ProxylessNAS

E899018

ProxylessNAS is a neural architecture search method that directly optimizes neural network architectures on target tasks and hardware without relying on proxy models, enabling efficient and hardware-aware network design.

Jump to: Statements Referenced by

Statements (47)

Predicate Object
instanceOf NAS algorithm
deep learning technique
neural architecture search method
aimsTo directly optimize architectures on target hardware
directly optimize architectures on target tasks
codeAvailability open source implementation
comparedWith AmoebaNet NERFINISHED
DARTS NERFINISHED
NASNet NERFINISHED
doesNotUse proxy networks
proxy tasks
evaluatedOn ImageNet NERFINISHED
field computer vision
machine learning
fullName Proxyless Neural Architecture Search NERFINISHED
handles discrete architecture choices via relaxation
hardwarePlatform CPUs
GPUs
mobile phones
improvesOn deployment realism
search efficiency
keyIdea jointly consider accuracy and hardware efficiency
remove proxy gap between search and deployment
latencyModel measured on real hardware
objective learn architecture parameters
learn network weights
optimizedFor target hardware
target tasks
proposedBy Han Cai NERFINISHED
Ligeng Zhu NERFINISHED
Song Han NERFINISHED
publicationYear 2019
publishedIn International Conference on Learning Representations NERFINISHED
relatedTo efficient deep learning
model compression
neural architecture search NERFINISHED
searchSpace ResNet-like architectures
mobile CNN architectures
searchStrategy differentiable NAS
supports hardware-aware neural architecture search
targetMetric FLOPs
accuracy
latency
model size
uses gradient-based optimization
path-level binarization
stochastic binary gates

Referenced by (1)

Full triples — surface form annotated when it differs from this entity's canonical label.