This is a C# implementation of a handwritten digit classifier neuron network in Michael Nielsen's Neural Networks and Deep Learning book.
Currently, the trained network has about 95% accuracy, which is to be improved in future iterations.
This project is written with .Net Framework 4.7 using Visual Studio 2017.
- Download the MNIST data set. Save the 4 archive files (train-images-idx3-ubyte.gz, train-labels-idx1-ubyte.gz, t10k-images-idx3-ubyte.gz, t10k-labels-idx1-ubyte.gz) to a folder of your choice.
- Modify the first few lines of code in Program.cs to point to correct data package files.
- Compile and run. The code trains on 60,000 images from the training set, and then allows you to classify an individual image from the 10,000-image test set by entering an image id (0-9,999). Or, you can enter '-1' to run against the entire test set. Enter '-2' to exit. Enter '-3' to save the trained model to a text file. Later on, you can launch the program with the file name as a parameter to reload the trained model.
There are few trained models under the trained folder:
| File | Accuracy | Cost Function | Learning Rate | Regularization | Epoch | Hidden Layer Size |
|---|---|---|---|---|---|---|
| quadratic-9640.txt | 96.4% | quadratic | 3.0 | 0.0 | 30 | 30 |
| crossentropy-9800.txt | 98.0% | cross-entropy | 0.5 | 0.0 | 30 | 100 |
| crossentropy-regulated-9810.txt | 98.1% | cross-entropy | 0.1 | 5.0 | 60 | 100 |
| crossentropy-regulated-9820.txt | 98.2% | cross-entropy | 0.1 | 5.0 | 120 | 100 |
To use a pre-trained model, pass the file name to DigitalClassifier as a parameter:
DigitalClassifier.exe /path/to/trained/model/file