Introduction

PC-DARTS has been accepted for spotlight presentation at ICLR 2020!

PC-DARTS is a memory-efficient differentiable architecture method based on DARTS. It mainly focuses on reducing the large memory cost of the super-net in one-shot NAS method, which means that it can also be combined with other one-shot NAS method e.g. ENAS. Different from previous methods that sampling operations, PC-DARTS samples channels of the constructed super-net. Interestingly, though we introduced randomness during the search process, the performance of the searched architecture is better and more stable than DARTS! For a detailed description of technical details and experimental results, please refer to our paper:

Partial Channel Connections for Memory-Efficient Differentiable Architecture Search

Yuhui Xu, Lingxi Xie, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Qi Tian and Hongkai Xiong.

This code is based on the implementation of DARTS.

Updates

• The implementation of random sampling is also uploaded for your consideration.
• The main file for search on ImageNet has been uploaded train_search_imagenet.py.

Results

Results on CIFAR10

Method	Params(M)	Error(%)	Search-Cost
AmoebaNet-B	2.8	2.55	3150
DARTSV1	3.3	3.00	0.4
DARTSV2	3.3	2.76	1.0
SNAS	2.8	2.85	1.5
PC-DARTS	3.6	2.57	0.1

Only 0.1 GPU-days are used for a search on CIFAR-10!

Results on ImageNet

Method	FLOPs	Top-1 Error(%)	Top-5 Error(%)	Search-Cost
NASNet-A	564	26.0	8.4	1800
AmoebaNet-B	570	24.3	7.6	3150
PNAS	588	25.8	8.1	225
DARTSV2	574	26.7	8.7	1.0
SNAS	522	27.3	9.3	1.5
PC-DARTS	597	24.2	7.3	3.8

Search a good arcitecture on ImageNet by using the search space of DARTS(First Time!).

Usage

Search on CIFAR10

To run our code, you only need one Nvidia 1080ti(11G memory).

python train_search.py \\

Search on ImageNet

Data preparation: 10% and 2.5% images need to be random sampled prior from earch class of trainingset as train and val, respectively. The sampled data is save into ./imagenet_search. Note that not to use torch.utils.data.sampler.SubsetRandomSampler for data sampling as imagenet is too large.

python train_search_imagenet.py \\
       --tmp_data_dir /path/to/your/sampled/data \\
       --save log_path \\

The evaluation process simply follows that of DARTS.

Here is the evaluation on CIFAR10:

python train.py \\
       --auxiliary \\
       --cutout \\

Here is the evaluation on ImageNet (mobile setting):

python train_imagenet.py \\
       --tmp_data_dir /path/to/your/data \\
       --save log_path \\
       --auxiliary \\
       --note note_of_this_run

Pretrained models

Coming soon!.

Notes

• For the codes in the main branch, python2 with pytorch(3.0.1) is recommended （running on Nvidia 1080ti）. We also provided codes in the V100_python1.0 if you want to implement PC-DARTS on Tesla V100 with python3+ and pytorch1.0+.

• You can even run the codes on a GPU with memory only 4G. PC-DARTS only costs less than 4G memory, if we use the same hyper-parameter settings as DARTS(batch-size=64).

• You can search on ImageNet by model_search_imagenet.py! The training file for search on ImageNet will be uploaded after it is cleaned or you can generate it according to the train_search file on CIFAR10 and the evluate file on ImageNet. Hyperparameters are reported in our paper! The search cost 11.5 hours on 8 V100 GPUs(16G each). If you have V100(32G) you can further increase the batch-size.

• We random sample 10% and 2.5% from each class of training dataset of ImageNet. There are still 1000 classes! Replace input_search, target_search = next(iter(valid_queue)) with following codes would be much faster:

    try:
      input_search, target_search = next(valid_queue_iter)
    except:
      valid_queue_iter = iter(valid_queue)
      input_search, target_search = next(valid_queue_iter)

• The main codes of PC-DARTS are in the file model_search.py. As descriped in the paper, we use an efficient way to implement the channel sampling. First, a fixed sub-set of the input is selected to be fed into the candidate operations, then the concated output is swaped. Two efficient swap operations are provided: channel-shuffle and channel-shift. For the edge normalization, we define edge parameters(beta in our codes) along with the alpha parameters in the original darts codes.

• The implementation of random sampling is also provided model_search_random.py. It also works while channel-shuffle may have better performance.

• As PC-DARTS is an ultra memory-efficient NAS methods. It has potentials to be implemented on other tasks such as detection and segmentation.

Related work

Progressive Differentiable Architecture Search

Differentiable Architecture Search

Reference

If you use our code in your research, please cite our paper accordingly.

@inproceedings{
xu2020pcdarts,
title={{\{}PC{\}}-{\{}DARTS{\}}: Partial Channel Connections for Memory-Efficient Architecture Search},
author={Yuhui Xu and Lingxi Xie and Xiaopeng Zhang and Xin Chen and Guo-Jun Qi and Qi Tian and Hongkai Xiong},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=BJlS634tPr}
}