mnist-hub 0.1.3

No project description provided

MNIST Hub

A collection of MNIST classifiers and tools including:

• Generic methods for training, saving and evaluating models
• Interactive GUI for drawing and recognition
• A mini contest mode for comparing different models

Installation

Have Python 3.10 of higher installed. Use Python for which numpy is already compiled which means 3.12 on most platform.

# Create a new conda environment
micromamba create -y -n mnist python=3.12

# Activate the environment
micromamba activate mnist

Install the package:

pip install mnist-hub --upgrade

To install in development mode, clone the source and in the mnist_hub directory type:

pip install -e .

The main runner

The first run will a long time a bit because it has to parse the entire scipy/numpy codebase.

mnist

then it will print somethinw like

Usage: mnist [OPTIONS] COMMAND [ARGS]...

  MNIST Hub - A collection of MNIST classifiers and tools.

Options:
  -h, --help  Show this message and exit.

Commands:
  train    Train a model and save the fitted model to a file
  eval     Load a model stored in a file and evaluate it
  gui      Launch the MNIST GUI application.
  contest  Run the MNIST contest evaluation.

Usage

The package provides a command line interface with several subcommands:

# Launch the GUI application that you can use
# to test the performance of the models.
mnist gui

# Train a custom model, on just 100 training samples and save to a file.
mnist train --name mnist.svm.Network --limit 100 --fname foo.gz

# Evaluate the saved model
mnist eval --fname foo.gz 

# Run the contest evaluation
mnist contest

Development

In development mode clone the archive and install in editable mode.

# Create environment
micromamba create -y -n mnist python

# Activate the environment
micromamba activate mnist

# Install in editable mode
pip install -e .

Training different models

Your can train any valid Python class that implements the train(trainig_data)and predict(data) methods.

Look at the minst.toy.Toy class for an example API.

• src/mnist/toy.py

You can train new models using the CLI:

# Train the new model
mnist train --name mnist.toy.Toy --fname toy_model.gz

# Evaluate the model in the file.
mnist eval --fname toy_model.gz

Interesting observations

Neural Network models are massively faster to evaluate than SVM models.

Evaluating a pre-trained Neural Network model on 10000 test data takes just 0.15 seconds, while an SVM model takes about 61 seconds for the same task.

The MNIST Contest

Can you beat a neural model? Try the contest mode:

mnist contest

Keep guessing!