MNIST Classifier
This repository contains a simple convolutional neural-network classifier for the mnist-dataset.
- The algorithm uses 2 convolutional- and 2 max-pooling-layers with a final fully conncted layer.
- The network is trained using softmax-cross-entropy as loss function and an Adamoptimizer with a set learning rate of 0.001.
- It reaches an accuracy of 97.96% after 2 epochs of training (1 epoch being the full dataset)
Contact me if you have any questions or want to use the code.
Different stable-versions of the algorithm can be found in different commits,
e.g. git checkout 398690e checks out the first working version using an MLP instead of a CNN.
Go back to newest commit using git checkout master.
Prerequisites
- Python 3.5.2 or newer
Packages
All packages can be installed using pip
- tensorflow 1.3.0
- numpy 1.13.1
Run Locally
- Clone the repo
- Run
python3 main.py - The code will download the mnist-dataset if none is provided
Some hyperparameters can be set using parameters:
python3 main.py --batchsize 10
A list of all hyperparameters and their use can be found using:
python3 main.py --help