Skip to content

Latest commit

 

History

History

cnn

README.md

Searching CNN architectures with SGAS

We run 10 independent searches to get 10 architectures with Criterion 1 or Criterion 2 on on CIFAR-10. We evaluate the discovered architectures on CIFAR-10 and report the mean and standard deviation of the test accuracy across those 10 models and the performance of the best model.

We choose the 3 best performing cell architectures on CIFAR-10 for each Criterion and train them on ImageNet.

CIFAR-10 will be downloaded automatically. To obtain ImageNet dataset, please follow the instructions here.

Search on CIFAR-10

The search takes about 0.25 day (6 hours) on a single NVIDIA GTX 1080Ti. To search with SGAS Cri2, run:

python train_search.py --use_history

To search with SGAS Cri1, run:

python train_search.py

Train on CIFAR-10

To train the best architectures (Cri1_CIFAR_Best, Cri2_CIFAR_Best) from scratch, run:

python train.py --auxiliary --cutout --arch Cri1_CIFAR_Best

or

python train.py --auxiliary --cutout --arch Cri2_CIFAR_Best

Train on ImageNet

To train the best architectures (Cri1_ImageNet_Best, Cri1_ImageNet_Best) from scratch, run:

python train_imagenet.py --auxiliary --arch Cri1_ImageNet_Best --batch_size 1024 --learning_rate 0.5

or

python train_imagenet.py --auxiliary --arch Cri2_ImageNet_Best --batch_size 1024 --learning_rate 0.5
  • We run these experiments on 8 Nvidia Tesla V100 GPUs for three days by setting --batch_size 1024 and --learning_rate 0.5.

Set --arch to any architecture you want. (One can find more architectures from genotyps.py).

Pretrained models

Our pretrained models can be found from Google Cloud.

Test pretrained models on CIFAR-10

Use the parameter --model_path to set a specific pretrained model to load. For example, to test the best architecture Cri1_CIFAR_Best or Cri2_CIFAR_Best, run:

python test.py --auxiliary --arch Cri1_CIFAR_Best --model_path Cri1_CIFAR_Best.pt
  • Expected result: 2.39% test error with 3.8M model params.

or

python test.py --auxiliary --arch Cri2_CIFAR_Best --model_path Cri2_CIFAR_Best.pt
  • Expected result: 2.44% test error with 4.1M model params.

Test pretrained models on ImageNet

To test the best architecture Cri1_ImageNet_Best or Cri2_ImageNet_Best, run:

python test_imagenet.py --auxiliary --arch Cri1_ImageNet_Best --model_path Cri1_ImageNet_Best.pt
  • Expected result: 24.2% top1 test error with 5.3M model params.

or

python test_imagenet.py --auxiliary --arch Cri2_ImageNet_Best --model_path Cri2_ImageNet_Best.pt
  • Expected result: 24.1% top1 test error with 5.4M model params.

Ablation

In our ablation study on hyper-parameters, we find setting --decision_freq as 7 yield more stable results. Run:

python train_search.py --use_history --decision_freq 7