Companion code for the ICLR2021 paper: Zero-Cost Proxies for Lightweight NAS
tl;dr A single minibatch of data is used to score neural networks for NAS instead of performing full training.
In this README, we provide:
If you have any questions, please open an issue or email us. (last update: 02.02.2021)
Intro. To perform neural architecture search (NAS), deep neural networks (DNNs) are typically trained until a final validation accuracy is computed and used to compare DNNs to each other and select the best one. However, this is time-consuming because training takes multiple GPU-hours/days/weeks. This is why a proxy for final accuracy is often used to speed up NAS. Typically, this proxy is a reduced form of training (e.g. EcoNAS) where the number of epochs is reduced, a smaller model is used or the training data is subsampled.
Proxies. Instead, we propose a series of "zero-cost" proxies that use a single-minibatch of data to score a DNN.
These metrics are inspired by recent pruning-at-initialization literature, but are adapted to score an entire DNN and work within a NAS setting.
When compared against econas
(see orange pentagon in plot below), our zero-cost metrics take ~1000X less time to run but are better-correlated with final validation accuracy (especially synflow
and jacob_cov
), making them better (and much cheaper!) proxies for use within NAS.
Even when EcoNAS is tuned specifically for NAS-Bench-201 (see econas+
purple circle in the plot), our vote
zero-cost proxy is still better-correlated and is 3 orders of magnitude cheaper to compute.
Figure 1: Correlation of validation accuracy to final accuracy during the first 12 epochs of training (blue line) for three CIFAR-10 on the NAS-Bench-201 search space. Zero-cost and EcoNAS proxies are also labeled for comparison.
Zero-Cost NAS We use the zero-cost metrics to enhance 4 existing NAS algorithms, and we test it out on 3 different NAS benchmarks. For all cases, we achieve a new SOTA (state of the art result) in terms of search speed. We incorporate zero-cost proxies in two ways: (1) warmup: Use proxies to initialize NAS algorithms, (2) move proposal: Use proxies to improve the selection of the next model for evaluation. As Figure 2 shows, there is a significant speedup to all evaluated NAS algorithms.
Figure 2: Zero-Cost warmup and move proposal consistently improves speed and accuracy of 4 different NAS algorithms.
For more details, please take a look at our paper!
- Install PyTorch for your system (v1.5.0 or later).
- Install the package:
pip install .
(add-e
for editable mode) -- note that all dependencies other than pytorch will be automatically installed.
The main function is find_measures
below. Given a neural net and some information about the input data (dataloader
) and loss function (loss_fn
) it returns an array of zero-cost proxy metrics.
def find_measures(net_orig, # neural network
dataloader, # a data loader (typically for training data)
dataload_info, # a tuple with (dataload_type = {random, grasp}, number_of_batches_for_random_or_images_per_class_for_grasp, number of classes)
device, # GPU/CPU device used
loss_fn=F.cross_entropy, # loss function to use within the zero-cost metrics
measure_names=None, # an array of measure names to compute, if left blank, all measures are computed by default
measures_arr=None): # [not used] if the measures are already computed but need to be summarized, pass them here
The available zero-cost metrics are in the measures directory. You can add new metrics by simply following one of the examples then registering the metric in the load_all function. More examples of how to use this function can be found in the code to reproduce results (below). You can also modify data loading functions in p_utils.py
- Download the NAS-Bench-201 dataset and put in the
data
directory in the root folder of this project. - Run python
nasbench2_pred.py
with the appropriate cmd-line options -- a pickle file is produced with zero-cost metrics (seenotebooks
folder on how to use the pickle file. - Note that you need to manually download ImageNet16 and put in
_datasets/ImageNet16
directory in the root folder. CIFAR-10/100 will be automatically downloaded.
- Download the
data
directory and save it to the root folder of this repo. This contains pre-cached info from the NAS-Bench-101 repo. - [Optional] Download the NAS-Bench-101 dataset and put in the
data
directory in the root folder of this project and also clone the NAS-Bench-101 repo and install the package. - Run
python nasbench1_pred.py
. Note that this takes a long time to go through ~400k architectures, but precomputed results are in thenotebooks
folder (with a link to the results).
- Run python
ptcv_pred.py
Coming soon...
For the full list of NAS algorithms in our paper, we used a different NAS tool which is not publicly released. However, we included a notebook nas_examples.ipynb
to show how to use zero-cost proxies to speed up aging evolution and random search methods using both warmup and move proposal.
@inproceedings{
abdelfattah2021zerocost,
title={{Zero-Cost Proxies for Lightweight NAS}},
author={Mohamed S. Abdelfattah and Abhinav Mehrotra and {\L}ukasz Dudziak and Nicholas D. Lane},
booktitle={International Conference on Learning Representations (ICLR)},
year={2021}
}