Skip to content

Latest commit

 

History

History
27 lines (27 loc) · 1.68 KB

1608.06993 - Densely Connected Convolutional Networks.md

File metadata and controls

27 lines (27 loc) · 1.68 KB

♥ DL , CV , CNN, Recognition

  • One of CVPR2017 best paper, a new CNN archetecture
  • Advantages:
    • Alleviate gradient vanishing
    • strengthen feature propagation
    • encourage feature reuse
    • reduce number of parameters
    • Can be trained as similar steps in ResNet
  • Limitations: from DPN
    • Width of the densely connected path linearly increases as the depth rises
    • This may cause the number of parameters to grow quadratically compared with the residual networks if the implementation is not specifically optimized
  • Archetecture:
    • Difference bwtween ResNet and DenseNet:
      • ResNet adds the input features to the output features through the residual path: $x(l) = H_l(l-1) + x_{l-1}$
      • DenseNet uses a densely connected path to concatenate the input features with the output features : $x(l) = H_l([x_0,...,x_{l-1}])$
      • This enables each micro-block to receive raw information from all previous micro-blocks
    • Preactivation, i.e BN->ReLU->1x1Conv->BN->ReLU->3x3Conv
    • To make pooling easier(the dimensions may increase too fast), halve feature dimension using conv before pooling
    • Growth rate k(the number of 3x3 kernels after each part):
      • After each part, the dimension of next part will increase by k
      • Larger k means more information will be accessed, while the computational complexity will be increased
      • In this paper k = 32/48
  • Implementation: