Convolução em profundidade
A convolução depthwise é uma forma especializada de convolução usada principalmente em aprendizado profundo, particularly in the context of redes neurais convolucionais (CNNs). Unlike standard convolution, which combines multiple input channels into a single canal de saída, depthwise convolution operates on each input channel independently. This process significantly reduces the number of parameters and computations required, making it an efficient alternative for mobile and embedded systems.
Em uma convolução depthwise típica, cada canal de entrada é convolvido com its own set of filters. For example, if an input tensor has C channels and each filter has a size of KxK, then C separate filters are applied, one for each channel. The output of this operation is a new tensor with the same number of channels, but the spatial dimensions may vary depending on the stride and padding used.
Depthwise convolution is often used in conjunction with pointwise convolution, which applies a 1×1 convolution to mix the outputs of the depthwise convolution across channels. This combination is commonly found in architectures like MobileNet, where maintaining high performance while minimizing computational load is crucial.
Overall, depthwise convolution enhances the efficiency of deep learning models, allowing them to run faster and require less memory, making them suitable for real-time applications and devices with limited resources.