Code for paper:

CATE: Computation-aware Neural Architecture Encoding with Transformers
Shen Yan, Kaiqiang Song, Fei Liu, Mi Zhang.
ICML 2021 (Long Talk).

Overview of CATE: It takes computationally similar architecture pairs as the input and trained to predict masked operators given the pairwise computation information. Apart from the cross-attention blocks, the pretrained Transformer encoder is used to extract architecture encodings for the downstream search.

The repository is built upon pybnn and nas-encodings.

conda create -n tf python=3.7
source activate tf
cat requirements.txt | xargs -n 1 -L 1 pip install

Install nasbench and download nasbench_only108.tfrecord in ./data folder.

python preprocessing/gen_json.py

Data will be saved in ./data/nasbench101.json.

python preprocessing/data_generate.py --dataset nasbench101 --flag extract_seq
python preprocessing/data_generate.py --dataset nasbench101 --flag build_pair --k 2 --d 2000000 --metric params

The corresponding training data and pairs will be saved in ./data/nasbench101/.

Alternatively, you can download the data train_data.pt, test_data.pt and pair indices train_pair_k2_d2000000_metric_params.pt, test_pair_k2_d2000000_metric_params.pt from here.

bash run_scripts/pretrain_nasbench101.sh

The pretrained models will be saved in ./model/.

Alternatively, you can download the pretrained model nasbench101_model_best.pth from here.

python inference/inference.py --pretrained_path model/nasbench101_model_best.pth.tar --train_data data/nasbench101/train_data.pt --valid_data data/nasbench101/test_data.pt --dataset nasbench101

The extracted embeddings will be saved in ./cate_nasbench101.pt.

Alternatively, you can download the pretrained embeddings cate_nasbench101.pt from here.

bash run_scripts/run_search_nasbench101.sh

Search results will be saved in ./nasbench101/.

Install nasbench301 and download the xgb_v1.0 and lgb_runtime_v1.0 file. You may need to make pytorch_geometric compatible with Pytorch and CUDA version.

python preprocessing/gen_json_darts.py # randomly sample 1,000,000 archs

Data will be saved in ./data/nasbench301_proxy.json.

Alternatively, you can download the json file nasbench301_proxy.json from here.

python preprocessing/data_generate.py --dataset nasbench301 --flag extract_seq
python preprocessing/data_generate.py --dataset nasbench301 --flag build_pair --k 1 --d 5000000 --metric flops

The correspoding training data and pairs will be saved in ./data/nasbench301/.

Alternatively, you can download the data train_data.pt, test_data.pt and pair indices train_pair_k1_d5000000_metric_flops.pt, test_pair_k1_d5000000_metric_flops.pt from here.

bash run_scripts/pretrain_nasbench301.sh

The pretrained models will be saved in ./model/.

Alternatively, you can download the pretrained model nasbench301_model_best.pth from here.

python inference/inference.py --pretrained_path model/nasbench301_model_best.pth.tar --train_data data/nasbench301/train_data.pt --valid_data data/nasbench301/test_data.pt --dataset nasbench301 --n_vocab 11

The extracted encodings will be saved in ./cate_nasbench301.pt.

Alternatively, you can download the pretrained embeddings cate_nasbench301.pt from here.

bash run_scripts/run_search_nasbench301.sh

Search results will be saved in ./nasbench301/.

Download the pretrained embeddings cate_darts.pt from here.

python search_methods/dngo_ls_darts.py --dim 64 --init_size 16 --topk 5 --dataset darts --output_path bo  --embedding_path cate_darts.pt

Search log will be saved in ./darts/. Final search result will be saved in ./darts/bo/dim64.

python darts/cnn/train.py --auxiliary --cutout --arch cate_small
python darts/cnn/train.py --auxiliary --cutout --arch cate_large

• Expected results (CATE-Small): 2.55% avg. test error with 3.5M model params.
• Expected results (CATE-Large): 2.46% avg. test error with 4.1M model params.

python darts/cnn/train_imagenet.py  --arch cate_small --seed 1 
python darts/cnn/train_imagenet.py  --arch cate_large --seed 1

• Expected results (CATE-Small): 26.05% test error with 5.0M model params and 556M mult-adds.
• Expected results (CATE-Large): 25.01% test error with 5.8M model params and 642M mult-adds.

python darts/cnn/visualize.py cate_small
python darts/cnn/visualize.py cate_large

bash run_scripts/generate_oo.sh

Data will be saved in ./data/nasbench101_oo_train.json and ./data/nasbench101_oo_test.json.

python preprocessing/data_generate_oo.py --flag extract_seq
python preprocessing/data_generate_oo.py --flag build_pair

The corresponding training data and pair indices will be saved in ./data/nasbench101/.

python run.py --do_train --parallel --train_data data/nasbench101/nasbench101_oo_trainSet_train.pt --train_pair data/nasbench101/oo_train_pairs_k2_params_dist2e6.pt  --valid_data data/nasbench101/nasbench101_oo_trainSet_validation.pt --valid_pair data/nasbench101/oo_validation_pairs_k2_params_dist2e6.pt --dataset oo

The pretrained models will be saved in ./model/.

# Adjacency encoding
python inference/inference_adj.py
# CATE encoding
python inference/inference.py --pretrained_path model/oo_model_best.pth.tar --train_data data/nasbench101/nasbench101_oo_testSet_split1.pt --valid_data data/nasbench101/nasbench101_oo_testSet_split2.pt --dataset oo_nasbench101

The extracted encodings will be saved as ./adj_oo_nasbench101.pt and ./cate_oo_nasbench101.pt.

Alternatively, you can download the data, pair indices, pretrained models, and extracted embeddings from here.

for s in {1..500}; do python search_methods/oo_mlp.py --dim 27 --seed $s --init_size 16 --topk 5 --dataset oo_nasbench101 --output_path np_adj  --embedding_path adj_oo_nasbench101.pt; done
for s in {1..500}; do python search_methods/oo_mlp.py --dim 64 --seed $s --init_size 16 --topk 5 --dataset oo_nasbench101 --output_path np_cate  --embedding_path cate_oo_nasbench101.pt; done

Search results will be saved in ./oo_nasbench101.

If you find this useful for your work, please consider citing:

@InProceedings{yan2021cate,
  title = {CATE: Computation-aware Neural Architecture Encoding with Transformers},
  author = {Yan, Shen and Song, Kaiqiang and Liu, Fei and Zhang, Mi},
  booktitle = {ICML},
  year = {2021}
}