Dubhe/dubhe-tadl/cream/README.md

• Cream of the Crop: Distilling Prioritized Paths For One-Shot Neural Architecture Search
• • 0x01 requirements
  • 0x02 Quick Start

Cream of the Crop: Distilling Prioritized Paths For One-Shot Neural Architecture Search

0x01 requirements

• Install the following requirements:

future
thop
timm<0.4
yacs
ptflops==0.6.4
#tensorboardx
#tensorboard
#opencv-python
#torch-scope
#git+https://github.com/sovrasov/flops-counter.pytorch.git
#git+https://github.com/Tramac/torchscope.git

• (required) Build and install apex to accelerate the training
  (see yuque),
  a little bit faster than pytorch DistributedDataParallel.

• Put the imagenet data in ./data Using the following script:

cd TADL_DIR/pytorch/cream/
ln -s /mnt/data .

0x02 Quick Start

• Run the following script to search an architecture.

python trainer.py

• Selector (deprecated)

python selector.py

• Train searched architectures.

Note: exponential moving average(model_ema) is not available yet.

python retrainer.py