/deepLearning

MNIST, SVHN, Transfered Learning, DNN, CNN

Primary LanguageJupyter Notebook

Keras + Sklearn Wrapper for grid searching parameters

  • Task1 Iris+DNN
    highest accuracy is 0.97 achieved with 100 epoches (12 as first hidden layer size, and 6 as the second hidden layer size). For details, please see task1/Task1-Iris-DNN.ipynb

  • Task2 MINIST+DNN
    vanila baseline achieves 0.9657, and adding dropout achieves 0.9795 on a DNN with the following parameters(64 epochs, 64 as the first hidden layer size, 32 as the second hidden layer size, 128 batch size, and 0.1 dropout rate)
    For output details please see task2/Task2-MNIST-DNN.ipynb and task2/Task2-gpu.py, the learning curve plot can be found in task2/learning_curve_comparison.png

  • Task3 SVHN+CNN+bachnormalization
    the baseline model achives 0.893 accuracy; after bachnorm, achieves 0.952 accuracy on the test set. For details, please see task3/Task3-SVHN-CNN.py

  • Task4 VGG+Pets+transferredLearning
    The model achieves 0.755 test accuracy with a retrained MLP (32 epochs, 128 batch size, 32 hidden layer size). For details please see task4/task4.py and simply task4/task4-notebook.ipynb, which contains all output display.
    To execute task4/task4.py make sure you change path_to_pets = "../pets/" to let path_to_pets point to where pets folder sits on your computer.