Python 如何在 Keras 模型中初始化偏差？
声明：本页面是StackOverFlow热门问题的中英对照翻译，遵循CC BY-SA 4.0协议，如果您需要使用它，必须同样遵循CC BY-SA许可，注明原文地址和作者信息，同时你必须将它归于原作者（不是我）：StackOverFlow
原文地址: http://stackoverflow.com/questions/40708169/
Warning: these are provided under cc-by-sa 4.0 license. You are free to use/share it, But you must attribute it to the original authors (not me):
StackOverFlow
How to initialize biases in a Keras model?
提问by Mohammad Amin
I am trying to build a synthetic model in Keras, and I need to assign values for the weights and biases. Assigning the weights is easy, I am using the instructions provided here: https://keras.io/initializations/. However, I could not find any instructions on how to assign the biases. Any ideas?
我正在尝试在 Keras 中构建一个合成模型，我需要为权重和偏差分配值。分配权重很容易，我使用这里提供的说明：https: //keras.io/initializations/。但是，我找不到有关如何分配偏差的任何说明。有任何想法吗？
回答by StatsSorceress
You can also use bias_initializer like this:
你也可以像这样使用bias_initializer：
model.add(Dense(64,
kernel_initializer='random_uniform',
bias_initializer='zeros')
This is from https://keras.io/initializers/
回答by Hengda Qi
You can find the answer here. https://keras.io/layers/core/
你可以在这里找到答案。 https://keras.io/layers/core/
weights: list of Numpy arrays to set as initial weights. The list should have 2 elements, of shape (input_dim, output_dim) and (output_dim,) for weights and biases respectively.
weights：要设置为初始权重的 Numpy 数组列表。该列表应该有 2 个元素，形状分别为 (input_dim, output_dim) 和 (output_dim,) 用于权重和偏差。
When adding a new layer, you can define the argument "weights", a list that contains initial wand bwith shape speicified.
添加新层时，您可以定义参数“权重”，这是一个包含指定形状的初始w和b的列表。
model.add(Dense(50, input_dim= X_train.shape[1], weights = [np.zeros([692, 50]), np.zeros(50)]))
model.add(Dense(50, input_dim= X_train.shape[1], weights = [np.zeros([692, 50]), np.zeros(50)]))
回答by daoliker
Initialize biases with small positive value such as 0.1
用小的正值（例如 0.1）初始化偏差
Since we're using ReLU neurons, it is also good practice to initialize them with a slightly positive initial bias to avoid "dead neurons".
由于我们使用的是 ReLU 神经元，因此将它们初始化为稍微正的初始偏置以避免“死神经元”也是一种很好的做法。
回答by Trevor Witter
Weight and bias initialization for each layer can be set via kernel_initializerand bias_initializerkeyword arguments respectively within layers.Dense(). If undefined by user, default settings of kernel_initializer='glorot_uniform'and bias_initializer='zeros'are applied.
每层的权重和偏置初始化可以分别通过kernel_initializer和bias_initializer关键字参数在layers.Dense(). 如果用户未定义kernel_initializer='glorot_uniform'，bias_initializer='zeros'则应用和 的默认设置。
For example, if you wanted to initialize a layer's weight initialization to random uniform instead of glorot and bias initialization to 0.1 instead of 0, you could define a given layer as follows:
例如，如果您想将一个层的权重初始化初始化为随机统一而不是 glorot 并将偏置初始化初始化为 0.1 而不是 0，您可以按如下方式定义给定层：
from keras import layers, initializers
layer = layers.Dense(64,
activation='relu',
kernel_initializer='random_uniform',
bias_initializer=initializers.Constant(0.1))(previous_layer)
See layers/core/for details on Dense layer keyword arguments and initializers/for preset and customizable initializer options
