CNN for Digit Recognition

Experiment with different pre-processing methods to learn the performance of the CNN model.

Code style

https://www.python.org/dev/peps/pep-0008/

Packages

Important packages:

• tensorflow==2.1.0
• numpy==1.19.3
• h5py==2.10.0

Built with

• Python

Features

There are 3 methods provided (in this repo) to scale pixel values in images. Pixels size set for this model is 28x28. Please refer process_data.py to see data pre-processing flow.

Methods:

1. Normalization
2. Mean
3. Standardize

Note:

• For development environment only. Not suitable for production.
• Modify the input shape and type in data.py (if required).
• Modify the layers in load.py (if required).

Installation

Install Python packages.

Predict

The result is saved in h5.

python LeNet <method>

Training

Run modelling to create new model.

Note:

• There are 3 types of activation functions (ReLu, Tan, Sigmoid) tested (in this repo) and model with ReLu has the highest accuracy.

Testing

Include testing data in data.py file. Assign to

test_y

How to use?

Modelling

Find the best activation function.

Prediction

Use different pre-processing methods.

Credits

Dataset

http://yann.lecun.com/exdb/mnist/

License

https://github.com/ami-sm/cnn-models/blob/master/LICENSE

MIT © AMI-SM