Awesome

Can GPT-4 Perform Neural Architecture Search?

For details, see Paper Link by Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie.

ImageNet Experiments

Please refer to the ImageNet experiments for further information.

Reproduce

The results presented in the paper for NAS-Bench-Macro, Channel-Bench-Macro, and NAS-Bench-201 were generated using the code provided below. Although we set the temperature to 0 in the code, it is important to acknowledge that some level of randomness may persist. Consequently, the results obtained from executing the code may not perfectly match a specific experimental run in the paper. Nevertheless, as the code is run multiple times, the average performance should align with the overall performance reported in the paper.

* NAS-Bench-Macro

python nas_bench_macro.py --openai_key {YOUR_OPENAI_API_KEY} --openai_organization {YOUR_OPENAI_ORGANIZATION}

* Channel-Bench-Macro

# For ResNet
python channel_bench_res.py --openai_key {YOUR_OPENAI_API_KEY} --openai_organization {YOUR_OPENAI_ORGANIZATION}

# For MobileNet
python channel_bench_mob.py --openai_key {YOUR_OPENAI_API_KEY} --openai_organization {YOUR_OPENAI_ORGANIZATION}

* NAS-Bench-201

python nas_bench_201.py --openai_key {YOUR_OPENAI_API_KEY} --openai_organization {YOUR_OPENAI_ORGANIZATION} --dataset {DATASET}

For {DATASET}, you can select from ['cifar10', 'cifar100', 'imagenet']. Please note that this is an interactive program, requiring manual interpretation of GPT-4's suggested model into six numbers to retrieve the performance and provide feedback to GPT-4.

Retrieve Performance From Benchmark

* NAS-Bench-Macro

python get_performance.py --benchmark nas-macro --arch xxxxxxxx

xxxxxxxx is 8 numbers (e.g. 01201201) which representes the operation for each layer. There are three different choices for each layer, you can use [0, 1, 2] to represents the operations.

* Channel-Bench-Macro

python get_performance.py --benchmark channel-res --arch 'xx, xx, xx, xx, xx, xx, xx'
python get_performance.py --benchmark channel-mob --arch 'xx, xx, xx, xx, xx, xx, xx'

Use channel-res for ResNet base model and channel-mob for MobileNet base model. xx represents the channel numers of each layer.

* NAS-Bench-201

python get_performance.py --benchmark 201-cifar10  --arch xxxxxx
python get_performance.py --benchmark 201-cifar100 --arch xxxxxx
python get_performance.py --benchmark 201-imagenet --arch xxxxxx

Use 201-cifar10, 201-cifar100, and 201-imagenet for CIFA10, CIFAR100, and ImageNet16-120 respectively. xxxxxx is 6 numbers (e.g. 213401) which representes the operation for each edge. There are three different choices for each layer, you can use [0, 1, 2, 3, 4] to represents the operations.

Reference

@misc{zheng2023gpt4,
    title={Can GPT-4 Perform Neural Architecture Search?}, 
    author={Mingkai Zheng and Xiu Su and Shan You and Fei Wang and Chen Qian and Chang Xu and Samuel Albanie},
    year={2023},
    eprint={2304.10970},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}