zfnet

Implementing ZFNet

In this paper implementation, we are going to implement ZFNet from the paper Visualizing and Understanding Convolutional Networks by Matthew D Zeiler, Rob Fergus

This paper utilizes Deconvolution which also introduced in the paper Deconvolutional Network (deconvnet) (Zeiler et al., 2011)

This produces visualizations per layer of the activations produced by specific Convolution layers on a particular Neural Network

Also, I've used AlexNet (Krizhevsky et al., 2012) again here since the original paper uses it as their architecture of choice.

Implementation Details

Python Libraries

• Pytorch
• Torchvision, Torchtext (For download utilities)
• TorchInfo (for summarizing models)
• Numpy
• Matplotlib
• Seaborn

Result of Implementation

Visualization of Activations for Low-Loss Predictions

Visualization of Activations for High-Loss Predictions

Hardware Used

GPU : NVIDIA RTX 3060

Dataset Used

• Kaggle Cats and Dogs Dataset

References

• Visualizing and Understanding Convolutional Networks (Original Paper)