MobileNet is a family of efficient convolutional neural networks designed for mobile and edge deployment — using depthwise separable convolutions and width multipliers to dramatically reduce parameters and computation while maintaining competitive accuracy for vision tasks.

What Is MobileNet?

• Definition: Lightweight CNN architecture for efficient inference.
• Key Innovation: Depthwise separable convolutions.
• Goal: Deploy vision models on mobile/edge devices.
• Versions: MobileNetV1, V2, V3 (progressive improvements).

Why MobileNet

• Size: 10-20× smaller than VGG/ResNet.
• Speed: Real-time inference on mobile CPUs.
• Accuracy: Competitive with much larger models.
• Flexibility: Width/resolution multipliers for tuning.

Depthwise Separable Convolutions

Standard Convolution:

Input: H × W × C_in
Kernel: K × K × C_in × C_out
Output: H × W × C_out

Computation: H × W × K² × C_in × C_out

Depthwise Separable (MobileNet):

Step 1: Depthwise (spatial filtering per channel)
  Input: H × W × C_in
  Kernels: K × K × 1 (one per channel)
  Output: H × W × C_in
  Computation: H × W × K² × C_in

Step 2: Pointwise (1×1 convolution)
  Input: H × W × C_in
  Kernel: 1 × 1 × C_in × C_out
  Output: H × W × C_out
  Computation: H × W × C_in × C_out

Total: H × W × (K² + C_out) × C_in
Savings: ~K² (typically 8-9×)

Visual:

<svg viewBox="0 0 502 245" xmlns="http://www.w3.org/2000/svg" style="max-width:100%;height:auto" role="img"><rect x="0" y="0" width="502" height="245" rx="12" fill="#0d1117"/><g font-family="ui-monospace,SFMono-Regular,Menlo,Consolas,&quot;Liberation Mono&quot;,monospace" font-size="14"><text xml:space="preserve" x="20" y="31.7"><tspan fill="#c9d1d9">Standard Conv:</tspan></text><text xml:space="preserve" x="20" y="50.7"><tspan fill="#6e7681">┌─────────┐</tspan><tspan fill="#c9d1d9">     </tspan><tspan fill="#6e7681">┌─────────┐</tspan></text><text xml:space="preserve" x="20" y="69.7"><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9">  Input  </tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> </tspan><tspan fill="#6e7681">→</tspan><tspan fill="#c9d1d9"> K×K×C_in ×C_out </tspan><tspan fill="#6e7681">→</tspan><tspan fill="#c9d1d9"> </tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> Output </tspan><tspan fill="#6e7681">│</tspan></text><text xml:space="preserve" x="20" y="88.7"><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> H×W×C_in</tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9">                </tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> H×W×C_out</tspan><tspan fill="#6e7681">│</tspan></text><text xml:space="preserve" x="20" y="107.7"><tspan fill="#6e7681">└─────────┘</tspan><tspan fill="#c9d1d9">                </tspan><tspan fill="#6e7681">└─────────┘</tspan></text><text xml:space="preserve" x="20" y="126.7"></text><text xml:space="preserve" x="20" y="145.7"><tspan fill="#c9d1d9">Depthwise Separable:</tspan></text><text xml:space="preserve" x="20" y="164.7"><tspan fill="#6e7681">┌─────────┐</tspan><tspan fill="#c9d1d9">   K×K×1    </tspan><tspan fill="#6e7681">┌─────────┐</tspan><tspan fill="#c9d1d9">   1×1   </tspan><tspan fill="#6e7681">┌─────────┐</tspan></text><text xml:space="preserve" x="20" y="183.7"><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9">  Input  </tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> </tspan><tspan fill="#6e7681">→</tspan><tspan fill="#c9d1d9"> per ch </tspan><tspan fill="#6e7681">→</tspan><tspan fill="#c9d1d9"> </tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> H×W×C_in</tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> </tspan><tspan fill="#6e7681">→</tspan><tspan fill="#c9d1d9"> conv </tspan><tspan fill="#6e7681">→│</tspan><tspan fill="#c9d1d9"> H×W×C_out</tspan><tspan fill="#6e7681">│</tspan></text><text xml:space="preserve" x="20" y="202.7"><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9"> H×W×C_in</tspan><tspan fill="#6e7681">│</tspan><tspan fill="#c9d1d9">            </tspan><tspan fill="#6e7681">└─────────┘</tspan><tspan fill="#c9d1d9">         </tspan><tspan fill="#6e7681">└─────────┘</tspan></text><text xml:space="preserve" x="20" y="221.7"><tspan fill="#6e7681">└─────────┘</tspan><tspan fill="#c9d1d9">           Depthwise         Pointwise</tspan></text></g></svg>

