Neural Networks for MNIST digits starting with pure Python

The expected order to inspect/develop/test these examples is as given below.

The idea is to first develop some NN code that works in pure Python, and then continue to develop from there until we have a fairly comprehensive framework in pure Python supporting multiple layer types and a simple class-based declarative definition of the network.

On completion of the mnist_conv pure Python step of this development process we can define networks using our own classes such as:

net_conv = [
    Conv2D(3,1,8,{ "pad": 2, "stride": 2}),
    Pooling({"f": 3, "stride": 3, "mode": "max"}),
    ReLU(),
    Flatten(),
    Dense(200,10)
]

We can then discard our home-grown Python classes and move over to the tensorflow/keras API, now with a better comprehension of similarly defined networks such as:

model_lewis = Sequential()
model_lewis.add(Dense(256, input_shape=(784,), activation="relu"))
model_lewis.add(Dense(128, activation="relu"))
model_lewis.add(Dense(10, activation="softmax"))

MNIST Weisberg - simplest neural network from scratch in Python

Build a single fully-connected-layer network with numpy to recognize hand-written digits in the MNIST data set.

Readme and source files here.

Ref: Building a Neural Network from Scratch: Part 1

See source file in this repo mnist_weisberg.py

MNIST (Aayush) Agrawal - restructure NN code using Python classes for each layer type

Similar to Weisberg, but abstracting the layers into Python classes.

Readme and source files here.

There are copied and changed versions of this article but the reference below is preferred.

Ref: Building Neural Network from scratch

See source file in this repo mnist_agrawal.py

MNIST Conv - add classes for Convolution

Adding new layer classes for

• Convolution
• Pooling (max & average)
• Softmax

Readme and source files here.

MNIST Keras

MNIST digit recognition using keras/tensorflow API - intentionally a very similar layout to the pure Prolog class-based layers used in mnist-conv above, which were developed specifically to illustrate the similarity with what you would code using the keras API.

Readme and source files here.

Initial setup steps

Build a Linux server

username james hostname: james-dell5090 IP address 192.168.1.37 Netmask 255.255.255.0

adduser ijl20 / sudo

install openssh-server

check python

install python3-pip

sudo apt install python-venv

make git repo / push to github

copy .gitignore from this repo

make/copy requirements.txt - include "wheel"

Make virtual env in development directory python3 -m venv venv

python3 -m pip install pip --upgrade

python3 -m pip install -r requirements.txt

Jupyter use/config notes

run with $ jupyter notebook

Set new password with

$ jupyter notebook password

Check config (e..g password) with

$ atom ~/.jupyter/jupyter_notebook_config.json

numpy

matplotlib load image

display bar chart

display graph