PyTorch implementation of HMAX

PyTorch implementation of the HMAX model that closely follows that of the MATLAB implementation of The Laboratory for Computational Cognitive Neuroscience:

http://maxlab.neuro.georgetown.edu/hmax.html

The S and C units of the HMAX model can almost be mapped directly onto TorchVision's Conv2d and MaxPool2d layers, where channels are used to store the filters for different orientations. However, HMAX also implements multiple scales, which doesn't map nicely onto the existing TorchVision functionality. Therefore, each scale has its own Conv2d layer, which are executed in parallel.

Here is a schematic overview of the network architecture:

layers consisting of units with increasing scale
S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 S1 S1
  /    /    /    /    /    /    /    /
  C1    C1    C1    C1    C1    C1    C1    C1
                 |     /     /     /     /
           ALL-TO-ALL CONNECTIVITY
   /     /     /    |                    
  S2    S2    S2    S2    S2    S2    S2    S2
   |     |     |     |     |     |     |     |
  C2    C2    C2    C2    C2    C2    C2    C2

Installation

This script depends on the NumPy, SciPy, PyTorch and TorchVision packages.

Clone the repository somewhere and run the example.py script:

git clone https://github.com/wmvanvliet/pytorch_hmax
python example.py

Usage

See the example.py script on how to run the model on 10 example images.