MobileNet Versions

V1 (2017):

# Core block
class MobileNetV1Block(nn.Module):
    def __init__(self, in_ch, out_ch, stride=1):
        super().__init__()
        self.depthwise = nn.Conv2d(in_ch, in_ch, 3, stride, 1, groups=in_ch)
        self.bn1 = nn.BatchNorm2d(in_ch)
        self.pointwise = nn.Conv2d(in_ch, out_ch, 1)
        self.bn2 = nn.BatchNorm2d(out_ch)
        self.relu = nn.ReLU6(inplace=True)
    
    def forward(self, x):
        x = self.relu(self.bn1(self.depthwise(x)))
        x = self.relu(self.bn2(self.pointwise(x)))
        return x

V2 (2018) - Inverted Residuals:

# Inverted residual with linear bottleneck
class InvertedResidual(nn.Module):
    def __init__(self, in_ch, out_ch, stride, expand_ratio):
        super().__init__()
        hidden_dim = in_ch * expand_ratio
        self.use_residual = stride == 1 and in_ch == out_ch
        
        layers = []
        if expand_ratio != 1:
            # Expansion
            layers.append(nn.Conv2d(in_ch, hidden_dim, 1))
            layers.append(nn.BatchNorm2d(hidden_dim))
            layers.append(nn.ReLU6())
        
        # Depthwise
        layers.append(nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim))
        layers.append(nn.BatchNorm2d(hidden_dim))
        layers.append(nn.ReLU6())
        
        # Projection (linear, no activation)
        layers.append(nn.Conv2d(hidden_dim, out_ch, 1))
        layers.append(nn.BatchNorm2d(out_ch))
        
        self.conv = nn.Sequential(*layers)
    
    def forward(self, x):
        if self.use_residual:
            return x + self.conv(x)
        return self.conv(x)

V3 (2019) - Neural Architecture Search:

Improvements:
- NAS-discovered architecture
- Hard-swish activation
- Squeeze-and-excite attention
- Modified last layers

Width and Resolution Multipliers

Scaling Options:

Width multiplier (α): Scale channels
  Channels = base_channels × α
  α ∈ {0.25, 0.5, 0.75, 1.0}

Resolution multiplier (ρ): Scale input size
  Input = 224 × ρ
  ρ ∈ {0.57, 0.71, 0.86, 1.0} → {128, 160, 192, 224}

Trade-off: Smaller = faster but less accurate

Using MobileNet

PyTorch:

import torch
from torchvision.models import mobilenet_v3_small, mobilenet_v3_large

# Small version
model = mobilenet_v3_small(pretrained=True)

# Large version
model = mobilenet_v3_large(pretrained=True)

# Modify for custom classes
model.classifier[-1] = nn.Linear(1024, num_classes)

TensorFlow/Keras:

from tensorflow.keras.applications import MobileNetV3Small

model = MobileNetV3Small(
    input_shape=(224, 224, 3),
    include_top=True,
    weights="imagenet"
)

Performance Comparison

Model            | Params  | MACs   | Top-1 Acc
-----------------|---------|--------|----------
VGG-16           | 138M    | 15.5G  | 71.5%
ResNet-50        | 25M     | 4.1G   | 76.1%
MobileNetV1 1.0  | 4.2M    | 569M   | 70.6%
MobileNetV2 1.0  | 3.4M    | 300M   | 72.0%
MobileNetV3-Large| 5.4M    | 219M   | 75.2%
MobileNetV3-Small| 2.5M    | 66M    | 67.4%

MobileNet is the foundational efficient architecture for mobile AI — its depthwise separable convolution innovation enabled practical on-device computer vision and inspired subsequent efficient architectures like EfficientNet and MobileViT.