ZFNet¶

ConvNet Image Classification

class lucid.models.ZFNet(num_classes: int = 1000)¶

The ZFNet module in lucid.nn implements the Zeiler and Fergus Net, an improvement over AlexNet with smaller convolutional filters and enhanced visualization techniques for understanding feature learning.

Class Signature¶

class ZFNet(nn.Module):
    def __init__(self, num_classes: int = 1000)

Parameters¶

num_classes (int, optional): The number of output classes for classification. Default is 1000.

Attributes¶

features (nn.Sequential): The convolutional layers, including pooling and ReLU activations.
avgpool (nn.AdaptiveAvgPool2d): Adaptive average pooling layer to reduce spatial dimensions to (6, 6).
classifier (nn.Sequential): The fully connected layers with dropout and ReLU activations for classification.

Architecture¶

The architecture of ZFNet is as follows:

Convolutional Layers: - 5 convolutional layers with ReLU activations. - Smaller kernel sizes (7x7 in the first layer) compared to AlexNet for better feature learning. - MaxPooling layers after the 1st, 2nd, and 5th convolutional layers.
Fully Connected Layers: - 2 hidden fully connected layers, each with 4096 units and ReLU activations. - Output layer with num_classes units for classification.
Regularization: - Dropout is applied to fully connected layers to reduce overfitting.

Examples¶

Basic Example

import lucid.nn as nn

# Create ZFNet with default 1000 classes
model = nn.ZFNet(num_classes=1000)

# Input tensor with shape (1, 3, 224, 224)
input_ = Tensor.randn(1, 3, 224, 224)

# Perform forward pass
output = model(input_)

print(output.shape)  # Shape: (1, 1000)

Explanation

The model processes the input through its convolutional and fully connected layers, producing logits for 1000 classes.

Custom Number of Classes

# Create ZFNet with custom 10 classes
model = nn.ZFNet(num_classes=10)

input_ = Tensor.randn(1, 3, 224, 224)

output = model(input_)
print(output.shape)  # Shape: (1, 10)