Yifan Wang, Yueru Chen
[email protected], [email protected] last update 2019.09.25

From arXiv:1909.08190. There are well packed Pixelhop unit and LAG unit with simple usage. It uses Saab (arXiv:1810.02786) inside it, part of the Saab code is modified from https://github.com/davidsonic/Interpretable_CNN.

Run the code in Python3 in ./src folder.

Pixelhop Unit:

x -> Input, 4-D tensor (N,H,W,D), the same as channel_last mode in Keras.

dilate -> Controls location of chooesn neghbour pixels.

pad -> Padding, support none, reflect, zeros (default: reflect)

num_AC_kernels -> Number of AC components to be kept.

weight_name -> Saab kernel file location to be saved or loaded. (default: ../weight/+weight_name)

getK -> If use input to compute Saab kernel. (default: True)

useDC -> If add DC component. (default: False)

x1 = PixelHop_Unit(x, dilate=1, pad='reflect', num_AC_kernels=9, weight_name='pixelhop1.pkl', getK=True, useDC=False)

LAG Unit:

x -> Input data matrix, 2-D tensor (N,D)

train_labels -> class labels of each training sample

class_list -> list of object classes

SAVE -> store parameters

num_clusters -> output feature dimension (default: 50)

alpha -> A parameter to determine the relationship between the Euclidean distance and the likelihood for a sample belonging to a cluste (default: 5)

Train -> True: training stage; False: testing stage (default: True)

x1=LAG_Unit(x,train_labels=train_labels, class_list=class_list,SAVE=SAVE,num_clusters=50,alpha=5,Train=True)

One example is shown in ./src/example.py.