SqueezeNet

E431007

convolutional neural network architecture deep learning model image classification model

SqueezeNet is a compact deep convolutional neural network architecture designed to achieve AlexNet-level image classification accuracy with dramatically fewer parameters, making it efficient for deployment on resource-constrained devices.

Try in SPARQL Jump to: Surface forms Statements Referenced by

All labels observed (1)

Label	Occurrences
SqueezeNet canonical	1

Statements (51)

Predicate	Object
instanceOf	convolutional neural network architecture ⓘ deep learning model ⓘ image classification model ⓘ
accuracyComparableTo	AlexNet NERFINISHED ⓘ
benchmarkedOn	ImageNet NERFINISHED ⓘ
designedBy	Forrest N. Iandola NERFINISHED ⓘ Khalid Ashraf NERFINISHED ⓘ Kurt Keutzer NERFINISHED ⓘ Matthew W. Moskewicz NERFINISHED ⓘ Song Han NERFINISHED ⓘ William J. Dally NERFINISHED ⓘ
developedAt	DeepScale NERFINISHED ⓘ Stanford University NERFINISHED ⓘ University of California, Berkeley NERFINISHED ⓘ
hasComponent	Fire module ⓘ expand layer with 1x1 and 3x3 filters ⓘ squeeze layer with 1x1 filters ⓘ
hasGoal	achieve AlexNet-level accuracy with dramatically fewer parameters ⓘ enable deployment on resource-constrained devices ⓘ
hasVersion	SqueezeNet v1.0 NERFINISHED ⓘ SqueezeNet v1.1 NERFINISHED ⓘ
implementedIn	Caffe NERFINISHED ⓘ PyTorch NERFINISHED ⓘ TensorFlow NERFINISHED ⓘ
influencedBy	AlexNet NERFINISHED ⓘ
keyIdea	decrease number of input channels to 3x3 filters ⓘ downsample late in the network to maintain large activation maps ⓘ replace many 3x3 filters with 1x1 filters ⓘ
license	permissive open-source license (via reference implementations) ⓘ
modelSize	less than 0.5 MB (for some configurations) ⓘ
notableProperty	amenable to further compression techniques such as pruning and quantization ⓘ suitable for deployment over low-bandwidth networks ⓘ very small model size compared to AlexNet ⓘ
openSource	true ⓘ
optimizedFor	embedded devices ⓘ mobile devices ⓘ resource-constrained hardware ⓘ
parameterCountRelativeTo	50x fewer parameters than AlexNet ⓘ
publicationTitle	SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size NERFINISHED ⓘ
publicationYear	2016 ⓘ
supportsOperation	distributed training ⓘ model compression ⓘ
targetTask	image classification ⓘ
usesLayerType	1x1 convolution ⓘ 3x3 convolution ⓘ Fire module ⓘ convolutional layer ⓘ expand layer ⓘ max pooling layer ⓘ squeeze layer ⓘ
v1.1Characteristics	faster with similar accuracy compared to SqueezeNet v1.0 ⓘ

Referenced by (1)

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

torchvision (ecosystem) → modelFamily → SqueezeNet ⓘ

subject surface form: torchvision