A rudimentary Convolutional Neural Network in Python

This post consists of an implementation of a three layer Convolutional Neural Network in Python. Please note that this network is neither effiecient nor achieves state of the art classification performance. Since the purpose is solely educational I tried to keep it as simple as possible with very few dependencies.

Installation

Using Anaconda

• Anaconda - The Anaconda Python distribution

Without Anaconda

• Python
• Numpy
• Scipy
• scikit-learn
• matplotlib

Usage

First instantiate a model object and pass reg_lambda (regularization strength) and epsilon (learning rate) which default to 0.009 and 0.00001 respectively.

model = Model(reg_lamda=0.01, epsilon=0.00005)

Then call the fit method passing the training data and the number of epochs (20-30 seems like a good choice, depending on your machine).

model.fit(X_train, y_train, 30)

Testing your model

You can check the performance of your model after training:

model.test(X_test, y_test)

which uses the now trained parameters to predict labels for the test data X_test and compares them to the correct labels in y_test.

Save weights for later use

You can save the trained parameters as numpy arrays to /nets with

model.dumpParams()

and load previously saved nets with

model.fromParams(nets/nameOfYourNet)

which overrides the current weight matrices W1 and W2 of model.

Results

TODO: show results

Convolutions

TODO: write something about why this actually works

Credits

TODO: write credits