP, such as /home/data/LIP/prepare_file.zip, and unzip it.
3. pip install -r requirements.txt

Search

We only release the search of interaction (without encoder-decoder search) now. Searching with 4 gpus: CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 search_lip_sync.py --cfg ./experiments/lip/384_384.yaml

Train

Download the pretrained backbone from link: https://pan.baidu.com/s/1sIcNXZnM4_U6nthmJAo6NA?pwd=2050 with the extracing code: 2050, and modify the model path in the line 205 of augment_lip_sync.py. The newest model weight can be also downloaded from the link.

Training with 4 gpus: CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 augment_lip_sync.py --cfg ./experiments/lip/384_384.yaml

Q&A

If you have any questions, please concat with [email protected].

Thanks

Thanks to the public repo: mula(ECCV2018) for providing the base code.