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.
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.
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](https://github.com/D-X-Y/NAS-Bench-201).
[deprecated] The benchmark file of NAS-Bench-201 can be downloaded from [Google Drive](https://drive.google.com/open?id=1SKW0Cu0u8-gb18zDpaAGi0f74UdXeGKs) or [Baidu-Wangpan (code:6u5d)](https://pan.baidu.com/s/1CiaNH6C12zuZf7q-Ilm09w).
[recommended] The benchmark file of NAS-Bench-201 can be downloaded from [Google Drive](https://drive.google.com/open?id=1OOfVPpt-lA4u2HJrXbgrRd42IbfvJMyE). The files for model weight are too large (431G) and I need some time to upload it. Please be patient, thanks for your understanding.
- [2020.02.25] APIv1.0/FILEv1.0: `NAS-Bench-201-v1_0-e61699.pth` (2.2G), where `e61699` is the last six digits for this file. It contains all information except for the trained weights of each trial.
- [2020.02.25] APIv1.0/FILEv1.0: The full data of each architecture can be download from [
NAS-BENCH-201-4-v1.0-archive.tar](https://drive.google.com/open?id=1X2i-JXaElsnVLuGgM4tP-yNwtsspXgdQ) (about 226GB). This compressed folder has 15625 files containing the the trained weights.
- [2020.02.25] APIv1.0/FILEv1.0: Checkpoints for 3 runs of each baseline NAS algorithm are provided in [Google Drive](https://drive.google.com/open?id=1eAgLZQAViP3r6dA0_ZOOGG9zPLXhGwXi).
- [2020.03.09] APIv1.2/FILEv1.0: More robust API with more functions and descriptions
- [2020.03.16] APIv1.3/FILEv1.1: `NAS-Bench-201-v1_1-096897.pth` (4.7G), where `096897` is the last six digits for this file. It contains information of more trials compared to `NAS-Bench-201-v1_0-e61699.pth`, especially all models trained by 12 epochs on all datasets are avaliable.
The training and evaluation data used in NAS-Bench-201 can be downloaded from [Google Drive](https://drive.google.com/open?id=1L0Lzq8rWpZLPfiQGd6QR8q5xLV88emU7) or [Baidu-Wangpan (code:4fg7)](https://pan.baidu.com/s/1XAzavPKq3zcat1yBA1L2tQ).
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.
config = api.get_net_config(123, 'cifar10') # obtain the network configuration for the 123-th architecture on the CIFAR-10 dataset
from models import get_cell_based_tiny_net # this module is in AutoDL-Projects/lib/models
network = get_cell_based_tiny_net(config) # create the network from configurration
print(network) # show the structure of this architecture
```
If you want to load the trained weights of this created network, you need to use `api.get_net_param(123, ...)` to obtain the weights and then load it to the network.
6.`api.get_more_info(...)` can return the loss / accuracy / time on training / validation / test sets, which is very helpful. For more details, please look at the comments in the get_more_info function.
7. For other usages, please see `lib/nas_201_api/api.py`. We provide some usage information in the comments for the corresponding functions. If what you want is not provided, please feel free to open an issue for discussion, and I am happy to answer any questions regarding NAS-Bench-201.
`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):
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.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.
To obtain the training and evaluation information (please see the comments [here](https://github.com/D-X-Y/AutoDL-Projects/blob/master/lib/nas_201_api/api.py#L172)):
api.get_more_info(112, 'ImageNet16-120', None, False, True) # the info of last training epoch for 112-th architecture (use 200-epoch-hyper-parameter and randomly select a trial)
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.
2. 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).
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.
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.
- [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`.
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.
**Note that** since 2020 March 16, in these scripts, the default NAS-Bench-201 benchmark file has changed from `NAS-Bench-201-v1_0-e61699.pth` to `NAS-Bench-201-v1_1-096897.pth`, and thus the results could be slightly different from the original paper.
