From 3079463bc0ca4f880c3c6c86e919e6f885320205 Mon Sep 17 00:00:00 2001
From: Luc Giffon <luc.giffon@lis-lab.fr>
Date: Fri, 12 Jul 2019 15:50:35 +0200
Subject: [PATCH] add documentation to python files
---
keras_kernel_functions.py | 28 +++++++++++++++++++++++++---
nystrom_layer.py | 21 +++++++++++++++++++--
2 files changed, 44 insertions(+), 5 deletions(-)
diff --git a/keras_kernel_functions.py b/keras_kernel_functions.py
index 9ebd4cb..b967df1 100644
--- a/keras_kernel_functions.py
+++ b/keras_kernel_functions.py
@@ -19,7 +19,7 @@ def keras_linear_kernel(args, normalize=True, tanh_activation=False):
:param args: list of size 2 containing x and y
:param normalize: if True, normalize the input with l2 before computing the kernel function
:param tanh_activation: if True apply tanh activation to the output
- :return:
+ :return: The linear kernel between args[0] and args[1]
"""
X = args[0]
Y = args[1]
@@ -35,6 +35,17 @@ def keras_linear_kernel(args, normalize=True, tanh_activation=False):
def keras_chi_square_CPD(args, epsilon=None, tanh_activation=True, normalize=False):
+ """
+ Chi square kernel (equivalent to `additive_chi2_kernel` in scikit-learn):
+
+ $k(x, y) = -Sum [(x - y)^2 / (x + y)]$
+
+ :param args: list of size 2 containing x and y
+ :param epsilon: very small value to add to the denominator so that we do not have zeros here
+ :param tanh_activation: if True apply tanh activation to the output
+ :param normalize: if True, normalize the input with l2 before computing the kernel function
+ :return: The chi square kernel between args[0] and args[1]
+ """
X = args[0]
Y = args[1]
if normalize:
@@ -59,6 +70,17 @@ def keras_chi_square_CPD(args, epsilon=None, tanh_activation=True, normalize=Fal
def keras_chi_square_CPD_exp(args, gamma, epsilon=None, tanh_activation=False, normalize=True):
+ """
+ Exponential chi square kernel (equivalent to `chi2_kernel` in scikit-learn):
+
+ $k(x, y) = exp(-gamma Sum [(x - y)^2 / (x + y)])$
+
+ :param args: list of size 2 containing x and y
+ :param epsilon: very small value to add to the denominator so that we do not have zeros here
+ :param tanh_activation: if True apply tanh activation to the output
+ :param normalize: if True, normalize the input with l2 before computing the kernel function
+ :return: The exponential chi square kernel between args[0] and args[1]
+ """
result = keras_chi_square_CPD(args, epsilon, tanh_activation, normalize)
result *= gamma
return K.exp(result)
@@ -68,12 +90,12 @@ def keras_rbf_kernel(args, gamma, normalize=True, tanh_activation=False):
"""
Compute the rbf kernel between each entry of X and each line of Y.
- tf_rbf_kernel(x, y, gamma) = exp(- (||x - y||^2 * gamma))
+ $(x, y, gamma) = exp(- (||x - y||^2 * gamma))$
:param X: A tensor of size n times d
:param Y: A tensor of size m times d
:param gamma: The bandwith of the kernel
- :return:
+ :return: The RBF kernel between args[0] and args[1]
"""
X = args[0]
Y = args[1]
diff --git a/nystrom_layer.py b/nystrom_layer.py
index ebeb7f5..a610db6 100644
--- a/nystrom_layer.py
+++ b/nystrom_layer.py
@@ -10,7 +10,18 @@ from keras.preprocessing.image import ImageDataGenerator
from keras_kernel_functions import keras_linear_kernel
-def datagen_fixed_batch_size(x, y, x_sub=None, p_datagen=ImageDataGenerator()):
+def datagen_fixed_batch_size(x, y, batch_size, x_sub=None, p_datagen=ImageDataGenerator()):
+ """
+ Wrap a data generator so that:
+ - it always output batches of the same size
+ - it gives a subsample along with each batch
+
+ :param x: observation data
+ :param y: label data
+ :param x_sub: list of base of subsample (each base must be of size batch_size)
+ :param p_datagen: the initial data generator to wrap
+ :return:
+ """
if x_sub is None:
x_sub = []
for x_batch, y_batch in p_datagen.flow(x, y, batch_size=batch_size):
@@ -59,6 +70,8 @@ def init_number_subsample_bases(nys_size, batch_size):
return quotient + 1, batch_size - remaining
if __name__ == "__main__":
+ # model meta parameters
+ # ---------------------
batch_size = 128
epochs = 1
num_classes = 10
@@ -126,11 +139,15 @@ if __name__ == "__main__":
# weight matrix of the nystrom layer
input_classifier = Dense(nys_size, use_bias=False, activation='linear')(kernel_vector) # metric matrix of the Nyström layer
+ # final softmax classification layer
+ # ----------------------------------
classif = Dense(num_classes, activation="softmax")(input_classifier)
+ # finalization of model, compilation and training
+ # -----------------------------------------------
model = Model([input_x] + input_repr_subsample, [classif])
adam = Adam(lr=.1)
model.compile(loss='categorical_crossentropy', optimizer=adam, metrics=['accuracy'])
- model.fit_generator(datagen_fixed_batch_size(x_train, y_train, list_subsample_bases, datagen),
+ model.fit_generator(datagen_fixed_batch_size(x_train, y_train, batch_size, list_subsample_bases, datagen),
steps_per_epoch=int(x_train.shape[0] / batch_size),
epochs=epochs)
--
GitLab