python - 如何在Python中找到分类神经网络的预测输出？

Question

我是Python和学习神经网络的新手。我有一个训练有素的 3 层前馈神经网络，隐藏层有 2 个神经元，输出层有 3 个神经元。我想知道如何计算输出层值/预测输出

我从网络中提取了权重和偏差，并计算了隐藏层的激活值。我只是想确认一下如何使用softmax函数来计算输出层神经元的输出？

我的实现如下:

weights_from_hiddenLayer_to_OutputLayer = [
    [x, y],  # two weights connected to the output neuron 1 from hidden neurons 1 and 2
    [a, b],  # two weights connected to the output neuron 2 from hidden neurons 1 and 2
    [c, d]   # two weights connected to the output neuron 3 from hidden neurons 1 and 2I
    ]

# output layer biases extracted from the neural network
biases_output_layer = [a, b, c]

act1 = m  # activation value of hidden neuron 1
act2 = n  # activation value of hidden neuron 2
arr = []
for i, weights in enumerate(weights_from_hiddenLayer_to_OutputLayer):
            arr.append(m*weights[0]+n*weights[1] +
                       biases_output_layer[i])
# i believe this will be the brightest neuron / predicted neural networks output ?  
print(np.argmax(arr))

我在互联网上搜索了在 python 中使用 softmax 的信息，我已经找到了。我的预测输出与神经网络的预测大部分不同。而我使用的是来自同一训练模型的完全相同的值。

Answer 1

您的输出将是 weights_from_hiddenLayer_to_OutputLayer 与之前激活值的矩阵乘法。 然后，您可以将其传递给 softmax 函数来获取概率分布，并根据您的猜测使用 argmax 来获取相应的类。

weights_from_hiddenLayer_to_OutputLayer = np.array([
    [x, y],  # two weights connected to the output neuron 1 from hidden neurons 1 and 2
    [a, b],  # two weights connected to the output neuron 2 from hidden neurons 1 and 2
    [c, d]   # two weights connected to the output neuron 3 from hidden neurons 1 and 2I
    ])

act = np.array([m, n])
biases_output_layer = [a, b, c]
arr = np.dot(weights_from_hiddenLayer_to_OutputLayer, act)    # matrix multiplication of weights and activations
arr = arr + biases_output_layer
     
probability = np.exp(arr) / np.sum(np.exp(arr), axis=0)       # softmax
print(np.argmax(probability))

请注意，从技术上讲，您不需要使用 softmax，除非您要反向传播或尝试评估输出的置信度，因为 np.argmax() 的结果无论如何都是相同的是否传入 arr 还是相应的概率。