Python 神经网络：运行 10 次迭代，但我得到相同的输出

Question

我是编码新手，所以我遇到了一些简单的问题。当我运行 10 次迭代时，我得到相同的数字。-0.5 用于激活，0.0 用于输入，即使在底部我将节点列表中每个对应节点的起始激活设置为 1.0、1.0 和 0.0。

我想通过设置初始状态。他们向另一个节点发送一个输入：这是 sender.activation * 权重为 1。我应该得到一个新的输入值。然后将其应用于我的激活，然后我将能够 -0.5 并获得节点的新激活。

至少那是我尝试做的。不知何故，它只是吐出 0.0 和 -0.5。

# 
#                               Preparations 
# 

nodes=[] 
NUMNODES=3

# 
#                                   Defining Node Class
# 

class Node(object): 

    def __init__(self,name=None): 
        self.name=name 
        self.activation_threshold=1.0
        self.net_input=0.0
        self.outgoing_connections=[] 
        self.incoming_connections=[] 
        self.connections=[] 
        self.activation=None

    def addconnection(self,sender,weight=0.0):
        self.connections.append(Connection(self,sender,weight)) 

    def update_input(self): 
        self.net_input=0.0
        for conn in self.connections: 
            self.net_input += conn.weight * conn.sender.activation 
        print 'Updated Input is', self.net_input 

    def update_activation(self): 
        self.activation = self.net_input - 0.5
        print 'Updated Activation is', self.activation 

# 
#                                   Defining Connection Class
# 

class Connection(object): 
    def __init__(self, sender, reciever, weight=1.0): 
        self.weight=weight 
        self.sender=sender 
        self.reciever=reciever 
        sender.outgoing_connections.append(self) 
        reciever.incoming_connections.append(self) 
# 
#                                 Other Programs 
# 


def set_activations(act_vector): 
    """Activation vector must be same length as nodes list"""
    for i in xrange(len(act_vector)): 
        nodes[i].activation = act_vector[i] 


for i in xrange(NUMNODES): 
    nodes.append(Node()) 


for i in xrange(NUMNODES):#go thru all the nodes calling them i 
    for j in xrange(NUMNODES):#go thru all the nodes calling them j 
        if i!=j:#as long as i and j are not the same 
            nodes[i].addconnection(nodes[j])#connects the nodes together
#
#                                         Setting Activations
#
set_activations([1.0,1.0,0.0])

#
#                                        Running 10 Iterations
#

for i in xrange(10): 
    for thing in nodes: 
        thing.update_activation() 
        thing.update_input()

Answer 1

所以，你编码了

def addconnection(self,sender,weight=0.0):
    self.connections.append(Connection(self,sender,weight)) 
    print "Node", str(self), "now contains", str(self.connections[-1])

你用它来称呼它

nodes[i].addconnection(nodes[j]) #connects the nodes together

您没有在此处指定权重。因此，您可能会认为您使用的是 Connections 类的默认值 weight = 1.0，但实际上并非如此。
如果您仔细观察，您确实在定义addconnection 时将weight = 0.0 指定为默认参数，对吧？ :
def addconnection(self,sender,weight=0.0):

并且由于您使用
self.connections.append(Connection(self,sender,weight))
调用 Connection 类 __init__ 方法 你实际上传递了一个权重值：你在 addconnection 方法中指定的默认 0.0。所以所有连接都将具有默认权重0.0。这有效地将输入的所有值锁定在 0.0，激活锁定在 -0.5。

要改变这一点，你也许可以在第 75 行指定一个权重，在那里你使用addconnection 方法，和/或只让addconnection 有一个权重的默认值（并让它为 1.0），而 Connection 类__init__ 方法应该总是需要一个 weight 值，没有默认值。这就是我在下面的代码中所做的，加上一些__str__ 方法来检查。

（这是addconnection中默认值为1.0的版本，而连接__init__中没有默认值）：

[编辑：添加了net_input的第一次初始化。]

# 
#                               Preparations 
# 

nodes=[] 
NUMNODES=3

# 
#                                   Defining Node Class
# 

class Node(object): 

    def __init__(self,name=None): 
        self.name=name 
        self.activation_threshold=1.0
        self.net_input=0.0
        self.outgoing_connections=[] 
        self.incoming_connections=[] 
        self.connections=[] 
        self.activation=None

    def __str__(self):
        return self.name

    def addconnection(self,sender,weight=1.0):
        self.connections.append(Connection(self,sender,weight)) 
        print "Node", str(self), "now contains", str(self.connections[-1])

    def update_input(self): 
        self.net_input=0.0
        for conn in self.connections: 
            self.net_input += conn.weight * conn.sender.activation 
        print 'Updated Input for node', str(self), 'is', self.net_input 

    def update_activation(self): 
        self.activation = self.net_input - 0.5
        print 'Updated Activation for node', str(self), 'is', self.activation 

# 
#                                   Defining Connection Class
# 

class Connection(object): 
    def __init__(self, sender, reciever, weight): 
        self.weight=weight 
        self.sender=sender 
        self.reciever=reciever 
        sender.outgoing_connections.append(self) 
        reciever.incoming_connections.append(self) 
        print 'Created', str(self)

    def __str__(self):
        string = "Connection from " + str(self.sender) + " to " + str(self.reciever) + ", weight = " + str(self.weight)
        return string
# 
#                                 Other Programs 
# 


def set_activations(act_vector): 
    """Activation vector must be same length as nodes list"""
    for i in xrange(len(act_vector)): 
        nodes[i].activation = act_vector[i] 


for i in xrange(NUMNODES): 
    nodes.append(Node(str(i)))
    print "Created node:", nodes[i]


for i in xrange(NUMNODES):#go thru all the nodes calling them i 
    for j in xrange(NUMNODES):#go thru all the nodes calling them j 
        if i!=j:#as long as i and j are not the same 
            nodes[i].addconnection(nodes[j])#connects the nodes together
#
#                                         Setting Activations
#
set_activations([1.0,1.0,0.0])

#
#                                        Running 10 Iterations
#
for thing in nodes:
    thing.update_input() #initializing inputs

for i in xrange(10): 
    for thing in nodes: 
        thing.update_activation() 
        thing.update_input()