Skip to content

Latest commit

 

History

History
57 lines (36 loc) · 1.48 KB

README.md

File metadata and controls

57 lines (36 loc) · 1.48 KB
Raw

Gradient Based NAS

Blocks

  • weighted_sum : Like fbnet, output of block is the weighted sum of units in the block.
  • sample : Like proxyless, output of block is one unit sampled from all the units in the block.
  • dag : Like DARTS, treat block as a DAG, every edge's output is the weighted sum of all the operations specified.

Head

  • classification : For classification task, output ce, acc.
  • detection : For detection task, highly depend on mmdetection and mmcv.
  • regress(TODO) : For regression task, e.g. landmark
  • segmentation(TODO) : For segmentation task

Models

Assembly class, stack base, block, head.

  • darts : snas, but no shared architecture parameters.
  • fbnet_faster_rcnn : fbnet + faster rcnn
  • proxyless : proxyless, but use reward = - (ce + time cost)

Search

Do search, train model parameters and architecture parameters iteratively.

Demos

Assume you are under direcotry ${nas_directory}/, you can run some demo with following scipts. NOTE: You may need to modify path for dataset and log in ${nas_directory}/nas/demo/*.py.

  • proxyless
python -m nas.demo.proxyless_nas_cifar10.py --log-frequence 50 --gpus 0,1 --batch-size 128
  • fbnet
python -m nas.demo.fbnet_cus_mxnet_rec.py --gpus 0,1,2,3,4,5,6,7 --log-frequence 100 --batch-size 192
  • snas
python -m nas.demo.darts_cifar10.py --gpus 0,1 --log-frequence 50 --batch-size 32
  • detection
python -m nas.demo.fbnet_faster_rcnn.py --gpus 0,1,2,3