#install apex intsll apex in ./

https://blog.csdn.net/jacke121/article/details/103252605

#Pipeline

0 train supernet

if the gpu memory >15G: use single gpu is nenough

  run      CUDA_VISIBLE_DEVICES=2,3  python train_autodeeplab.py --dataset 2d --no_val --filter_multiplier 2

##1 decode supernet The check point stored in ./run/2d/deeplab-resnet/experiment_xx（the last epoch）

CUDA_VISIBLE_DEVICES=0 python decode_autodeeplab.py --dataset 2d --resume ./run/2d/deeplab-resnet/experiment_16/checkpoint.pth.tar

2 retrain arch

CUDA_VISIBLE_DEVICES=0 python train.py --net_arch ./run/2d/deeplab-resnet/experiment_xx/network_path.npy  
--cell_arch ./run/2d/deeplab-resnet/experiment_xx/genotype.npy  --filter_multiplier 2 --dataset 2d

##derive model ##retrain model