simple-MNIST-MLP

ATPESC 2021

2 layer MLP for classifying MNIST using different levels of manual implementation and TensorFlow, Keras API.

• mnist_tf_variable_v1.py: network and layer definitions bassed on tf.Variable, autodiff via tf.GradientTape, manual implementation of loss function and SGD, no Keras used beyond loading MNIST dataset.
• mnist_tf_variable_v2.py: same as v1, but variables collected in a tf.Module subclass. Should update tape.gradient() call to use model.trainable_variables and output a single dictionary of wgrads.
• mnist_tf_variable_v3.py: same as v1, but using Keras tf.keras.optimizers and loss function.
• mnist_keras.py: using Keras for everything, including tf.keras.Sequential for the model. No use of tf.Variable or tf.Module