MNIST Classifier in TensorFlow

Just a simple neural network MNIST classifier I built to teach myself. Nothing fancy! Everything is done in TensorFlow, and matplotlib is used to check out the losses periodically.

Consistently gets 98% accuracy or more on the test set with only a couple minutes of training on mycpu (and approximately 30 seconds on GPU).

Model Summary

• 3 layers:

  • 2 dense (512 and 256 units, relu activation) and one dropout (with rate of 0.2)

• Loss Function: softmax cross entropy

• Initializer: Xavier

• Optimizer: Adam (rate 0.001)

• Batch Size: 200

Requirements:

• TensorFlow (tested with 1.12.0, and 1.15.0)
• Numpy
• Matplotlib (visualizes the losses periodically during training)