深度学习模型绘图模板 .pptxVIP

卷积神经网络结构（1）（2）............................01895×5Convolution2×2Subsampling5×5Convolution2×2SubsamplingFullyConnectedFullyConnectedInput1×28×28Featuremaps4×24×24Featuremaps4×12×12Featuremaps8×8×8Featuremaps8×4×4n1Outputn2C1S1C2S2FeatureExtractionClassification卷积池化池化卷积全连接全连接输出卷积层全连接层输出层输入层123

??~??^Split^~Kernel3×3Kernel5×5??gp??fcsoftmaxSelectFuseelement-wisesummationelement-wiseproductSelectiveKernelConvolutionOriginalimage/abs/1903.06586

103012211023220112011111030200200120100000200····4··55·····5·输入图像6×6卷积核3×3特征图4×4卷积11110001110011001100111000101011014=5×5imageconvolutionfilterfeaturemap（1）（2）

CONV11x11.stride=4，96kernelsOverlappingMaxPOOL3x3,stride=2CONV5x5,pad=2256kernelsOverlappingMaxPOOL3x3,stride=2CONV3x3,pad=1374kernelsCONV3x3,pad=1384kernelsOverlappingMaxPOOL3x3,stride=2CONV3x3,pad=1256kernelsFCFC

OverlappingMaxPOOL3x3,stride=21313CONV11x11，stride=4，96kernelsOverlappingMaxPOOL3x3,stride=2CONV5x5,pad=2256kernelsCONV3x3,pad=1384kernelsCONV3x3,pad=1384kernelsOverlappingMaxPOOL3x3,stride=2CONV3x3,pad=1256kernelsFCFC22722711113555596279627272725613132561313384131338425666256409640961000softmax

CONV3x3，stride=1，32kernels3232332323233323232MaxPOOL2x2,stride=2321616CONV3x3，stride=1，64kernels1616161664MaxPOOL2x2,stride=2CONV3x3,128kernels86488888128MaxPOOL2x2,stride=244128FCFC204851210softmaxSENet166416

MaxPOOL2x2,stride=288CONV3x3，stride=1，96kernelsMaxPOOL2x2,stride=2CONV3x3,256kernelsCONV3x3,384kernelsCONV3x3,384kernelsMaxPOOL2x2,stride=2CONV3x3,256kernelsFCFC32323323296169616161625688256888838425644256409651210softmax33SENetSENetSENet384

4681539723156332005070600012000085955267PoolingIndices（池化索引）MaxPooling（最大池化）Unpooling（反池化）池化2413509423080744平均池化步长为2，大小为2×2最大池化步长为2，大小为2×245893326（1）（2）

转置卷积,k=3,s=2,p=0卷积,k=3,s=1,p=0卷积/转置卷积

Dropout网络模型BeforeDropoutAfterDropout

全连接层结构.........输入层隐藏层输出层（1）...InputLayerFullyConnectedLayerOutputLayer（2）

MLP多层感知器（1）（2）ΣΣΣ1Softmax1x1x2ΣΣΣΣSoftmaxOutputLa