This is a tiny experiment to visualize the activations of each unit of a neural network based image classifier as a graphical plot.

The image classifier in this experiment is based on a deep neural network that has 3 hidden layers with 10 units each and a single output layer. The hidden layers use the ReLU activation function and the output layer uses the sigmoid activation function.

Although the model trained in this experiment works with about 80% accuracy, that's not the primary concern of this experiment. The primary concern of this experiment is to visualize the activations in each unit of a trained model.

The development steps here are written for a Linux or Mac system. All steps mentioned below assume that Python 3 is installed and you are at the top-level directory of this project.

1. Enter the following command to create a Python 3 virtual environment with numpy, matplotlib and h5py.

   make venv
   

2. Enter the following command to enter the virtual environment.

   . venv
   

3. Enter the following command to train a model, test it and write the model to a file named model.json.

   ./model.py
   

   To alter the learning parameters, look for the train() function in this file, edit the values of iterations and alpha variables and run this script again.

4. Classify arbitrary 64x64 PNG images in the extra-set directory with the following command. You can copy any image into this directory as long as it is a 64x64 PNG and run the following command.

   ./classify.py
   

5. To generate graphical plots of the learned model, enter the following command.

   ./plotmodel.py
   

   This generates activation plots for each unit in the neural network. This is explained further in the next section.

Here are the graphical plots of the activations in each unit in each layer for every pixel component (i.e. R, G and B components). Each image is a visualization of what the activations in a specific unit looks like. For example, the first image for layer 1 is the visualization of the activations of the first unit in the first hidden layer.

Each pixel in an image below represents the activation in a specific unit for the corresponding pixel in the input image. The activation for each component (red, green and blue) for each pixel in each unit is computed separately. Then the activations of red, green and blue components in each pixel is combined and shown as a single pixel in an image below.

Out of 50 test samples, 41 were correctly classified.

The test accuracy is: 82.00%.

Alter the learning parameters in iterations and alpha variables in train() and backward functions, respectively, of model.py to alter the test accuracy.

The training images and test images are present in train-set and test-set directories.

The training and test data were obtained from a few HDF5 files shared by Andrew Ng. The original H5 files are present in the h5data directory.

The script h5toimg.py converts this data to separate PNG image files and writes them to train-set and test-set directories.