Neuron Feature (Pytorch)

This repo provides a Pytorch implementation of the Neuron Feature (NF) CNN filter visualization method introduced in Color encoding in biologically-inspired convolutional neural networks. For a given Pytorch model, it generates a weighted average of the highest activation input patches for each filter in the final layer.

The code is compatible with Group Equivariant Convolutions in which case a separate NF is generated for each group element.

Uses pytorch-receptive-field for auto calculating the receptive field of the final layer. Tested on pytorch 1.3.0 and torchvision 0.4.1.

Usage

1. git clone https://github.com/Attila94/neuron-feature

2. Run example.py.

This generates the neuron features for VGG16 on the CIFAR100 test set.

Example outputs for filters 20 and 33:

Note: The generated NFs contain input padding (black borders), required due to the large receptive field of VGG16 and small resolution of CIFAR100. A higher resolution dataset is better suitable for generating NFs.

Known limitations

• Receptive field calculation for residual connections not implemented / tested

Please report all other issues.