/MNIST-MLP

Project to learn Multi Layer Perceptron model and Stochastic Gradient Descent with some heuristics as learning rate decay, momentum and Nesterov

Primary LanguagePythonMIT LicenseMIT

MNIST-MLP

Project to learn Multi Layer Perceptron model and Stochastic Gradient Descent with some heuristics as learning rate decay, momentum and Nesterov

From scratch:

Was created MLP with 2 hidden layers. Error metric: Mean Squared Error. Optimisation: SGD.

Could be found at https://github.com/RikiTikkiTavi/MNIST-MLP/tree/master/src/self_realisation

Using keras:

Was created MLP with 2 hidden layers. Error metric: Mean Squared Error. Optimisation: SGD with decay, momentum and Nesterov.

Parameters where tuned using GridSearch implementation from Talos library.

Accuracy 95.69% is achieved on test data.

Code could be found at https://github.com/RikiTikkiTavi/MNIST-MLP/tree/master/src/tf_realisation

Trained model: https://github.com/RikiTikkiTavi/MNIST-MLP/blob/master/src/tf_realisation/mnist_mlp_sgd_mse_model.h5