Figure 1.
>The image shows a vertical architectural diagram of a convolutional neural network, illustrating the flow of data through successive layers. At the top is the input layer, displaying the input shape as None, 224, 224, 3, followed by zero padding and a convolution layer that reduces spatial dimensions. Subsequent layers include batch normalisation and R e L U activation, maintaining feature map dimensions. The diagram then presents an expanded depthwise convolution block, consisting of depthwise convolution, batch normalisation, and R e L U activation, all preserving the same spatial resolution. This is followed by a projection convolution layer that reduces the number of channels, along with batch normalisation. At the bottom, an expansion convolution layer increases the channel depth again. Each block lists both input and output tensor shapes, allowing clear tracking of dimensional changes through the network. The layout emphasises the sequential structure and functional components of the model.Layers of the component of the first sequential part
Source: Figure by authors