This is a demo code to learn how to use neural networks for image recognition. It uses the MNIST dataset of handwritten digits. Technological we use Python, Tensorflow, Keras and Jupyter Notebooks to do the job.
- 01-MNIST-dense.ipynb is the first example. It uses a fully connected neural network.
- 02-MNIST-conv.ipynb is the second example. It is based on the first example, but uses a convolutional neural network. Here you can examine the improvement of accuracy, although the network has fewer params.
If you want to learn more about this topic, I very much recommend the book Deep Learning with Python from Francois Chollet.
- install Conda (Python 3.7): https://docs.conda.io/en/latest/miniconda.html
- make sure to have direct internet access or configure a web-proxy (for git and conda) - see web-proxy with conda below
- check out this project:
git clone https://github.com/t-systems-on-site-services-gmbh/nn-demo-keras.git - change to the
nn-demo-kerasproject directory - create new conda environment with python version 3.6:
conda create --name nn-demo python=3.6 - activate the new conda environment:
conda activate nn-demo - install pip packages:
pip install -r pip-packages.txt - install tool to use conda with jupyter:
conda install nb_conda_kernels - Start jupyter notebook:
jupyter notebook - load the first notebook and execute it
To use a web-proxy with conda, you have to create a .condarc with the following content:
proxy_servers:
http: http://user:[email protected]:8080
https: https://user:[email protected]:8080
ssl_verify: False