Depthwise Separable Convolution
Depthwise Separable Convolution: Depthwise separable convolution factorizes a standard convolution into spatial filtering per channel followed by channel mixing. It is a common building block in efficient convolutional networks. The depthwise stage applies one k × k filter independently to each input channel.
In italiano: Depthwise Separable ConvolutionLegge III · Come una rete impara dai datiThe depthwise stage applies one k × k filter independently to each input channel. The pointwise stage then uses a 1 × 1 convolution to combine channels and set the output width. Separating these operations substantially reduces parameters and multiply-accumulate operations compared with a standard convolution.
Formula
Standard: k²CᵢₙCₒᵤₜ parameters; depthwise separable: k²Cᵢₙ + CᵢₙCₒᵤₜ
Numerical example
import torch.nn as nn
# 32 input channels, 64 output channels, 3x3 spatial filters
block = nn.Sequential(
nn.Conv2d(32, 32, 3, padding=1, groups=32, bias=False),
nn.Conv2d(32, 64, 1, bias=False)
)
# 3*3*32*64 = 18,432 standard parameters
# 3*3*32 + 32*64 = 2,336 factorized parametersWhen to use it
Use it in mobile, edge and latency-sensitive vision models, including MobileNet-style architectures, when reducing compute and parameter count is a primary constraint.
Limitations
The factorization restricts how spatial and channel information interact and may reduce representational capacity. Actual latency also depends on hardware support, memory access and operator implementation, not only parameter count.
Examples
- MobileNet architecture
- Efficient CNNs
- Mobile deployment
Autovalutazione
Tre domande per verificare la comprensione della voce. Le spiegazioni restano disponibili anche senza JavaScript.