We present the first large-scale energy-aware benchmark that allows studying AutoML methods to achieve better trade-offs between performance and search energy consumption, named EA-HAS-Bench. EA-HAS-Bench provides a large-scale architecture/hyperparameter joint search space, covering diversified configurations related to energy consumption. Furthermore, we propose a novel surrogate model specially designed for large joint search space, which proposes a Bezier curve-based model to predict learning curves with unlimited shape and length.
Most of the existing conventional benchmarks like NAS-Bench-101 do not directly provide training energy cost but use model training time as the training resource budget, which as verified by our experiments, is an inaccurate estimation of energy cost. HW-NAS-bench provides the inference latency and inference energy consumption of different model architectures but also does not provide the search energy cost
Unlike the search space of existing mainstream NAS-Bench that focuses only on network architectures, our EA-HAS-Bench consists of a combination of two parts: the network architecture space-
The EA-HAS-Bench dataset provides the following three types of metrics to evaluate different configurations.
- Model Complexity: parameter size, FLOPs, number of network activations (the size of the output tensors of each convolutional layer), as well as the inference energy cost of the trained model.
- Model Performance: full training information including training, validation, and test accuracy learning curves.
- Search Cost: energy cost (in kWh) and time (in seconds)
The left plot shows the validation accuracy box plots for each NAS benchmark in CIFAR-10. The right plot shows a comparison of training time, training energy consumption (TEC), and test accuracy in the dataset.
Although training the model for a longer period is likely to yield a higher energy cost, the final cost still depends on many other factors including power (i.e., consumed energy per hour). The left and right plots of Figure also verifies the conclusion, where the models in the Pareto Frontier on the accuracy-runtime coordinate (right figure) are not always in the Pareto Frontier on the accuracy-TEC coordinate (left figure), showing that training time and energy cost are not equivalent.
Clone this repository and install its requirements.
git clone https://github.com/microsoft/EA-HAS-Bench
cd EA-HAS-Bench
cat requirements.txt | xargs -n 1 -L 1 pip install
pip install -e .Download the pretrained surrogate models and place them into BSC/checkpoints/. The current models are v0.1.
NOTE: This codebase is still subject to changes. Upcoming updates include improved versions of the surrogate models and code for all experiments from the paper. The API may still be subject to changes.
Besides providing a large-scale proxy benchmark and the tens of thousands of sampling points used to construct it, we also provide a small real tabular benchmark. We redefine a very small joint search space with a size of 500. The latest benchmark file of NAS-Toy can be downloaded from One Drive.
The api is located in api.py.
Here is an example of how to use the API:
def get_ea_has_bench_api(dataset):
full_api = {}
# load the ea-nas-bench surrogate models
if dataset=="cifar10":
ea_has_bench_model = load_ensemble('checkpoints/ea_has_bench-v0.2'))
train_energy_model = load_ensemble('checkpoints/ea_has_bench-trainE-v0.2')
test_energy_model = load_ensemble('checkpoints/ea_has_bench-testE-v0.1')
runtime_model = load_ensemble('checkpoints/ea_has_bench-runtime-v0.1')
elif dataset=="tiny":
ea_has_bench_model = load_ensemble('checkpoints/ea-nas-bench-tiny-v0.2')
train_energy_model = load_ensemble('checkpoints/ea-nas-bench-trainE-v0.1')
test_energy_model = load_ensemble('checkpoints/ea-nas-bench-testE-v0.1')
runtime_model = load_ensemble('checkpoints/ea-nas-bench-runtime-v0.1')
full_api['ea_has_bench_model'] = [ea_has_bench_model, runtime_model, train_energy_model, test_energy_model]
return full_api
ea_api = get_ea_has_bench_api("cifar10")
# output the learning curve, train time, TEC and IEC
lc = ea_api['ea_has_bench_model'][0].predict(config=arch_str)
train_time = ea_api['ea_has_bench_model'][1].predict(config=arch_str)
train_cost = ea_api['ea_has_bench_model'][2].predict(config=arch_str)
test_cost = ea_api['ea_has_bench_model'][3].predict(config=arch_str)# Supported optimizers: rs, re, {EA}-(ls bananas), hb, bohb
cd naslib
bash naslib/benchmarks/nas/run_nbgree.sh
bash naslib/benchmarks/nas/run_nbtoy.sh
Results will be saved in results/.
For EA-HAS-Bench’s search space that contains both model architectures and hyperparameters, we use random search (RS) to sample unbiased data to build a robust surrogate benchmark.
The following command will train all architecture candidate in the search space.
cd RegNet+HPO
python tools/azure_sweep.py --mode amlt --config_path configs\sweeps\cifar\mb_v0.4.yaml
python tools/azure_sweep.py --mode amlt --config_path configs\sweeps\tinyimagenet\mb_v0.1.yaml
After training these candidate architectures, please use the following command to re-organize all logs into the single file.
python tools/sweep_collect.py
To fit a Bézier Curves-based surrogated model surrogate model run
cd BSC
python fit_model.py --search_space regnet --model bezier_nn_STAR
@inproceedings{ea23iclr,
title={{EA}-{HAS}-Bench: Energy-aware Hyperparameter and Architecture Search Benchmark},
author={Dou, Shuguang and JIANG, XINYANG and Zhao, Cai Rong and Li, Dongsheng},
booktitle={International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=n-bvaLSCC78},
note={ICLR 2023 notable top 25%}
}
