Visualizing VGG-16

Visualizing the weights and feature/activation maps of a pretrained VGG-16

Dependencies

refer requirements.txt

Implementation

1. For visualizing weights, I started with this [1].

• The layerwise weights are visualized here. Weights at first Convolution layer look like this

• The numpy arrays were also saved and plotted as stacks of histograms. The three channels at the first layer produce the following histogram.

2. For Visualizing activations/features, I used the guided backprop method [2].

• The feature map at layer 1 and filter 0 on an example image looks

• Here the histograms produced were layerwise, to compare how a filter is activated across different layers.

3. Also tried deep dream on an image, which looks very interesting!

References

[1] https://github.com/pedrodiamel/nettutorial/blob/master/pytorch/pytorch_visualization.ipynb

[2] https://github.com/utkuozbulak/pytorch-cnn-visualizations#convolutional-neural-network-filter-visualization