Profiling Neural Blocks and Design Spaces for Mobile Neural Architecture Search

This code accompanies the paper which has been accepted as an Applied Full Paper at the 30th ACM International Conference on Information and Knowledge Management (CIKM `21). The paper is available at [arXiv:2109.12426].

[Poster][Video][Slides]

In the paper we profile several well-known Design Spaces like Once-for-All (OFA-MBv3), ProxylessNAS and ResNet50 in terms of metrics like accuracy or inference latency on a number of target devices, like the Huawei Kirin 9000 NPU, Nvidia RTX 2080 Ti GPU, AMD Threadripper 2990WX CPU and Samsung Note10 mobile processor.

We can compare the profiling results to draw conclusions about different Design Spaces perform on varying hardware, e.g. do specific operations require more latency, or how sensitive the network is to block choice in different locations.

Dependencies

python 3.6 or 3.7
pytorch >= 1.4.0
torchvision >= 0.4.0
ofa==0.1.0.post202012082159
torchprofile==0.0.2

Preliminaries:

Data for available in-house predictors (e.g. NPU, GPU and CPU) must be downloaded and used to generate appropriate .pt files by training the predictor before profiling can be done.

1. Download the .csv files from Google Drive and place them in search/rm_search/data/
2. For each .csv file, run the following command:

python -u search/rm_search/run_ofa_op_graph_lat_predictor.py -sub_space <SPACE> -lat_device <DEVICE>

Where the value for the -sub_space flag is the design space, e.g., mbv3 or pn and the value for the -lat_device is the target device, e.g., npu or gpu. Doing so should place the appropriate _best.pt files in models/Latency/.

If you have trouble running run_ofa_op_graph_lat_predictor.py, see the corresponding README in search/rm_search

Note: At this time the GPU and CPU predictor data for ResNet50 has not been publicly released.

Contents:

This code consists of two parts:

1. Block Profiling, performed using main.py
2. Search experiments, performed using scripts in the /search/rm_search directory

Commands to perform profiling

python3 main.py
    --space {OFAPred, OFASupernet, ProxylessSupernet, ResNet50Supernet}
    --num_archs 10 # Number of architectures to evaluate per unit-layer-block fix
    --blocks # Can be used to specify individual unit-layer-block combinations to profile, see the block_sample method in each search space for details
    --all # Evaluate all possible unit-layer-block combinations
    --device # Which device, e.g., CPU or GPU, to use
    --metrics # Can specify profiling specific metrics (e.g. only certain latencies), see the metrics field in each file in /search_spaces/ for details
    --data # Location of ImageNet data. Ignored when --fast is specified
    --fast # Use fast RAM imagenet validation data loader, see /models/imagenet_RAM_saver.py for details
    --save # Name of experiment, where to save. Information will be saved in /logs/{Space}/{Save}
    --no-log # Binary; do not log experiment information

Output data is formatted to be comma separated such that it can easily be transferred to a CSV file for further processing and analysis if desired, e.g., plotting trends.

We ask that if you use this code for your research, please cite our [paper]:

@inbook{mills2021profiling,
author = {Mills, Keith G. and Han, Fred X. and Zhang, Jialin and Changiz Rezaei, Seyed Saeed and Chudak, Fabian and Lu, Wei and Lian, Shuo and Jui, Shangling and Niu, Di},
title = {Profiling Neural Blocks and Design Spaces for Mobile Neural Architecture Search},
year = {2021},
isbn = {9781450384469},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://dl.acm.org/doi/10.1145/3459637.3481944},
booktitle = {Proceedings of the 30th ACM International Conference on Information & Knowledge Management},
pages = {4026–4035},
numpages = {10}
}