Visualizing CNN features

Description

Simple iTorch example for visualizing CNN features as described in "Visualizing and Understanding Convolutional Networks" by Matthew Zeiler and Rob Fergus.

The implementation is quite quick-and-dirty: it seems to work, but I can't exclude the presence of methodological mistakes which by luck don't prevent the implementation to show something apparently meaningful.

Requirements

• Torch 7
• nn package
• image package
• iTorch

Resources needed

To run the example without modifications, you will need the OverFeat Torch wrapper from jhjin/overfeat-torch.

In short, you will need to:

• Clone the repository.
• Follow the instruction to download the weights files, and put the net_weight_0 file in the models directory (you can ignore the other files in the tgz file).
• Compile the ParamBank.c file through the provided Makefile.
• Put ParamBank.lua and the compiled libParamBank.so in the same directory as my visualize_features_example.ipynb notebook.

Also you will need to bee.jpg test image from OverFeat repository and put it into the repository root directory. However, it is straightforward to edit the notebook to choose any other picture you'd like.

Compatiblity with other networks

You should also be able to use any other model, provided that:

• Layers are wrapped in a nn.Sequential container.
• It only has nn.SpatialConvolutionMM, nn.SpatialMaxPooling, nn.ReLU layers.
• Convolutional layers must be proper convolutions; i.e. not 1x1 convolutions for fully-connected layers.

However, if you look at the code, you'll see easily how to adapt it for other layers. I just don't need them now.