numpy-neuralnet-exercise

All codes and slides are based on the online book neuralnetworkanddeeplearning.com.

numpy-only-implementation

From example1.py to example8.py is implemented via only numpy and use the same architecture of a simple network called multilayer perceptrons (MLP) with one hidden layer.

example1.py : Quadratic Loss, SGD, Sigmoid (BASE LINE)

n is the number of unit in a hidden layer in following results.

example2.py : BASE LINE + Cross Entropy Loss

n=30	n=100

example3.py : BASE LINE + Cross Entropy Loss + L2 regularization

n=30	n=100

example4.py : BASE LINE + Cross Entropy Loss + L1 regularization

n=30	n=100

example5.py : BASE LINE + Cross Entropy Loss + Droput

n=30	n=100

example6.py : BASE LINE + Cross Entropy Loss + Xavier Initializer

n=30	n=100

example7.py : BASE LINE + Cross Entropy Loss + Momentum based SGD

n=30	n=100

example8.py : BASE LINE + Cross Entropy Loss + Xavier Initializer + ReLU

n=30	n=100

lauchers for other resources of numpy-only-implementation

There are also good resources for numpy-only-implementation and laucher for each recourse is provided.

Resource	Launcher
neuralnetworkanddeeplearning.com	launcher_package1.py
Stanford CS231 lectures	launcher_package2.py

simple tensoflow code for CNN

Code in tf_code_mnist folder is for CNN implmentation.
ch6_summary.pdf is related slide.

Command	Description	MNIST acc.
`train --model v0`	model v0 : BASE LINE + Softmax Layer + Cross Entropy Loss	97.80%
`train --model v1`	model v1 : model v0 + 1 Convolutional/Pooling Layers	98.78%
`train --model v2`	model v2 : model v1 + 1 Convolutional/Pooling Layers	99.06%
`train --model v3`	model v3 : model v2 + ReLU	99.23%
`train --model v4`	model v4 : model v3 + Data Augmentation	99.37%
`train --model v5`	model v5 : model v4 + 1 Fully-Connected Layer	99.43%
`train --model v6`	model v6 : model v5 + Dropout	99.60%