LucianoPAlmeida / mnist-number-classification

A simple classification of mnist handwritten digits

mnist-number-classification

This is a Convolutional Neural Network model trained with the MNIST Dataset of handwritten digits and this model was implemented to support the CoreML-MNIST Demo Application.

Tools

• Tensorflow 1.8
• tf-coreml
• coremltools
• Jupyter Notebook
• Python 3.6
• matplotlib

The Model

The model was trainned with 70 epochs with a batch size of 512. Achieving 0.984400 of validation accuracy and 0.9861225328947368 of test accuracy. The AdamOptimizer was used to train this network with a learning rate of 0.00001.

Architecture

• conv2d with filter size 32, strides 5, padding same and relu activation
• max_pooling2d with pool size of 2 and strides 2
• conv2d with filter size 64, strides 5, padding same and relu activation
• max_pooling2d with pool size of 2 and strides 2
• fully_connected with number of outputs 1024 and relu activation
• fully_connected with number of outputs 10 and no activation function
• softmax activation layer

The CoreML Model

With a trained model and saved .pb file, tf-coreml was used generate a CoreML model. The code is available on coreml_converter.py

Credits and Thanks

• Udacity for their awesome Deep Learning Nanodegree Foundation course.

Licence

mnist-number-classification is released under the MIT License.