gptkbp:instanceOf
|
gptkb:convolutional_neural_network
|
gptkbp:activatedBy
|
gptkb:ReLU
|
gptkbp:advantage
|
reduces number of parameters
encourages feature reuse
improves information flow between layers
reduces vanishing-gradient problem
|
gptkbp:arXivID
|
arXiv:1608.06993
|
gptkbp:block
|
dense block
|
gptkbp:citation
|
over 10,000
|
gptkbp:developedBy
|
gptkb:Laurens_van_der_Maaten
gptkb:Gao_Huang
gptkb:Kilian_Q._Weinberger
gptkb:Zhuang_Liu
|
gptkbp:github
|
https://github.com/liuzhuang13/DenseNet
|
gptkbp:growthForm
|
number of filters added per layer
|
gptkbp:hasConcept
|
connect each layer to every other layer in a feed-forward fashion
|
gptkbp:hasVariant
|
gptkb:DenseNet-121
gptkb:DenseNet-169
gptkb:DenseNet-201
gptkb:DenseNet-264
|
https://www.w3.org/2000/01/rdf-schema#label
|
DenseNet
|
gptkbp:inputToLayer
|
concatenation of outputs from all previous layers
|
gptkbp:license
|
gptkb:MIT_License
|
gptkbp:maximumDepth
|
can be very deep (e.g., 121, 169, 201 layers)
|
gptkbp:normalization
|
Batch Normalization
|
gptkbp:notableFor
|
gptkb:Keras_Applications
gptkb:Torchvision
|
gptkbp:notablePublication
|
gptkb:Densely_Connected_Convolutional_Networks
|
gptkbp:outputLayer
|
Fully Connected Layer
|
gptkbp:parameterEfficiency
|
high
|
gptkbp:platform
|
gptkb:TensorFlow
gptkb:Keras
gptkb:PyTorch
|
gptkbp:pooling
|
gptkb:Max_Pooling
Average Pooling
|
gptkbp:publicationDate
|
gptkb:CVPR_2017
|
gptkbp:publishedIn
|
2017
|
gptkbp:relatedTo
|
gptkb:ResNet
gptkb:VGGNet
gptkb:AlexNet
|
gptkbp:trainer
|
gptkb:SVHN
gptkb:CIFAR-10
gptkb:CIFAR-100
gptkb:ImageNet
|
gptkbp:transitionLayer
|
used between dense blocks
|
gptkbp:usedFor
|
image classification
object detection
semantic segmentation
|
gptkbp:bfsParent
|
gptkb:convolutional_neural_network
|
gptkbp:bfsLayer
|
5
|