Digits

Digit classification with tensorflow: https://d-costa.github.io/digits/

Features

• Optional centering: understand what areas are more relevant to identify digits;
• Probability for each digit;

Neural Net Architecture

The MNIST dataset is used for training a CNN network with the following architecture:

Layer	Output Shape	Param #
InputLayer	[(None, 28, 28, 1)]	0
Conv2D (32, (3, 3), relu)	(None, 26, 26, 32)	320
MaxPooling2D (2,2)	(None, 13, 13, 32)	0
Conv2D (16, (3, 3), relu)	(None, 11, 11, 16)	4624
MaxPooling2	(None, 5, 5, 16)	0
Flatten	(None, 400)	0
Dense (relu)	(None, 128)	51328
Dense (softmax)	(None, 10)	1290

• Total params: 57,562
• Trainable params: 57,562
• Non-trainable params: 0
• ~15.000 examples of the training set are used as validation
• Adam is used as the optimizer with 15 epochs and batch size of 64.

Test loss: 0.050
Test accuracy: 98.97%

Development

1. Install the required python packages:

   pip install -r requirements.txt

2. Modify the model architecture in train.py and train the model:

   python train.py

   The model is saved in a file named model.h5.

3. Test the trained model:

   python test.py

4. Convert the model to use it in tensorflowjs:

   ./convert.sh

   The model is saved in /public/tfjs

5. Open index.html in your browser

Acknowledgements

• Carlos Aguayo : https://github.com/carlos-aguayo/carlos-aguayo.github.io
• CNN architecture render: https://netron.app/