Slowest-Neural-Net-in-the-West

My custom neural net architecture written in Python to practice my ML skills.

Install

Make a new Python virtual environment (version >= 3.6). Install this project with pip. In terminal, run

pip3 install snnw

Download MNIST dataset

Open a new Python shell or Jupyter Notebook and run

import snnw

Set raw_dir = the directory where you want to save the MNIST raw data files. Run

snnw.dataset.mnist.download.raw(raw_dir)

Set png_dir = the directory where you want to save the MNIST .png images and path text files. Run

snnw.dataset.mnist.convert.raw_to_png(raw_dir, png_dir)

Set npy_dir = the directory where you want to save the MNIST .npy image and label numpy arrays. Run

snnw.dataset.mnist.convert.png_to_npy(png_dir, npy_dir)

Note: To use a custom dataset, please use the same format that the MNIST 'png' folder is in. Otherwise, you can feed in numpy arrays of similar form to the ones in the MNIST 'npy' folder.

Please let me know if you have any issues running your custom dataset, and I will be happy to fix it!

Get model config

Set config_path = the path to where you want to save the training/testing model's config file. Run

snnw.nn.config.get(config_path, 1) or snnw.nn.config.get(config_path, 2)

to write sample config file 1 or 2 to config_path, or write your own config file and place it where config_path points to.

If you write a custom config file, make sure that it follows the given format to prevent parsing errors!

Train model

Set model_dir = the directory where the trained model's weights and biases will be stored.

Set train_image_path = the path to where the .npy file for the training image arrays are located. This file should be located inside the npy_dir you specified earlier.

Set train_label_path = the path to where the .npy file for the training label arrays are located. This file should be located inside the npy_dir you specified earlier.

Set steps = the number of training steps you would like to train for. The default is 60,000.

Set learning_rate = the learning rate you would like to train with. The default is 5e-4.

Run snnw.run.train(model_dir, config_path, train_image_path, train_label_path, steps, learning_rate).

Note: if you get a NaN error or "not a probability array" error, then you probably have a vanishing or exploding gradient problem. To fix this, try adjusting the learning rate. The default learning rate and number of steps included have been tested to work with both included models.

Test model

Set test_image_path = the path to where the .npy file for the training image arrays are located. This file should be located inside the npy_dir you specified earlier.

Set test_label_path = the path to where the .npy file for the training label arrays are located. This file should be located inside the npy_dir you specified earlier.

Run snnw.run.evaluate(model_dir, config_path, test_image_path, test_label_path).