ResNet(channel_config, block_config, shortcut_config;
       block, connection = addrelu, nclasses = 1000)

Create a ResNet model (reference). See also resnet.

Arguments

• channel_config: the growth rate of the output feature maps within a residual block
• block_config: a list of the number of residual blocks at each stage
• shortcut_config: the type of shortcut style (either :A, :B, or :C). shortcut_config can also be a vector of symbols if different shortcut styles are applied to different residual blocks.
• block: a function with input (inplanes, outplanes, downsample=false) that returns a new residual block (see Metalhead.basicblock and Metalhead.bottleneck)
• connection: the binary function applied to the output of residual and skip paths in a block
• nclasses: the number of output classes

ResNet(depth = 50; pretrain = false, nclasses = 1000)

Create a ResNet model with a specified depth (reference) following these modification referred as ResNet v1.5.

See also Metalhead.resnet.

Arguments

• depth: depth of the ResNet model. Options include (18, 34, 50, 101, 152).
• nclasses: the number of output classes

For ResNet(18) and ResNet(34), the parameter-free shortcut style (type :A) is used in the first block and the three other blocks use type :B connection (following the implementation in PyTorch). The published version of ResNet(18) and ResNet(34) used type :A shortcuts for all four blocks. The example below shows how to create a 18 or 34-layer ResNet using only type :A shortcuts:

using Metalhead

resnet18 = ResNet([1, 1], [2, 2, 2, 2], :A; block = Metalhead.basicblock)

resnet34 = ResNet([1, 1], [3, 4, 6, 3], :A; block = Metalhead.basicblock)

The bottleneck of the orginal ResNet model has a stride of 2 on the first convolutional layer when downsampling (instead of the second convolutional layers as in ResNet v1.5). The architecture of the orignal ResNet model can be obtained as shown below:

resnet50_v1 = ResNet([1, 1, 4], [3, 4, 6, 3], :B; block = Metalhead.bottleneck_v1)