Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
203 lines (160 sloc) 11.7 KB

NAS-BENCH-201: Extending the Scope of Reproducible Neural Architecture Search

We propose an algorithm-agnostic NAS benchmark (NAS-Bench-201) with a fixed search space, which provides a unified benchmark for almost any up-to-date NAS algorithms. The design of our search space is inspired by that used in the most popular cell-based searching algorithms, where a cell is represented as a directed acyclic graph. Each edge here is associated with an operation selected from a predefined operation set. For it to be applicable for all NAS algorithms, the search space defined in NAS-Bench-201 includes 4 nodes and 5 associated operation options, which generates 15,625 neural cell candidates in total.

In this Markdown file, we provide:

Note: please use PyTorch >= 1.2.0 and Python >= 3.6.0.

You can simply type pip install nas-bench-201 to install our api. Please see source codes of nas-bench-201 module in this repo.

If you have any questions or issues, please post it at here or email me.

Preparation and Download

The benchmark file of NAS-Bench-201 can be downloaded from Google Drive or Baidu-Wangpan (code:6u5d). You can move it to anywhere you want and send its path to our API for initialization.

  • v1.0: NAS-Bench-201-v1_0-e61699.pth, where e61699 is the last six digits for this file. It contains all information except for the trained weights of each trial.
  • v1.0: The full data of each architecture can be download from Google Drive (about 226GB). This compressed folder has 15625 files containing the the trained weights.
  • v1.0: Checkpoints for 3 runs of each baseline NAS algorithm are provided in Google Drive.

The training and evaluation data used in NAS-Bench-201 can be downloaded from Google Drive or Baidu-Wangpan (code:4fg7). It is recommended to put these data into $TORCH_HOME (~/.torch/ by default). If you want to generate NAS-Bench-201 or similar NAS datasets or training models by yourself, you need these data.

How to Use NAS-Bench-201

  1. Creating an API instance from a file:
from nas_201_api import NASBench201API as API
api = API('$path_to_meta_nas_bench_file')
api = API('NAS-Bench-201-v1_0-e61699.pth')
api = API('{:}/{:}'.format(os.environ['TORCH_HOME'], 'NAS-Bench-201-v1_0-e61699.pth'))
  1. Show the number of architectures len(api) and each architecture api[i]:
num = len(api)
for i, arch_str in enumerate(api):
  print ('{:5d}/{:5d} : {:}'.format(i, len(api), arch_str))
  1. Show the results of all trials for a single architecture:
# show all information for a specific architecture
api.show(1)
api.show(2)

# show the mean loss and accuracy of an architecture
info = api.query_meta_info_by_index(1)  # This is an instance of `ArchResults`
res_metrics = info.get_metrics('cifar10', 'train') # This is a dict with metric names as keys
cost_metrics = info.get_comput_costs('cifar100') # This is a dict with metric names as keys, e.g., flops, params, latency

# get the detailed information
results = api.query_by_index(1, 'cifar100') # a dict of all trials for 1st net on cifar100, where the key is the seed
print ('There are {:} trials for this architecture [{:}] on cifar100'.format(len(results), api[1]))
print ('Latency : {:}'.format(results[0].get_latency()))
print ('Train Info : {:}'.format(results[0].get_train()))
print ('Valid Info : {:}'.format(results[0].get_eval('x-valid')))
print ('Test  Info : {:}'.format(results[0].get_eval('x-test')))
# for the metric after a specific epoch
print ('Train Info [10-th epoch] : {:}'.format(results[0].get_train(10)))
  1. Query the index of an architecture by string
index = api.query_index_by_arch('|nor_conv_3x3~0|+|nor_conv_3x3~0|avg_pool_3x3~1|+|skip_connect~0|nor_conv_3x3~1|skip_connect~2|')
api.show(index)
  1. For other usages, please see lib/nas_201_api/api.py

Detailed Instruction

In nas_201_api, we define three classes: NASBench201API, ArchResults, ResultsCount.

ResultsCount maintains all information of a specific trial. One can instantiate ResultsCount and get the info via the following codes (000157-FULL.pth saves all information of all trials of 157-th architecture):

from nas_201_api import ResultsCount
xdata  = torch.load('000157-FULL.pth')
odata  = xdata['full']['all_results'][('cifar10-valid', 777)]
result = ResultsCount.create_from_state_dict( odata )
print(result) # print it
print(result.get_train())   # print the final training loss/accuracy/[optional:time-cost-of-a-training-epoch]
print(result.get_train(11)) # print the training info of the 11-th epoch
print(result.get_eval('x-valid'))     # print the final evaluation info on the validation set
print(result.get_eval('x-valid', 11)) # print the info on the validation set of the 11-th epoch
print(result.get_latency())           # print the evaluation latency [in batch]
result.get_net_param()                # the trained parameters of this trial
arch_config = result.get_config(CellStructure.str2structure) # create the network with params
net_config  = dict2config(arch_config, None)
network    = get_cell_based_tiny_net(net_config)
network.load_state_dict(result.get_net_param())

ArchResults maintains all information of all trials of an architecture. Please see the following usages:

from nas_201_api import ArchResults
xdata   = torch.load('000157-FULL.pth')
archRes = ArchResults.create_from_state_dict(xdata['less']) # load trials trained with  12 epochs
archRes = ArchResults.create_from_state_dict(xdata['full']) # load trials trained with 200 epochs

print(archRes.arch_idx_str())      # print the index of this architecture 
print(archRes.get_dataset_names()) # print the supported training data
print(archRes.get_comput_costs('cifar10-valid')) # print all computational info when training on cifar10-valid 
print(archRes.get_metrics('cifar10-valid', 'x-valid', None, False)) # print the average loss/accuracy/time on all trials
print(archRes.get_metrics('cifar10-valid', 'x-valid', None,  True)) # print loss/accuracy/time of a randomly selected trial

NASBench201API is the topest level api. Please see the following usages:

from nas_201_api import NASBench201API as API
api = API('NAS-Bench-201-v1_0-e61699.pth') # This will load all the information of NAS-Bench-201 except the trained weights
api = API('{:}/{:}'.format(os.environ['TORCH_HOME'], 'NAS-Bench-201-v1_0-e61699.pth')) # The same as the above line while I usually save NAS-Bench-201-v1_0-e61699.pth in ~/.torch/.
api.show(-1)  # show info of all architectures
api.reload('{:}/{:}'.format(os.environ['TORCH_HOME'], 'NAS-BENCH-201-4-v1.0-archive'), 3) # This code will reload the information 3-th architecture with the trained weights

weights = api.get_net_param(3, 'cifar10', None) # Obtaining the weights of all trials for the 3-th architecture on cifar10. It will returns a dict, where the key is the seed and the value is the trained weights.

Instruction to Re-Generate NAS-Bench-201

There are four steps to build NAS-Bench-201.

  1. generate the meta file for NAS-Bench-201 using the following script, where NAS-BENCH-201 indicates the name and 4 indicates the maximum number of nodes in a cell.
bash scripts-search/NAS-Bench-201/meta-gen.sh NAS-BENCH-201 4
  1. train earch architecture on a single GPU (see commands in output/NAS-BENCH-201-4/BENCH-201-N4.opt-full.script, which is automatically generated by step-1).
CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/NAS-Bench-201/train-models.sh 0     0   389 -1 '777 888 999'

This command will train 390 architectures (id from 0 to 389) using the following four kinds of splits with three random seeds (777, 888, 999).

Dataset Train Eval
CIFAR-10 train valid / test
CIFAR-10 train + valid test
CIFAR-100 train valid / test
ImageNet-16-120 train valid / test

Note that the above train, valid, and test indicate the proposed splits in our NAS-Bench-201, and they might be different with the original splits.

  1. calculate the latency, merge the results of all architectures, and simplify the results. (see commands in output/NAS-BENCH-201-4/meta-node-4.cal-script.txt which is automatically generated by step-1).
OMP_NUM_THREADS=6 CUDA_VISIBLE_DEVICES=0 python exps/NAS-Bench-201/statistics.py --mode cal --target_dir 000000-000389-C16-N5
  1. merge all results into a single file for NAS-Bench-201-API.
OMP_NUM_THREADS=4 python exps/NAS-Bench-201/statistics.py --mode merge

This command will generate a single file output/NAS-BENCH-201-4/simplifies/C16-N5-final-infos.pth contains all the data for NAS-Bench-201. This generated file will serve as the input for our NAS-Bench-201 API.

[option] train a single architecture on a single GPU.

CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/NAS-Bench-201/train-a-net.sh resnet 16 5
CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/NAS-Bench-201/train-a-net.sh '|nor_conv_3x3~0|+|nor_conv_3x3~0|nor_conv_3x3~1|+|skip_connect~0|skip_connect~1|skip_connect~2|' 16 5

To Reproduce 10 Baseline NAS Algorithms in NAS-Bench-201

We have tried our best to implement each method. However, still, some algorithms might obtain non-optimal results since their hyper-parameters might not fit our NAS-Bench-201. If researchers can provide better results with different hyper-parameters, we are happy to update results according to the new experimental results. We also welcome more NAS algorithms to test on our dataset and would include them accordingly.

Note that you need to prepare the training and test data as described in Preparation and Download

  • [1] CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/DARTS-V1.sh cifar10 1 -1, where cifar10 can be replaced with cifar100 or ImageNet16-120.
  • [2] CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/DARTS-V2.sh cifar10 1 -1
  • [3] CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/GDAS.sh cifar10 1 -1
  • [4] CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/SETN.sh cifar10 1 -1
  • [5] CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/ENAS.sh cifar10 1 -1
  • [6] CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/RANDOM-NAS.sh cifar10 1 -1
  • [7] bash ./scripts-search/algos/R-EA.sh cifar10 3 -1
  • [8] bash ./scripts-search/algos/Random.sh -1
  • [9] bash ./scripts-search/algos/REINFORCE.sh 0.5 -1
  • [10] bash ./scripts-search/algos/BOHB.sh -1

In commands [1-6], the first args cifar10 indicates the dataset name, the second args 1 indicates the behavior of BN, and the first args -1 indicates the random seed.

Citation

If you find that NAS-Bench-201 helps your research, please consider citing it:

@inproceedings{dong2020nasbench201,
  title     = {NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {International Conference on Learning Representations (ICLR)},
  url       = {https://openreview.net/forum?id=HJxyZkBKDr},
  year      = {2020}
}
You can’t perform that action at this time.