mathdatech/Python
1
0
Fork 0
mirror of https://github.com/TheAlgorithms/Python.git synced 2025-01-31 06:33:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update back_propagation_neural_network.py (#1342)

Browse Source 
* Update back_propagation_neural_network.py

Added comments below functions

* Update back_propagation_neural_network.py

Co-authored-by: Christian Clauss <cclauss@me.com>
This commit is contained in:
Yash Bhardwaj 2019-12-26 08:44:47 +05:30 committed by Christian Clauss
parent 725834b9bc
commit 1b3985837f
1 changed files with 6 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
neural_network/back_propagation_neural_network.py
View File

@ -31,7 +31,6 @@ class DenseLayer:
""" """
Layers of BP neural network Layers of BP neural network
""" """
def __init__( def __init__(
self, units, activation=None, learning_rate=None, is_input_layer=False self, units, activation=None, learning_rate=None, is_input_layer=False
): ):
@ -58,6 +57,7 @@ class DenseLayer:
self.activation = sigmoid self.activation = sigmoid
def cal_gradient(self): def cal_gradient(self):
# activation function may be sigmoid or linear
if self.activation == sigmoid: if self.activation == sigmoid:
gradient_mat = np.dot(self.output, (1 - self.output).T) gradient_mat = np.dot(self.output, (1 - self.output).T)
gradient_activation = np.diag(np.diag(gradient_mat)) gradient_activation = np.diag(np.diag(gradient_mat))
@ -78,7 +78,6 @@ class DenseLayer:
return self.output return self.output
def back_propagation(self, gradient): def back_propagation(self, gradient):
gradient_activation = self.cal_gradient() # i * i 维 gradient_activation = self.cal_gradient() # i * i 维
gradient = np.asmatrix(np.dot(gradient.T, gradient_activation)) gradient = np.asmatrix(np.dot(gradient.T, gradient_activation))
@ -89,11 +88,10 @@ class DenseLayer:
self.gradient_weight = np.dot(gradient.T, self._gradient_weight.T) self.gradient_weight = np.dot(gradient.T, self._gradient_weight.T)
self.gradient_bias = gradient * self._gradient_bias self.gradient_bias = gradient * self._gradient_bias
self.gradient = np.dot(gradient, self._gradient_x).T self.gradient = np.dot(gradient, self._gradient_x).T
# ----------------------upgrade # upgrade: the Negative gradient direction
# -----------the Negative gradient direction --------
self.weight = self.weight - self.learn_rate * self.gradient_weight self.weight = self.weight - self.learn_rate * self.gradient_weight
self.bias = self.bias - self.learn_rate * self.gradient_bias.T self.bias = self.bias - self.learn_rate * self.gradient_bias.T
# updates the weights and bias according to learning rate (0.3 if undefined)
return self.gradient return self.gradient
@ -101,7 +99,6 @@ class BPNN:
""" """
Back Propagation Neural Network model Back Propagation Neural Network model
""" """
def __init__(self): def __init__(self):
self.layers = [] self.layers = []
self.train_mse = [] self.train_mse = []
@ -144,8 +141,7 @@ class BPNN:
loss, gradient = self.cal_loss(_ydata, _xdata) loss, gradient = self.cal_loss(_ydata, _xdata)
all_loss = all_loss + loss all_loss = all_loss + loss
# back propagation # back propagation: the input_layer does not upgrade
# the input_layer does not upgrade
for layer in self.layers[:0:-1]: for layer in self.layers[:0:-1]:
gradient = layer.back_propagation(gradient) gradient = layer.back_propagation(gradient)
@ -176,7 +172,6 @@ class BPNN:
def example(): def example():
x = np.random.randn(10, 10) x = np.random.randn(10, 10)
y = np.asarray( y = np.asarray(
[ [
@ -192,17 +187,11 @@ def example():
[0.1, 0.5], [0.1, 0.5],
] ]
) )
model = BPNN() model = BPNN()
model.add_layer(DenseLayer(10)) for i in (10, 20, 30, 2):
model.add_layer(DenseLayer(20)) model.add_layer(DenseLayer(i))
model.add_layer(DenseLayer(30))
model.add_layer(DenseLayer(2))
model.build() model.build()
model.summary() model.summary()
model.train(xdata=x, ydata=y, train_round=100, accuracy=0.01) model.train(xdata=x, ydata=y, train_round=100, accuracy=0.01)