Skip to content
/ DNA Public

(CVPR 2020) Block-wisely Supervised Neural Architecture Search with Knowledge Distillation

Notifications You must be signed in to change notification settings

changlin31/DNA

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

26 Commits
 
 
 
 
 
 
 
 

Repository files navigation

DNA

This repository provides the code of our paper: Blockwisely Supervised Neural Architecture Search with Knowledge Distillation.

Illustration of DNA. Each cell of the supernet is trained independently to mimic the behavior of the corresponding teacher block.

Comparison of model ranking for DNA vs. DARTS, SPOS and MnasNet under two different hyper-parameters.

Our Trained Models

Usage

1. Requirements

2. Searching

The code for supernet training, evaluation and searching is under searching directory.

  • cd searching

i) Train & evaluate the block-wise supernet with knowledge distillation

  • Modify datadir in initialize/data.yaml to your ImageNet path.
  • Modify nproc_per_node in dist_train.sh to suit your GPU number. The default batch size is 64 for 8 GPUs, you can change batch size and learning rate in initialize/train_pipeline.yaml
  • By default, the supernet will be trained sequentially from stage 1 to stage 6 and evaluate after each stage. This will take about 2 days on 8 GPUs with EfficientNet B7 being the teacher. Resuming from checkpoints is supported. You can also change start_stage in initialize/train_pipeline.yaml to force start from a intermediate stage without loading checkpoint.
  • sh dist_train.sh

ii) Search for the best architecture under constraint.

Our traversal search can handle a search space with 6 ops in each layer, 6 layers in each stage, 6 stages in total. A search process like this should finish in half an hour with a single cpu. To perform search over a larger search space, you can manually divide the search space or use other search algorithms such as Evolution Algorithms to process our evaluated architecture potential files.

  • Copy the path of architecture potential files generated in step i) to potential_yaml in process_potential.py. Modify the constraint in process_potential.py.
  • python process_potential.py

iii) Searching with multiple cells in each block.

Please refer to the clarification from @MohanadOdema in this issue.

3. Retraining

The retraining code is simplified from the repo: pytorch-image-models and is under retraining directory.

  • cd retraining

  • Retrain our models or your searched models

    • Modify the run_example.sh: change data path and hyper-params according to your requirements
    • Add your searched model architecture to model.py. You can also use our searched and predefined DNA models.
    • sh run_example.sh
  • You can evaluate our models with the following command:
    python validate.py PATH/TO/ImageNet/validation --model DNA_a --checkpoint PATH/TO/model.pth.tar

    • PATH/TO/ImageNet/validation should be replaced by your validation data path.
    • --model : DNA_a can be replaced by DNA_b, DNA_c, DNA_d for our different models.
    • --checkpoint : Suggest the path of your downloaded checkpoint here.

About

(CVPR 2020) Block-wisely Supervised Neural Architecture Search with Knowledge Distillation

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published