Skip to content

Latest commit

 

History

History

server

README.md

Python Server for HWD-Perceptron

Files and folders

The following table is sorted by importance.

Name Description
spnn.py Simple Perceptron Neural Network: This is the main workhorse of the whole project. A Python class called SimplePerceptronNN, that implements a freely configurable Perceptron based Neural Network; you can specify the amount of nodes for each layer. Besides the obligatory training and querying features, the class also features saving and loading of networks including metadata.
hwdr_server.py Flask based REST API server: It uses loads a previously trained network, receives requests via REST and outputs the recognized digit.
hwdt.py Training tool: Uses CSV based training data and runs multiple epochs to train handwritten digits recognition networks. Each epoch is saved as a separate file and metadata is added, so that later, after possibly long running training sessions, the best network version can be selected.
nist_csv.py Helper functions: Load and parse CSV files in our special format. Additionally an elegant ASCII art printing function is included.
load_test.py Test program that checks loads a neural network and tests, if it performs according to the metadata stored within the file. This tool is a good test for the whole Python and training stack because it actually tests, if everything works as expcted.
test_cp.py Tinkering tool: You can copy/paste the web browser console output of the client and test, if the recognition works as expected. See also the tinkering section in the main README.md.
training Folder that stores training data. There is a README.md there worth reading.
saved_nn Folder that contains networks, that I experimented with. Have a look at the training results to learn more about them and to choose, which one you might want to test. The folder also contains by default all networks that you generate with hwdt.py.
first_perceptron_test.py This was the development testbed I used while developing SimplePerceptronNN (spnn.py). It trains a network with 4 input nodes, 2 hidden layers with 3 nodes each and 2 output nodes to output a certain value pair given certain inputs.
second_perceptron_test.py After first_perceptron_test.py worked, I used this more advanced development testbed to check, if deeper networks are working (I used [4, 8, 5, 4, 2]) and if this "low node count but deep" network could learn an excercise, that is a bit more complex. I also used this to get a first idea about learning times and CPU power requirements.
second_perceptron_test.txt Results achieved with second_perceptron_test.py.

SimplePerceptronNN: Flexible Python Perceptron Class

The heart of this project is the flexible Python Perceptron class called SimplePerceptronNN and located in spnn.py. It can not only be used for learning to recognize handwritten digits, but it can be used for "anything".

SimplePerceptronNN allows a flexible layout of the network by specifying the input nodes, the hidden nodes and the output nodes within a simple Python list. The following pseudo-code is illustrating this:

import spnn

learning_rate = 0.1
nn = spnn.SimplePerceptronNN([784, 100, 10], learning_rate)

# you need a deeper network? try this:
# nn = spnn.SimplePerceptronNN([784, 2000, 500, 250, 100, 10], learning_rate)

# make sure that the length of the 'training' list equals the amount of input
# nodes and the length of 'desired_output' equals the amount of output nodes
for t, o in zip(training, desired_output):
    nn.train(t, o)

# network is now ready to answer questions:
output = nn.query(the_input)

Design Decisions

  • The network can be arbitrarily big and/or deep. But keep in mind that the code is CPU-only code (no GPU acceleration), so it quickly becomes pretty CPU and RAM hungry.

  • I decided not to use any bias, because after all, this class is called Simple Perceptron Neural Network. And it works pretty well - even without a bias.

  • For a similar reason, I am using the Sigmoid Function as activation function: I wanted to work through the math behind creating the backpropagation algorithm by myself (which includes calculating partial derivatives of it to obtain the gradient and then simplifying and transforming it all to be doable in pure matrix multiplications) and it seemed to me, that the Signmoid Function is "kind of easy but still hard enough" for me personally and my current state of learning to actually be able to do it. I know, that other functions would work better.