Fig. 3

The mobile inverted bottleneck convolution (MBConv) block. (a) An initial 1x1 conv block expands the number of input channels according to the expansion factor hyper-parameter. (b) Depth-wise 3x3 conv block over channels. (c) Global average pooling shrinks the tensor along its spatial dimensions. (d, e) A squeeze conv (1x1 conv + swish) and an excitation conv (1x1 conv + sigmoid) first squeeze the channel dimension by a factor of 0.25, then expand it back to its original shape. The output is multiplied by the output tensor from step (b). (f) A final 1x1 conv block with a linear activation maps the tensor to the desired number of output channels, followed by a dropout layer for stochastic depth (dropout rate 0.2)